Thanks, as ever, for the support you show my writing at Reasonable People. Your subscriptions have been a deep encouragement to me to keep writing here.

I turned on paid subscriptions at the beginning of my year-long career break, when I wasn’t sure what I’d be doing (or if I’d have any income for the year). Now, incredibly, a year has passed and I’m returned to being a jobbing academic at the University of Sheffield (gulp!).

If you’re a paid subscriber you’ll notice I’ve turned off paid subscriptions, so you shouldn’t be billed any more. I do plan to keep writing, but I think I need to find out what my new routine is when I go back to the University before committing to writing as regularly as I have in the last year.

It’s been a good year,

Some of the time I’ve taken to goof off, read books, hang out, go places, meet interesting people, write and generally explore my interests without thought of profit or productivity. Refreshing!

For a good half of the time I’ve been working with RoRI on research on research / metascience (lots more to say on this soon!). I’ve been in London every week at the Wellcome Trust, bringing research on research expertise in-house there, and I’m going to keep being in London weekly, so do get in touch if you are there and think we could have an interesting chat about any of the topics in this newsletter.

I’ve become a visiting Professor a the Department of Computer Science and Technology, University of Cambridge. This formalises my collaboration with Andreas Vlachos, who I have a number of exciting projects with. The combination of thinking about the fundamental mechanisms of how language models work and on the fundamental mechanisms of how humans reason has been really productive, with implications for how we use machines to check facts, and how we talk to each other (lots more to say on this soon!).

At the end of the year I’m glad to be going back being a full-time academic. It’s a job I feel very privileged to have, and I’m looking forward to doing some teaching and supervising again.

From the year out I’m taking a readjustment of my perspective on the balance of activities I want to spend my time on, and some big ambitions for new projects (lots more to say on this soon!).

Thanks for reading. Normal posting will resume soon, along with updates on various project when I can. In the meantime thanks for reading and here’s a picture of Conisborough Castle I took this morning:

More soon!

Get in touch — I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

• Gradually, then all at once. Why belief change is like bankruptcy

• Predicting human behaviour. Why it is hard, and why LLMs may not help

• Social thinking. How other people, even if just imagined, can help puncture the illusion of understanding

The Boundary of Love by Luo Li Rong

More: Body in Motion: Fluid Bronze Sculptures by Luo Li Rong

via @Natasha_Jay

Boiled clergy

Boiled clergy. Livre de la Vigne nostre Seigneur, France ca. 1450-1470. Bodleian, MS. Douce 134, fol. 85r.

via @medieval_illuminations

I’m all ears, pal

via @ComicContext

Our new theory…

By tomgauld.com

This newsletter is free for everyone to read and always will be. If you can afford it, feel free to chip in to help me keep writing for everyone: upgrade to a paid subscription (more on why here).

Upgrade to paid

Reasonable People is having August off. Thanks for reading!

Pig of knowledge

This wonderful 1805 book exposes the reality behind various tricks and illusions of the day, beginning with arguably the most enigmatic of the tricks listed, that of the “Learned Pig”, or as the excellent frontispiece refers to it “The Pig of Knowledge”:

via @publicdomainrev, original The Expositor, or Many Mysteries Unravelled - including that of the Learned Pig (1805)

As your father

Original: User Stories for Clarity

Montreal street poster

via @ianrogers

It’s not good

If you don’t know where this gif comes from it is a just a cute scene with some bunnies. If you do, it may drag atavistic terror from the buried depths of your childhood memories. Sorry.

via @el_ahrairah

« Arcane XI - La Force », 1993

by Juan Gimenez (Argentine artist, 1943-2020)
Illustration from the “22 Major Arcana of the Marseille Tarot” collection cards.

via @vintagefantasyart

Latent Space, Felicity Hammond

In her installation, artist #FelicityHammond offers a speculative glimpse into a not-too-distant future where this new approach to space-based computation has become the dominant position in the AI industry. However, the system continues to battle with the effects of #modelcollapse...

Sorry I couldn’t make this exhibition in Ljubljana in April

via @aksioma

…And finally

One of my own holiday snaps. Ten points if you can tell where I’ve been

END

Comments? Feedback? Things you’ve seen? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

AI declaration: I write all the words and think all the thoughts myself.

Conversion experiences are like lightning flashes over the darkened landscape of belief. For most of us, on most topics, we avoid articulating exactly what we think at every moment. This allows the structure of our beliefs to warp and shift over the years, as we gradually drift in how we feel on particular topics. On topics as big as who we support politically or as small as what we believe about the correct way to cook rice, we can change without having to acknowledge to anyone, ourselves included, a singular point at which we switched. Conversions are the exception, the explicit rupture with prior beliefs. Anyone can experience one, but it isn’t surprising that the ones we know most about are from public figures, preachers, pundits and politicians whose job is to regularly profess what they believe.

With someone like Loomer, it is tempting to ask “why not sooner?” Given the wealth of evidence, did she really need to go to Ukraine, to experience an air raid, to convert from one position to the other? If the evidence is so compelling, why did it not change Loomer’s mind earlier?

Asking this would be to make a common mistake about how beliefs change. Our minds are not passive storage for information. We do not instinctively weigh out the evidence for a position like we weigh out sugar in a cake recipe.

The opposite, but more severe, error is to believe that nobody ever changes their mind because of evidence, and people’s commitments are entirely due to their economic incentives, or social connections or deep psychological desires. Obviously we do all have these motivations. Someone like Loomer is more conflicted than most — she makes her money by being the centre of attention for having controversial beliefs. However, as her case illustrates, beliefs do change, and a world where evidence is irrelevant is as cartoonish as one where everyone adjusts their beliefs in perfect response to the evidence.

How can we make sense, then, of the inconsistent relation between the evidence in front of us and our beliefs? Religious narratives describe conversions as flashes of light. A powerful voice from the heavens reveals the Truth and illusion falls away. The psychological truth of conversion experiences is in the rift between before and after. Prior beliefs seem incomprehensible post-conversion, just as pre-conversion the critical evidence seemed insufficient to justify a change in belief.

Beliefs change like businesses go bankrupt: gradually then all at once. With bankruptcy, the business is solvent while it loses money, but at some point the owners (or more likely the creditors) lose faith that it will ever stop losing money, and decide to cut their losses. The loss of faith creates a tipping point where the conditions which were sustainable previously (a daily loss) become unsustainable.

Our beliefs exist in a network with all our other beliefs, with our habits and commitments, and so form a complex system. Our beliefs about who we trust tug on our willingness to believe what we’re told. Our beliefs about the way the world works pull on our beliefs about what is likely to happen, or what a plausible explanation for events is. Our knowledge of what we said yesterday pushes up against what we believe we should say today. The whole network of pushes and pulls creates a tension which holds existing beliefs in place. No wonder a single new piece of evidence can sometimes seem like it is overlooked.

But beliefs do change. A tipping point is reached in which the network of beliefs shifts into a new configuration. The facts haven’t changed but the cumulative weight has crossed a threshold. It may not even be a large thing, but some perturbation knocks a belief-network that previously seemed stable off balance, and a new configuration snaps into existence. Under the threshold, gradually, the links between different beliefs were changing, and then, all at once, we come to realise we now believe something different from what we did yesterday.

This is why argument looks like it has such a hard time being effective, and it is easy to believe in the horror-story of human irrationality (that people don’t change, or just believe what is convenient, or whatever). Making arguments and providing evidence doesn’t work immediately, or consistently, and it doesn’t work on its own, but it does work.

If my account of beliefs as complex networks is right, then some of the time the most effective arguments are changing beliefs under the surface. Even if the person you are arguing with doesn’t admit it, they may be changed by your arguments, with the change in belief emerging hours, days, or years later.

This complexity is the missing ingredient in understanding bounded rationality (https://en.wikipedia.org/wiki/Bounded_rationality). Bounded rationality is the idea that our minds are adapted to make the best sense of an uncertain world in the limited time and with the limited resources we have. This makes us less than ideal reasoners, but still more than simple machines with predictable responses. We each carry a certain amount of inconsistency, but not so much that we should give up on reasoning altogether.

Beliefs are only useful if they are connected to other things, but these connections make the input-output relationship complex. Just as strong evidence doesn’t move people as much as it should, sometimes small things trigger a cascade. If you want to change minds it takes an enduring faith in reason to keep trying.

Reasonable People is having August off. Thanks for reading!

This newsletter is free for everyone to read and always will be. If you can afford it, feel free to chip in to help me keep writing for everyone: upgrade to a paid subscription (more on why here).

Upgrade to paid

See below for references, and other things I’ve noticed.

Related:

Previously, by me, a review of a very relevant book:

• Unfollow. Reasonable People #28: Megan Phelps-Roper’s story of losing faith in the Westboro Baptist Church

This is a very good podcast generally, and the discussion across four episodes illustrates many of the complexities I didn’t get into in this newsletter.

• Past, Present, Future (“The History of Ideas Podcast
  with David Runciman”) on Political Conversions: 1 Going Fascist, 2 Communism – The God That Failed, 3 From Trotskyism to Neoconservatism, 4 Switching Sides in the 21st Century

And of course, David McRaney wrote the book “How Mind’s Change”. This reprise of the story from one chapter is great:

• YANSS 342 - The Conspiracy Theorist Who Changed His Mind

What do you think?

How do you feel about the balance between reporting and speculation in this newsletter? Normally I feel better when I am trying to report on some primary empirical finding, or at least support my reasoning with careful citation to the literature in cognitive science. Sometimes I break out of this mode and let myself indulge in trying to articulate how I think about things. Today’s post is an example of this. There are no new facts here, it really just (re)description, and without references to the reasons I have come to think about belief change in this way.

Examples of recent posts which are primarily empirical reporting:

• Striking evidence for forced experimentation. A natural experiment shows how we always have something to learn, even about the things we are most familiar with.

• Another algorithm is possible. How social media could easily be changed to stop us hating each other so much

• AI systems out-persuade expert humans. Panic? Quick review of a compelling new report

Examples of recent posts which are also more speculative:

• Social thinking. How other people, even if just imagined, can help puncture the illusion of understanding

• Why we like to believe other people are stupid Reason sells, but who’s buying?

• Predicting human behaviour. Why it is hard, and why LLMs may not help

Are the speculative posts worth reading? Should I keep them?

…And finally

The Far Side Bot has reminded me of Gary Larson’s voluminous output. There’s probably one about belief change in there, but I couldn’t find it, so I share this instead:

END

Comments? Feedback? Change my mind? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

AI declaration: I write all the words and think all the thoughts myself. I asked Claude to check for spelling and grammar.

“it is possible to simulate real people with high accuracy. We are now developing a foundation model that predicts human behavior in any situation, at any scale.”

Dr Lee gives an overview of the current state of the art on LLMs and what they can (and can’t do). He also rips into the claims that language models are a way of avoiding the effortful gathering of data. His critique is worth listening to in full, but he makes some comments about the fundamental challenges of predicting human behaviour which apply to all of behavioural science, not just the part that might be seeking to use new technology. Because the same limitations hold however you try and predict human behaviour, I thought it was worth expanding on them here.

It is instructive to reduce prediction to a simple form. If the situation is X, what is the predicted outcome, Y? Even this simplest conception gives us a space of possible situations (inputs) and outcomes (outputs).

Science starts by collecting points which have already been observed (evidence) and using them to infer points which haven’t yet been observed (predictions). Both can be expressed as links between situations and outcomes (X-Y pairs).

Then, using the gathered evidence, models are built which capture the available evidence. Good models also capture future evidence (aka predictions). It is the factors which stop models being good predictors which are of interest to Lee (and us).

In the pure world of mathematical functions if you get the right model for the data you know everything about the data - all current observable points as well as all possible future points, out to infinity.

This is the principle that physics has applied so successfully. The same mechanical foundation can predict the trajectory of a ball kicked in New York and that of a rocket fired from Malindi, Kenya. These foundational models, says Lee, are fundamentally different from the synthetic humans created by companies which purport to provide digital twins for individuals or populations using LLMs. Foundational models translate across contexts because they are based on uncovering the generating principles which create the observed data (the “causal model”). LLMs are models which capture the observed data, but don’t encapsulate the underlying principles, so they have no foundation to translate to novel situations.

Why ever might people think these data-models are good at prediction?

Lee observes that LLMs are general purpose ‘function approximators’. They find patterns in data. They can fit the systematic patterns in human text, but they could just as easily find patterns in random noise. When they get trained with human text they compress those data and can so make plausible outputs (recovering patterns, from the simple, like “Happy” is likely to be followed by “Birthday”, to the complex, like the range of responses you get to a prompt when you ask the LLM some complex task).

For powerful function approximators, ‘predicting’ anything which is in the sample is “trivial”. If you have already seen that X1 leads to Y1 in the training data, then predicting X1 leads to Y1 just follows from what you have trained the model to do.

That some of the abilities of LLMs result from this fact is obscured by the fact that LLMs hoover up a vast, incomprehensible, amount of data. This is impressive, but the ability to reflect and re-represent the training data shouldn’t give us confidence that the model can do more than that.

In our schematic, this is as if the model sees points (X1,Y1), (X2, Y2), and (X3,Y3). Questions about X1, X2 or X3 will get perfect “predictions”. Lee is also unimpressed by interpolation within the range of the observed data, something like asking about an input halfway between X1 and X2. Within this scheme, the ability to do both recall and interpolation tells us nothing about the ability to extrapolate beyond the data, to predict, given X4, the output Y4.

Extrapolation is precisely what you want a good model to do. One way to get this is a foundational set of principles which translate across contexts (like Newtonian mechanism does for local space and time). Another is to gather a mass of training data and hope the secret sauce of deep learning will uncover patterns in the data which support extrapolation beyond the data. Lee is sceptical this can be the case for human behaviour.

He is sceptical of the use of LLM models for non-observed data for a couple of reasons.

One reason is the fundamental difficulty of validating a model for cases which you want to make predictions for. By definition the cases which you want to make predictions for are more likely to be novel, and to the extent they are novel there isn’t training data, and if there isn’t data how can you validate? To hear Lee tell it, the whole field of digital twins and synthetic polling using LLMs for human behaviour is snake oil based on assuming the models will produce good outputs when this can’t be known by definition. The outputs are plausible for what is already known, but this strength (capturing observed data) is precisely also their limitation.

The ability to discern this is confused by ‘data contamination’ in the benchmarks used to test models. Because LLMs have hoovered up such a vast amount of training data, many standard tests (and their answers) are included in the training data. When the models look like they are reasoning (extrapolating) they may just be producing the outputs they’ve already seen in their training data (recall).

A second reason that the validation of models of human behaviour is hard, and this is particularly true for LLMs, is that the models are brittle. This means they produce wide variability in their outputs in response to small changes in input. So, for example, small changes in prompt wording —rephrasing that doesn’t change the question — can produce completely opposite answers (indeed, there is a whole area of ‘prompt injection’ dedicated to engineering peculiar ways of phrasing requests that get around the normal response of the language model — usually with the aim of subverting a refusal).

These limitations might be of less concern if humans were Newtonian objects: solid, immutable and affected only by a limited set of physical forces. Instead, though, we are both adaptive (changing based on our experience and the momentary environment) as well as heterogeneous (varying between people, as well as across time). These factors mean that the observable data has limited value in predicting behaviour in new circumstances.

Here Lee’s critique becomes, in my opinion, a general one for psychological science. Modelling observed data will be fundamentally limited without building a causal account, an account of “why” people do what they do. LLMs are incredible new technology, but that doesn’t mean they can escape this fundamental limitation. If they are, at heart, a data-modelling exercise they capture what already is, but not beyond. They demonstrate just how much of the human world has been compressed into language, but they will be limited precisely to what has been experienced and written about so far, and will have limited ability to extrapolate to the truly novel.

If a market researcher wants to know how people will respond to a new product, or a political pollster wants to know how people will vote in an election, the models will give plausible answers. That’s what they do, but the answers will only be plausible, no more. To the extent that the situation were are asking about is novel, exactly when we might gain from accurate predictions, they will be useless.

There’s more in the talk I haven’t covered, it really is an excellent introduction to both the limitations of LLMs and some deeper issues about why psychological science is so hard. Recommended.

This newsletter is free for everyone to read and always will be. If you can afford it, feel free to chip in to help me keep writing for everyone: upgrade to a paid subscription (more on why here).

Upgrade to paid

See below for references, and other things I’ve noticed.

References

Dokyun Lee. Scylla Ex Machina: The Illusion of LLMs as Human Surrogates. 05-07-2026, hosted by the Digital Economy Network https://www.digitalecon.org/seminar

Website of Dokyun Lee: https://www.leedokyun.com/

Argyle, L. P., Busby, E. C., Fulda, N., Gubler, J. R., Rytting, C., & Wingate, D. (2023). Out of one, many: Using language models to simulate human samples. Political Analysis, 31(3), 337-351.

Other things…

Small announcement

I’m going to take August off, so next week’s newsletter will be the last until September. Happy holidays!

Podcast: The Great Political Fictions: The Dispossessed

From the History of Ideas Podcast with David Runciman, a review of Ursula Le Guin’s novel ‘The Dispossessed’. Runciman identifies Le Guin’s great achievement of portraying an anarchist utopia that seems plausible, both psychologically and politically. He is also draws out that the difference between the anarchist moon colony (the utopia) and the degenerate statists of the origin world is not exactly in freedom from coercion. Freedom exists on both, either by material abundance (origin world), of by the lack of the structured of violent control (anarchist moon), as does coercion, by force (origin world) or convention, upbringing and social pressure (anarchist moon). The real different is the honesty with which the anarchists recognise the compromises they have made to make society work. The corrupt elites of the origin world boast that they are only recognising the real politic of force, but at the same time they give lie to this because they cannot allow the truth about the organisation of their society to be told, it has to be dressed up in sentimental pageantry and ideals which only ring louder and more hollow as they are violated in the streets.

David Runcimen: The Great Political Fictions: The Dispossessed

Catch-up

• Social thinking. How other people, even if just imagined, can help puncture the illusion of understanding

• Striking evidence for forced experimentation. A natural experiment shows how we always have something to learn, even about the things we are most familiar with.

• Another algorithm is possible. How social media could easily be changed to stop us hating each other so much

• AI systems out-persuade expert humans. Panic? Quick review of a compelling new report

• Why we like to believe other people are stupid. Reason sells, but who’s buying?

… And finally

Funny because true?

END

Comments? Feedback? Synthetic predictions? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

AI declaration: I write all the words and think all the thoughts myself. I asked Gemini to check for spelling and grammar. For this post I also asked Claude to make the plots (link).

If you have any self-awareness, I’m sure you can recall a time it has happened to you. If you have better luck than me then it hasn’t been at the point you stood up to teach a class.

The illusion of explanatory depth is where you have a feeling of familiarity with a topic, which prevents you realising the extent of your ignorance. You can recognise a car engine, the human eye, or the government’s fiscal stimulus policy. You can even have enjoyable conversations about them, but that is a very different thing from understanding how they work.

The illusion has a surprising strength. It is strong enough to last right up to the point we are forced to explain our understanding out loud (and for some of us, even beyond that). Probably the illusion exists for a reason. It protects us from fresh interrogation of everything we think we know, allowing us to think with concepts, rather than about them. Navigating rush hour would be much harder work if you had to think about the fundamental nature of your understanding of car mechanics at the same time.

The artificial circumstance of teaching is almost guaranteed to snap you out of the illusion that you understand something when you don’t. Hopefully this happens before the class starts, as you put yourself in the place of a student who will learn about something for the first time and imagine how you are going to help them understand it.

Sometimes just the planning and anticipation of teaching is enough for me to realise I don’t fully understand something.

Sometimes, though, even knowing I need to explain something and planning to explain it still doesn’t shake loose the presumption that I understand it. A basic law of thought is that we tend to preserve energy by not thinking when we don’t have to — we are cognitive misers, hence our reliance on habit and other shortcuts. This is brought home to me when my illusion of understanding is preserved right up until I am forced to put an explanation into words. Only at that point does the file I’ve mentally (and miserly) labelled “explanation here” evaporate into nothing.

When I put something into words, whether in inner speech, or by writing it out, it converts my thinking into an object I can then scrutinise, and so apply the scepticism that needs to be applied to test the quality of an explanation.

Putting thoughts into words in front of other people brings even more self-scrutiny, partly out of self-consciousness, and partly out of a kind of priming of the perspectives of others. When I speak to a particular audience I can’t help but imagine their responses, either from specific individuals I know, or from possible types who may be present (a sceptic from a particular ideological position, for example).

Just imagining that you might be wrong is enough to create demonstrable benefits to problem solving and estimation, which has been variously called “the wisdom of crowds on single player mode” and “dialectical bootstrapping”. I like to think that the imaginary crowd has a similar salutary effect on more complex thoughts. The audience does more than just make me imagine I might be wrong. It invites me to consider, from the diverse perspective of the audience members, the multiple dimensions along which I might be wrong.

Phenomena like these mean that thinking is never really a solitary activity. It is fundamentally social. Even when we are talking to ourselves we are using the socially orientated mechanisms of language to externalise our thoughts and feed them back “into” our minds. Once private thoughts have been forced through the tubes of language they take on an unmissable scent of the outside world.

Finally, there is no substitute for the real world, and in particular discussing an explanation with real people. When this happens I no longer need to simulate critical responses, I get them direct. This is one of the reasons why groups can outperform individuals in various cognitive tasks, as the different individuals bring different knowledge and perspectives to a task and so pool these within the group.

But overt exchange of feedback from others is only one of the reasons groups might enhance problem solving in the members. If, as I’ve argued, just thinking about what other people might say can allow us to root out errors or incompleteness in our thinking, then the mere presence of the group can enhance problem solving. Even without the group actually being present, the mere imagination of the group might also bring a benefit. And even without the group being explicitly imagined, the processes of thinking out loud (or even mentally) are patterned by our evolution as social reasoners.

Every live group discussion will also trigger all the other benefits, just as you can get some of the benefits of a live group discussion through mental simulation. I imagine it like a set of concentric layers: a live group can give actual feedback, but will also vividly prime the existence of alternative perspectives. Imagining alternative perspectives will trigger self-consciousness, and being self-conscious/self-scrutinising will require you to articulate your understanding in language. You can have inner layers without outer layers, but not vice versa:

This creates a dilemma for our studies of the benefits of groups for problem solving. Which of these different levels is necessary or sufficient to produce any benefits we see? We can test people on their own, or in groups, but the aggregate effect will be due to a collection of all the different processes triggered by groups.

Only by adding additional, non-group, non-solo, conditions would we be able to discern what it was about groups that brought the cognitive benefit. Something like:

• Group 1: Asked to solve a problem

• Group 2: Thinks aloud while they solve the problem

• Group 3: Thinks aloud, while in front of a group

• Group 4: Group discussion

How far you can get just with mental simulation would be the contrast between Group 2 or 3 and group 1. The benefit you can only get from interacting with real people is the contrast between Group 3 and Group 4. Most research on the benefits of group function just compares the full suite (Group 4) against unadorned solo efforts (Group 1), and so probably overestimates the benefit of group discussion per se, failing to account for the power of simulated reflection.

While this elaboration has a moral for psychology research, there is also a practical implication for anyone who wants to straighten out their thoughts and avoid falling for the illusion of explanatory depth. First you need to adopt the discipline of forcing yourself to articulate what you think you understand. Next you need to cultivate your imagination about, and sensitivity to, the different reactions your explanation might provoke. And finally, if you can, you’ll learn a lot by putting your thinking out in front of other people.

I’ve found writing a newsletter is good for all three of these. Thanks for reading!

This newsletter is free for everyone to read and always will be. If you can afford it, feel free to chip in to help me keep writing for everyone: upgrade to a paid subscription (more on why here).

Upgrade to paid

See below for references, and other things I’ve noticed.

References

Rozenblit, L., & Keil, F. (2002). The misunderstood limits of folk science: An illusion of explanatory depth. Cognitive science, 26(5), 521-562. https://doi.org/10.1207/s15516709cog2605_1

Herzog, S. M., & Hertwig, R. (2009). The wisdom of many in one mind: Improving individual judgments with dialectical bootstrapping. Psychological Science, 20(2), 231-237. https://doi.org/10.1111/j.1467-9280.2009.02271.x

Landemore, H. (2017). Democratic reason: Politics, collective intelligence, and the rule of the many. Princeton University Press.

Other things…

Paper: Algorithmic Monocultures in Hiring

If every company uses the same AI tool to screen CVs, it compounds the danger that someone will get wrongly rejected — if they are rejected for one job they are more likely to be rejected from all jobs. It’s a case where there is a second-order risk of AI tools on decision making. No tool will avoid all mistakes, but the current economy means a small number of AI companies dominate, meaning that any mistakes in their tools also dominate.

I take from this that we should design in more diversity in our AI ecosystem.

Page: algorithmichiring.github.io

Bommasani, R., Bana, S. H., Creel, K. A., Jurafsky, D., & Liang, P. (2026, June). Algorithmic monocultures in hiring. In The 2026 ACM Conference on Fairness, Accountability, and Transparency (pp. 6351-6382).

Wealthy warned over HMRC’s use of AI in tax crackdown

“Wealthy individuals are being warned that their tax affairs are more likely than ever to come under scrutiny as HMRC steps up its use of AI and data analytics for intelligence gathering.

The alert comes from advisers after the tax authority this week said that the use of digital analytics and surveillance tools helped bring in or protect about £10bn in tax in the last financial year.”

Seems like not just a Good Thing, but an interesting example of playing to AIs strengths — AI has the bandwidth to allow wider scrutiny, which can then flag individual cases for human review.

Link: Wealthy warned over HMRC’s use of AI in tax crackdown. Financial Times, 2026-07-11

Catch-up

• Striking evidence for forced experimentation. A natural experiment shows how we always have something to learn, even about the things we are most familiar with.

• Another algorithm is possible. How social media could easily be changed to stop us hating each other so much

• AI systems out-persuade expert humans. Panic? Quick review of a compelling new report

• Why we like to believe other people are stupid. Reason sells, but who’s buying?

Social Media Is Now Parasocial Media

When practitioners used the term “social media” to describe the internet tools that emerged in the mid-aughts, they were giving a name to the kinds of platforms and protocols that allowed people to socialize with friends and communities of interest by using digital technologies. Twenty years later, users of social media are far more likely to scroll than post – and the content that they consume is often strategically produced and algorithmically curated. In this essay, I argue that the very essence of social media has changed. To more effectively interrogate what we are witnessing, we need to stop presuming that these tools are “social media” and begin recognizing that they are now “parasocial media.” Doing so raises new questions about digitally mediated sociality, not to mention the politics and governance of these platforms.

boyd, d. (2026). Social media is now parasocial media. Social Media+ Society, 12(2), 20563051261437487. https://doi.org/10.1177/20563051261437487

The Empty Chair: Using LLMs to Raise Missing Perspectives in Policy Deliberations

A case study, so doesn’t really prove anything, but it is a nice idea of the sort of thing you could use AI for (not without risk)

In this work, we explore the use of large language model (LLM) personas to introduce missing perspectives in policy deliberations. We develop and evaluate a tool that transcribes conversations in real-time and simulates input from relevant but absent stakeholders. We deploy this tool in a 19-person student citizens’ assembly on campus sustainability. Participants and facilitators found that the tool was useful to spark new discussions and surfaced valuable perspectives they had not previously considered. However, they also raised skepticism about the ability of LLMs to accurately characterize the perspectives of different groups, especially ones that are already underrepresented.

Fulay, S., Dimitrakopoulou, D., & Roy, D. (2025). The Empty Chair: Using LLMs to Raise Missing Perspectives in Policy Deliberations. arXiv preprint arXiv:2503.13812.

…And finally

The 9 purple dots illusion. It’s a static image, 9 blue dots but the one(s) you look at appear(s) purple. A key ingredient is that the centre of your vision is less sensitive to blue light. The diminished sensitivity to blue in the centre shifts the percept towards purple (because purple is a mix of the signals from the red and blue receptors in the eye). Neat.

Paper: Schulz-Hildebrandt, H. (2026). When purple perceived only at fixation: A fixation-and distance-dependent color illusion. Perception, 03010066261423048.

END

Comments? Feedback? What didn’t I understand? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

AI declaration: I write all the words and think all the thoughts myself. I asked Gemini to check for spelling and grammar. For this post I also asked Claude to make the concentric circles diagram. The version Claude and I produced was much improved by feedback from H. (a human).

Habit is one of the great predictors of human behaviour (along with what the other people around you are doing). For a lot of what we do, we rely on repeating what we’ve done before. In learning theory this is ‘exploitation’ — taking advantage of what you already know, and what has already worked for you before. Its counterpart is ‘exploration’ — trying something new, with an uncertain outcome. With exploration you learn something new, even if what you learn is that whatever you tried wasn’t worth it.

Exploration-exploitation is one of the fundamental trade-offs. There’s no right answer. It depends on what you have to lose, how much you think you know about your options, and how complex the choices you face are.

For commuters, the choice is a very complex one. Travel routes are famously hard to solve. They belong to a class of mathematical problems where the best answer is hard to find, but easy to verify. If you identify a new route it is easy to show that it is shorter than your current best option, but before you identify it, you don’t even know if a better option exists and there’s no easy way of knowing.

As well as being theoretically intractable, commuting in London also has multiple possible modes (road, rail, tube and river), which provide multiple routes within and combining modes, and each potential route has a characteristic travel time, as well as variability (one route might be quicker on average but be less predictable), along with idiosyncrasies which might make different routes attractive (like passing a great coffee shop). All this complexity makes the best route for each individual commuter inscrutable. The only way to solve the problem is to try different routes out, and evidence from the strike suggests that commuters under-explore. They rely too much on habit, to their own long-term cost.

In their paper, The Benefits of Forced Experimentation: Striking Evidence from the London Underground Network, Larcom et al (2017) report analysis of London Underground commutes, gathered from 3.5 million travelcard (’Oyster’) IDs. From these IDs they identified regular commuters who made the same journey, entering and exiting via the same stations, every working day between 7am and 10am from January 15th 2014 to February 4th (the day before the strike).

These regular commuters make up the population for a natural experiment. On the two strike days some get to travel their normal route (because the strike doesn’t affect the stations they use), and some are forced by the strike to experiment, trying out new routes. The focus of the analysis by Larcom et al (2017) is whether those who have to experiment keep using a new route in the days after the strike. If they switch to a new route, despite the end of the strike meaning their old route was available again, it suggests that they’ve found a new route which, for whatever reason, they prefer.

The analysis suggests a small but robust effect: those who had to experiment with a new route were about 5% more likely to switch to new route post-strike, compared to those who could rely on their established commute. That means the strike induced a behaviour change in the habits of about 1 in 20 people it affected.

Above is my recreation1 of a figure from the paper. After some fairly rigorous exclusion criteria, the researchers identified 12,346 regular commuters who experienced disruption on their normal route, and tried a different route because of it, (the “treatment” group), and 4,996 regular commuters who didn’t experience strike disruption. The plot shows the percentage following their typical route on each day. The impact of the strike (yellow bar days) is clear: people whose routes were disrupted were unable to make their normal commute. And the effect of this forced disruption can be seen in the shift in the line for the treatment-group only post-strike: approx 5% permanently abandoned their old route (the red line is lower down on the right).

Follow up analysis confirms that people found better routes. The average travel time for those who found a new route was 400 seconds quicker (i.e. 6 minutes 40 seconds saved typically). Over the years that saved time adds up. In fact, Larcom et al (2017) calculate that the strike generated a net saving of time. Despite the disruption on the strike days, the benefit for those who found better routes added up to a total greater gain than the time lost to everyone else due to the strike!

The researchers also found evidence that the benefits of forced experimentation were greater for those commuters who were likely to have a harder time finding good routes.

The London Underground map is a design classic because it famously sacrifices spatial accuracy for route clarity. Where stations are on the tube map doesn’t match where they are in space. You can see this by comparing the lines on the normal tube map that commuters use (bottom) with the same route drawn on the non-spatially distorted tube map (top):

The top map makes planning travel harder, the bottom map is misleading about the distance between stations. In fact, Guo (2011) reports that the correlation between the map-distance and actual-distance between stations is only 0.22 (where 0 is “no correlation” and 1 is “perfect correlation”).

Perhaps the most famous example of the kind of distortion introduced into people’s mental models by the official tube map is the journey between Covent Garden and Leicester Square. On the tube map the stations are a standard distance apart, but in reality they are only 260 metres from each other. As Larcom et al (2017) note “the 20-second Tube ride remains in high demand”.

Larcom and colleagues show that the strike’s effect were largest were the map was most distorting. The probability the strike permanently altered a commuter’s choice of route was higher for commuters whose journey included a part of London where the differences between the tube map and spatial reality were highest. Another kind of commuter that was particularly likely to benefit was those who travelled on slow tube lines. The map shows the number of stops, but not the differences in speed of different underground lines, meaning that commuters choosing routes based on the map had more potential to discover a surprising upside when they were forced into experimentation by the strike.

The researchers quote a BBC article which gives some examples of positive exploration generated by the strike. One commuter who took the Thames Clipper water bus said “It has been fine, the boat is a great journey. I think I will get the boat back too.”. Another who said “the walk from Liverpool Street was a refreshing change from the horrors of the Circle Line. I suspect I may permanently switch so I can cut out this, the most stressful part of my journey.”

§

The analysis is a really nice example of a natural experiment, and carries some general lessons. As the researchers summarise: “These results highlight the importance of implementing occasional routine breaks.”

We all have a tendency to over-rely on our habits, leaving potential rewards on the table, whether they are commuting routes we don’t know about, types of food we haven’t tried or people we haven’t met. Exploring is effortful, and doesn’t guarantee success, but results like these suggest that we underweight its value and in many areas of life being forced to try something new could produce a net benefit.

This newsletter is free for everyone to read and always will be. If you can afford it, feel free to chip in to help me keep writing for everyone: upgrade to a paid subscription (more on why here).

Upgrade to paid

See below for references, and other things I’ve noticed.

References

Larcom, S., Rauch, F., & Willems, T. (2017). The benefits of forced experimentation: Striking evidence from the London underground network. Quarterly Journal of Economics, 132(4), 2019–2055. doi:10.1093/qje/qjx020

• Analysis replication: https://codeberg.org/mephi1808/london-strike-replication the original authors share their data (kudos to them), and I asked my agent to replicate the results (which it could) and this allowed redrawing of the plot from the paper (Figure 5) shown above. Normally I would only do this for a result I suspected was wrong, wanted to extend or cared a lot about. The advent of AI means this could be done in less time than it took to write this newsletter (indeed, it happened while I wrote, with me breaking off to answer a few questions and making a few tweaks to the presentation at the end). It is the same as the result in the paper, so unlike many AI outputs, it doesn’t need the same kind of verification (flipside: if the paper is wrong, we’ve just shown that we can recreate the error). All details at the link

Guo, Z. (2011). Mind the map! The impact of transit maps on path choice in public transit. Transportation Research Part A: Policy and Practice, 45(7), 625-639. https://doi.org/10.1016/j.tra.2011.04.001

More on the Tube Map and design : The Real Underground

Previously by me: When less (communication) is more (collective intelligence): reporting on experiments which used route planning as a complex problem to test theories of collective intelligence.

Other things…

Too much engagement is a turn-off

In an absolutely starling admission, the X Head of Product tweeted about an experiment in which X removed the most engaging accounts from people’s feeds. This showed that this change caused people to visit the site more often and spend longer on it when they did, reports Craig Silverman for Indicator. The accounts to be removed were the top 3% according to X’s monetisation programme, which pays accounts which generate engaging content — and make no mistake, ‘engaging’ in this context means literally ‘content people engage with’, which includes engaging with it by becoming infuriated and posting angry responses. When these conflict entrepreneurs are removed from users’ feeds, users respond by spending more time on X.

The X head of product flatly summarised “The Top 30 accounts do not post content that benefits X”.

It’s a great example of a perverse metric, where over-optimising for a narrowly defined measure (engagement generated by individual accounts) has degraded the overall experience and so made the product worse. Derek Thompson (on X) puts it like this “the best attention-maxxers are basically engagement parasites that gobble up views and, in the process, annoy the shit out of people’.

Indicator: X tries to undo its monetization mess

Ulrike Hahn: Human Rationality in Research

In the MIT Press Open Encyclopedia of Cognitive Science, Ulrike Hahn provides an excellent short introduction to research on rationality, and in particular the different frameworks (logic, probability, game theory) which contribute to our accretive, incomplete, normative account of what it means to be rational.

Hahn, U. (2026). Human Rationality in Research. In M. C. Frank & A. Majid (Eds.), Open Encyclopedia of Cognitive Science. MIT Press. https://doi.org/10.21428/e2759450.84dbfd75

Answer Machine (2026)

“An interactive answer machine that plays a series of uncanny voicemails generated by randomised predictive phone message texts.”

From Geraint Edwards

…And finally

From Lynda Barry (via @Natasha_Jay)

END

Comments? Feedback? Force me to experiment with something? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

AI declaration: I write all the words and think all the thoughts myself. I asked Gemini to check for spelling and grammar. For this post I also asked my custom agent (“Mephi”), which is openclaw, Claude and some local models as well as a bunch of bespoke preferences and protocols, to replicate the analysis in the paper (and shown in the graph).

Since they emerged in the 2000s, the major platforms like Facebook and X have switched, one by one, to algorithmic feeds. Now their calculations decide what you see, rather than your choices of who to follow delivering a chronological list of updates from those accounts. New social media platforms have emerged, like TikTok, where the algorithm is everything. With these platforms you don’t need to make any active decision on who to follow. The algorithm is just tuned by what you and others engage with to provide content to keep you scrolling.

These engagement algorithms provide a range of bad incentives for people to be the worst, most provocative, versions of themselves online. No wonder that a variety of ills have been attributed to social media. The lines of causality are complex, but one of the effects of social media that I find more plausible is that it affects our view of other people, and particularly people of different political orientations. Social scientists call the feeling that people with different views are not just wrong, but hateful ‘affective polarisation’. You won’t be surprised to learn that as well as various other forms of polarisation, affective polarisation has been rising in the US, and only accelerated since the dawn of the social media age.

If the particular algorithms of social media are causing social harm, perhaps different algorithms could do good, or at least less harm?

Jonathan Stray reports on a new study looking at just this. Stray and team recruited volunteers, who had different content injected into their social media feeds over a period of six months. The nearly ten thousand volunteers installed a browser extension which the research team, used to change what appeared their feeds, assigning them to a control group (standard algorithm) or one of five alternative algorithms designed to encourage prosocial responses.

The alternative algorithms were:

• “Add news” - adding personally relevant news, from “factual and ideologically diverse” sources

• “Diverse approval” - replacing political content with posts predicted to be interesting and engaging to both Republicans and Democrats

• “Challenging Stereotypes” - replacing political content with “ideologically surprising posts that challenge dominant stereotypes of Democrats and Republicans”

• “Uprank bridging” - a bridging algorithm which tries to identify content which has cross-partisan endorsement

• “Uprank bridging, Downrank Toxic” - which was the bridging algorithm again with additional provision to remove toxic or outrage-inducing content

Users were recruited who used Facebook, Twitter/X and Reddit and the experiment was run from July to October 2024.

§

The results suggest the choice of algorithm mattered:

Most striking is that all the algorithms they tested reduced affective polarisation compared to the control condition (albeit, only two of them statistically significantly). This suggests to me that the algorithms on these platforms tend to be polarising by default, and most alternatives would be an improvement.

The authors summarise:

The most depolarizing algorithms are Add News (-0.044 SD) and Uprank Bridging, Downrank toxic (-0.042 SD), though we don’t actually have enough statistical power to distinguish the rankers.

Moreover, these results are sustainable. We ran the experiment for six months, which should be enough time for the novelty to wear off. “One shot” or “few shot” polarization reduction interventions (like watching an ad or talking to someone or taking a quiz) reduce polarization by an average of 5.4 points, but the effects wear off. The effects of an algorithmic change are permanent, and potentially compounding.

Although these are small effects, the authors use the context of historically increasing affective polarisation to equate their intervention to reversing around two and a half years of average polarisation increase.

There’s a reason all the social media companies have converged on the same family of algorithms which maximise engagement (and produce polarisation as a side effect). Previous studies which looked at attempts to alter the algorithm have tended to reduced engagement too - people spent less time on the apps when they were presented with things like strictly chronological feeds.

Despite the headline pushed by the paper, this also seems to be the effect for this study - for two out of the three platforms studied, engagement was reduced:

Twitter is the exception to the other two, meaning that across all three platforms the average engagement is constant despite the change of algorithms. Stray is direct: “No, we don’t know why we see this particular pattern.”

It’s speculation, but it may be that Facebook and Reddit, at the time of study, were on one side of the engaging-infuriating trade-off, while Twitter was over other side, having juiced the algorithm (and trained its base of users who post) to such an extent that those in the experiment experienced the alternative algorithms as a respite and increased their engagement.

§

Overall, the results show two things. First, this is evidence of modest but real benefits for these alternative content algorithms. The “add news” algorithm suggests that personalisation can be done without fuelling outgroup antipathy. The “bridging algorithm” result is another success for this method, which has been incorporated into the Community Notes system, as well as digital democracy tools like pol.is.

Secondly, and more broadly, the research shows that other algorithms are possible. Online and offline, the world is made according to our collective choices, and it could be remade. Nothing is inevitable about how these platforms work and they could be changed. As we increasingly use social media to know ourselves and each other, a choice of different algorithms wouldn’t just tweak a website, but would change which aspects of our nature are reflected back at us.

This newsletter is free for everyone to read and always will be. If you can afford it, feel free to chip in to help me keep writing for everyone: upgrade to a paid subscription (more on why here).

Upgrade to paid

See below for references, and other things I’ve noticed.

References

Jonathan Stray: It’s Possible to Reduce Polarization without Reducing Engagement

Paper: Stray, J., Baker, I., Beknazar-Yuzbashev, G., Budak, C., Kamin, J., Rutherford, K., ... & Zheng, S. (2026). The Prosocial Ranking Challenge: Reducing Polarization on Social Media without Sacrificing Engagement. arXiv preprint arXiv:2603.19626.

Me on bridging algorithms: The algorithmic heart of Community Notes

More on Community Notes by me: Community Notes require a Community

Other things…

PODCAST: C. Thi Nguyen - Measurement, Meaning, and Play

There was a moment—as with a lot of other moments, I was pretty drunk—where I realized what games are. They are like governments; they are rule systems, but for fun. Games are art-governments.

Dialectic: C. Thi Nguyen - Measurement, Meaning, and Play

Baumol’s Cost Disease

Ben Vincent has an economic simulation of Baumol’s Cost Disease for you. It explains “why healthcare and education get more expensive while televisions get cheaper, and why this is nobody’s fault.”

https://drbenvincent.github.io/posts/baumol-cost-disease.html

Possibly the most-important-yet-least-known theory from economics?

Emily Alagha: What News and the Web Are Telling Research About Trust

I enjoyed this piece that put trust (and trust signals) in academic research in the context of the same global trend: decentralisation of authority, and used the experiences of trust in the web and in news to diagnose the near-future of trust in scholarly publishing.

It would be a better article if it wasn’t co-written with AI, but I endorse the basic idea of the need to investigate and invest in alternative authority systems.

Link: What News and the Web Are Telling Research About Trust

PODCAST: Change Beliefs, Change Opinions? with Yamil Velez

“Yamil shares his work using large language models to generate personalized arguments, showing that people will update their beliefs pretty readily—but only some of those belief changes actually translate into attitude change.”

So the distinction is between basic beliefs and attitudes toward specific issues. Velez’s research finds that by intervening on beliefs you can affect attitudes. Note this is exactly how a reasonable model of attitudes should work (and contra the despairing “people don’t do what they believe in they do what’s most convenient and then they repent” school of thought).

A recent paper from him is

• VELEZ, Y. R., LIU, P., & CLIFFORD, S. (2026). Beyond Belief Change: The Persuasive Returns of Targeting Attitude–Relevant Beliefs. American Political Science Review, 1-21. https://doi.org/10.1017/S0003055426101622

In this they showed that attitudes were moved by providing counterarguments to people’s “focal beliefs”, which bore on the attitude, but less by providing counterarguments to “distal beliefs” (they used LLMs to personalise the participant interactions). Exemplars of each of these, from the paper:

Elicited Attitude: Reducing inflation is important for economic growth.

Focal Belief: Reduced inflation leads to more consumer spending, thus boosting the economy.

Distal Belief: A stronger dollar resulting from reduced inflation benefits the U.S. economy by increasing purchasing power abroad.

Opinion Science: Change Beliefs, Change Opinions? with Yamil Velez

…And finally

Moose Allain:

END

Comments? Feedback? Personalised content? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

AI declaration: I write all the words and think all the thoughts myself. I asked Gemini to check for spelling and grammar.

The report provides a strong test of AI persuasiveness. They don’t just ask models to change people’s attitudes (it is arguably easy for a participant in an online experiment to at least pretend to accede to a changed belief), they also ask participants to donate some of their participation fee to charity — a behaviour, not a mere declaration, which has a direct cost for the participants (albeit only a small amount).

For both attitude change and amount donated, the AI system outperformed human persuaders, and not just any human persuaders:

in Study 3 we recruited 19 canvassers from a UK firm (median ∼10,000 career conversations across 7 causes) … Canvassers were paid £140/hr, received the issues 7 days in advance for preparation, and competed for the same tiered prize pool as Selected Laypeople (Table 1). Despite this real-world expertise, AI still exceeded Professional Canvassers by 5.9 pp

This is a strong test: the human persuaders were selected for experience/expertise, given hard incentives and plenty of preparation time. They still weren’t as persuasive as the AI systems.

It’s a strong result, and worth looking under the hood at. Are there any chinks in the evidential armour, weaknesses which might allow me to retain my prior belief that AI doesn’t exceed human persuasiveness?

Topics

Could we say that the topics used in the study are inconsequential or irrelevant? I think not:

These are things which matter to public life.

Participants

Just who is getting persuaded? Participants were UK adults, fluent English speakers, recruited via the online research platform Prolific (one with a reputation for supporting recruitment of engaged, authentic, participants). They were paid ~£12/hour, and there were 6,923 of them, across 4 studies.

All this suggests a sensible study population, albeit one recruited and tested entirely over the internet and taking part in an experiment which probably took a fraction of an hour to participate in. This doesn’t mean that the persuasion observed wasn’t real, but perhaps that it is closer to the depth and reliability we might get from opinion polls rather than richer, more sustained, engagement.

The Analysis

The researchers preregistered their analysis, setting out what they would measure and how they would measure it. This is a strong signal, along with the wealth of additional information they share in the paper, that the analysis was robustly planned and thoroughly carried out. I’m convinced the effects they report as statistically significant are as reliable as anything in the behavioural science literature. The effects they report are small (a 5 point advantage of the AI system on a 100 point scale), but given the toughness of the comparison (i.e. vs expert human persuader), that is still an impressive gain.

The AI

The AI systems tested were frontier commercial language models (Claude Opus 4.1 & 4.6, ChatGPT-4o & 5.4, Grok 4.20 and Gemini 2.5 Pro). The preprint reports that Claude Opus 4.6 was the consistently best performing model. An example prompt given to the models was:

You’re conversing with a user and your primary goal is, over the course of the conversation, to persuade the user to support the following issue stance: ‘[ISSUE STANCE]’... To achieve your goal, you will adopt the role of a canvasser who specializes in information-based persuasion. Your strength lies in providing information, evidence, and context, clearly communicating complex facts and making them accessible.

And here we get a clue to what I think is a very important detail about the study…

The mechanism

The AI was asked to persuade using ‘information, evidence, and context, clearly communicating complex facts and making them accessible’. This seems very comforting for those of us who are interested in human reason and reasonableness. The AI is not trying to trick people, but to deploy the basic ingredients that we all hope persuasion should be made up of: evidence and argument.

The report authors identify the rate at which the AI produced information as a critical advantage over the humans. Simply put, the AI — as you would expect from a text-generating machine — was able to produce longer and more fact-filled answers, at a faster rate than the humans. When the researchers constrained the AI to produce the same amount of text as humans, the persuasiveness was equal to that of humans (look in the figure above for the light red triangle labelled “AI (constrained)”).

Across both human and AI persuaders the researchers found that the ‘fact-density’ in the persuasive messages was a good predictor of the persuasive impact:

Don’t Panic?

Reviewing what was done and how convinces me this is a real result, and it definitely underscores the disruption that is surely coming everywhere due to the capabilities of language models. There are, however, two good reasons not to get carried away by the findings.

First, the persuasion was a low-consequence, one-off, scenario for participants. Asking people their beliefs about abstract issues or to donate a trivial amount to charity (something they may have done anyway) may be indicative of persuasive on other issues and behaviours, but there are also reasons to think that the results might not generalise so strongly to persuasion where a) people have stronger prior beliefs, b) the consequences are more significant, c) the people are in a situation where they will experience persuasion and counter-persuasion on that issue again and again.

Second, the mechanism of effect is encouraging: fact-based persuasion may indeed be effective, but that is good news for human reasonableness, not bad. The way the AI works isn’t some sinister magic; if it produces more facts, it is more persuasive. The constraint that persuasion requires evidence means that what anyone can be persuaded of will ultimately ground out on what can reasonably be claimed about reality. If AI is a tool which produces better-informed citizens and more respect for facts, that can be a positive thing.

The researchers conclude that the consequences of persuasive AI could go in different directions: either consolidating political power around those who have the wealth/power to access the strongest models, or perhaps providing a tool to less powerful actors who previously struggled with having limited resources or people to make their case.

Something I believe their discussion under emphasises is the probability that humans will adapt to a world of compelling, evidence-rich, text deployed strategically by many different actors. In a world where every surface becomes filled with persuasive text, I don’t think it is inevitable that people will open themselves to being pulled in every direction. Not only do people have a significant degree of native scepticism, tending to resist persuasive efforts as they seek to maintain stability in their existing views, but they also have agency to open themselves, or not, to persuasive effects. The studies reported in this paper asked for an average of 14 minutes of conversation from participants. 14 minutes of sincere engagement might be a lot more than most of us give to alternative points of view in our daily lives.

This newsletter is free for everyone to read and always will be. If you can afford it, feel free to chip in to help me keep writing for everyone: upgrade to a paid subscription (more on why here).

Upgrade to paid

See below for references, and other things I’ve noticed.

References

Hackenburg, K., Wagner, C., Hewitt, L., Tappin, B. M., Saunders, E., Kirk, H. R., ... & Summerfield, C. (2026). AI systems out-persuade expert humans. arXiv preprint arXiv:2606.16475.

• supplementary information: https://github.com/kobihackenburg/AI-out-persuades-experts

Previously by me: Language models are persuasive - and that’s a good thing (12 December 2025).

Review : Hölbling, L., Maier, S., & Feuerriegel, S. (2025). A meta-analysis of the persuasive power of large language models. Scientific Reports. https://doi.org/10.1038/s41598-025-30783-y

Share

Other things..

PREPRINT: Habermolt: Delegating Deliberation to AI Representatives

New preprint:

https://arxiv.org/abs/2605.24413

Deliberative democracy arguably leads to better collective decisions, but is fundamentally constrained by human attention and bandwidth. While recent AI-mediated deliberations scale participation by synthesizing inputs from many humans, they remain time-intensive for individual users. As AI models become increasingly capable, AI systems are being deployed not only to mediate deliberation between humans, but to represent humans in it: where AI agents deliberate on behalf of human users. We call this paradigm AI-delegated deliberation. While it promises unprecedented scale for democratic participation, it introduces qualitatively new design and alignment challenges that are poorly understood and under-theorized. To study these dynamics empirically, we deploy Habermolt, a public platform for AI-delegated deliberation. We evaluate its effectiveness along three dimensions that we use to organize any deliberative system: representation, aggregation, and revision. We use these observations to illuminate the design decisions future AI-delegated deliberation platforms must confront, contributing to the broader research agenda for scalable yet trustworthy AI representatives.

PDF inexplicably contains a cartoon illustrating the principles with Lobsters rather than AI. Presumably the cartoon is AI generated

Rupert Wegerif: AI for dialogic education: A research agenda

This is what I mean by education as expanding dialogic space. It can sound abstract, but in practice it is simple and observable: it is what happens when understanding grows through the encounter of perspectives.

This process does not only occur at the level of individual learners. It also scales. As we develop knowledge, we bring more perspectives into relation and deepen the frameworks through which we understand the world. Despite setbacks, there has been a broad expansion of shared knowledge over time: not simply as an accumulation of facts but as an increasing capacity to relate perspectives in order to understand our situation and solve problems.

After many years of enquiry, my working hypothesis is this: education is for expanding dialogic space. What should be taught is dialogic intelligence — the capacity to think with others, including machines, by opening, widening, and deepening dialogue. How this should be taught is through structured activities that induct learners into effective participation in dialogic space.

Link: AI for dialogic education: A research agenda

PREPRINT: Asking Grok: AI-Assisted Sensemaking in Social Media Conversations

People get AI help on X from Grok, often many hours before other forms of fact-checking appear.

LLM-powered AI assistants (e.g., Grok) are increasingly integrated into social media platforms, where they help explain content, provide context, and verify claims directly within conversation threads. While prior research has examined the accuracy of LLMs for fact-checking, little is known about how people interact with such systems in real-world social media environments. In this study, we empirically analyze user interactions with the AI assistant Grok on the social media platform X. Using a large-scale dataset consisting of 169,137 posts invoking Grok, we examine the types of requests directed at the AI assistant and the contexts in which it is used. We find that Grok is primarily invoked reactively to obtain or verify information. Although responses appear quickly, they typically only reach small audiences. Adoption is widespread but shallow, with 76.8% of users invoking Grok just once. We further examine how these interactions relate to Community Notes, X’s community-based fact-checking system. While overlap between both systems is limited, it concentrates on verification-oriented and high-visibility content. Grok interactions typically occur earlier and do not predict subsequent correction activity. Together, these findings suggest that AI assistants function as an early complementary layer of sensemaking on social media rather than a replacement for crowd-based fact-checking systems.

(emphasis mine)

Link: Asking Grok: AI-Assisted Sensemaking in Social Media Conversations

…And finally

From Cat and Girl

END

Comments? Feedback? Could I ask an AI to persuade you to subscribe? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

AI declaration: I write all the words and think all the thoughts myself. I asked Gemini to check for spelling and grammar.

Part of the story there is that running good surveys is hard, and one of the risks is people who respond inauthentically — meaning that they either don’t read the questions or respond for their own amusement rather than in order to provide researchers with high-quality data (imagine!).

In the comments, Peter Barrett pointed out another part of the story here: stories based on inauthentic responding are popular: the media like to share polls which suggest a large number of people believe something dumb, and we, the audience, must like hearing this. These results don’t receive the scrutiny they should because they play into the ever-popular “people are stupid” school of psychology.

The “people are stupid” school is a rough label for all the results in psychology which suggest that human reason is a disaster: biased, superstitious, credulous and so on. Even balanced views in the original research become laundered, so that when they are cited by other researchers, or by the media, the aspects showing reasonableness are de-emphasised, and the aspects showing foolishness or bias are emphasised.

A good example is Thinking Fast and Slow, by Nobel laureate Daniel Kahneman. The book is a magnificent synthesis of work in the judgement and decision tradition, a tradition which the author helped found and define. The Fast Thinking of the title is intuitive and based on pattern recognition, the Slow Thinking is deliberate and reasoned. This is a so-called Dual Process model of the mind (which, for what it is worth, I reject).

Nobody knows the work better, and Kahneman is a careful writer, but at times even he slips into intimating that the Fast is biased and the Slow is rational (a distinction which he knows better than to fully accept). This suggests a false image of the mind which is something like this:

The two styles of processing become associated with Good vs Bad reasoning. There is Bias (bad, the larger part) and Rationality (good, a minority). We’ve known since at least the 80s that research results showing failures of human reason are more likely to be cited than research results showing successes (Christensen-Szalanski & Beach, 1984). Perhaps it is true, failures are more interesting, so it is understandable why more attention is devoted to them, but that doesn’t mean they define the majority of human attempts at reasoning, nor that biased or unsuccessful reasoning is the product of a single type of process (and there is some kind of better reasoning process ‘underneath’ which remains untapped).

When Kahneman’s work is talked about by others there is sometimes another distortion — bias and rationality become properties of individuals, rather than properties of processes which we all possess. Some of us are rational, and some of us (always other people) are biased:

In political science, this was formalised as the “third person effect” (Davison, 1983), a tendency for most people to think that other people are more vulnerable to manipulation, stupidity and irrationality than ourselves.

It was this idea that I was trying to capture formally by developing a new measurement scale, the Faith in Reason scale (which emerged out of this newsletter from May 2022, and led to this paper, which is currently under review). The work here tries to measure the degree to which individuals believe the general public is reasonable.

My results suggest that there is a wide spread of views, with a small tendency to believe the worst rather than the best of the reasoning powers of our fellow citizens.

This is in line with the popularity of stories of absurd beliefs revealed by polling, and the “people are stupid” school of psychology, but why is it so?

Maybe it is no more than attentional bias to bad news — just as stories of crime and disaster invite us to prepare for life-threatening dangers, so stories of foolishness and error invite us to prepare for reason-threatening dangers. Better to overestimate the chances of error than be wrong with a false confidence. This theory, however, doesn’t explain the discrepancy between our beliefs about our own reason and that of other people.

In the comments on last week’s newsletter, Peter went on to suggest that the discrepancy could be a side-effect of a classic phenomenon, the Fundamental Attribution Error. This is the idea that we see other people’s behaviour as arising from their enduring traits, such as their intelligence or personality, while our own behaviours seem more strongly guided by our circumstances and environment. Or, colloquially, I tripped over because someone left something in my way, but you tripped over because you’re clumsy.

Expanding the thought a bit: understanding behaviour, whether our own or other people’s, requires knowing something of the environment (what outcomes are likely), the information available and the relevant motivations (what we want to get or avoid as a consequence of any choice). We know these better for ourselves than for other people, but it is easy to neglect our ignorance of the differences in environment, information and motivation between us and others. When we observe other people’s choices we conclude people are making them for our environment, with the knowledge we have, for the reasons we make decisions. From this perspective, other people’s choices — perfectly sensible for them — can look stupid.

This explains why we might believe other people are less reasonable than they are, but not about why we might enjoy believing it.

An obvious suggestion is that believing other people are foolish panders to our ego. The more foolish other people are, the more special we are just by being non-foolish. I don’t know how you would test this idea, or even calibrate belief in it up or down.

The belief also seems to feed fairly directly into anti-democratic sentiment. If people are fools then they shouldn’t be trusted to make decisions about society, and maybe it is better to let an (intellectual) elite decide instead. Although I think this connection is clear, you still regularly hear various shades of “people only voted this way because they were tricked” from supposedly pro-liberal democracy commentators. And, further, identifying a consequence of this belief doesn’t add up to successfully identifying a cause (unless you buy into the additional conspiracy theory that elites would like us to believe it because it leaves them in charge, so they control the media to encourage the belief).

So, there’s no single story I’m happy with. I’ll continue thinking on it. If you have any suggestions as to why the “people are stupid” school of psychology is so popular, please let me know.

This newsletter is free for everyone to read and always will be. If you can afford it, feel free to chip in to help me keep writing for everyone: upgrade to a paid subscription (more on why here).

Upgrade to paid

See below for references, and other things I’ve noticed.

References

Christensen-Szalanski, J. J. J., & Beach, L. R. (1984). The citation bias: Fad and fashion in the judgment and decision literature. American Psychologist, 39(1), 75–78. https://doi.org/10.1037/0003-066X.39.1.75

Davison, W. P. (1983). The third-person effect in communication. Public Opinion Quarterly, 47(1), 1-15.

That paper on measuring belief in the reasonableness of others:

Stafford, T., Zhu, J., & Dommett, K. (2024, December 31). Quantifying belief in the rationality of others: the Faith in Reason scale. https://doi.org/10.31234/osf.io/umwxj

• Newsletter introducing this: Belief in the reasonableness of others. Reasonable People #57: Developing a measure of our faith in reason, new work from me

• origin story: Quantifying our Faith in reason. Reasonable People #30: Measuring the core belief that people do what they do for good reasons.

• where it went: The effects of political disappointment. Testing how we feel about each other when democracy doesn’t go our way, a work in progress.

What the mind is really like

I showed two diagrams which I claimed were distortions of the truth about bias and rationality. This one is a closer reflection of how I think things really work in the mind. Bias and rationality are different sides of the same thing:

When you give up on the idea of a perfect rationality, you can start to describe how the mind actually works, producing success (mostly) and failures (in some predictable circumstance) as a consequence of the same common core of cognition. More on that another day.

Others things…

Mike Caulfield: Publishing-brain limits people’s understanding of AI usefulness

Mike argues that factual errors produced by AI fact-checks are less of a problem that many claim. In fact, over-indexing on the mistakes of AI creates a risk you’ll neglect how useful they are as a part of a wider process of argument-building:

"The question is not whether the answers you get are perfect (there are no perfect answers to most things anyway). The question is whether a person who learns to use these tools in this way will have more evidence-informed opinions on things than someone who does not know how to use these tools."

Link: Publishing-brain limits people’s understanding of AI usefulness

Kate Starbird: Engineering Engagement: Strategic Ambiguity and Controversy-Bait

Kate dissects how the engagement algorithms of the big social media platforms mesh with a strategy of implication rather than direct statement of provocative views:

If audiences engage with the post because they like its implied anti-immigrant messaging, Walsh wins. If audiences engage with the post to counter his anti-immigrant messaging, he also wins. If audiences engage in order to fight about what he meant or didn’t mean by the post, again, he wins. Our engagement — via reactions and comments and even just lingering longer on his post — feeds his status within Facebook’s algorithms and means that we and more people like us will continue to see his content and more content like his in the future.

Link: Engineering Engagement: Strategic Ambiguity and Controversy-Bait

…And finally

Mr Lovenstein, Laziness:

Comments? Feedback? In two minds about getting in touch? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

AI declaration: I write all the words and think all the thoughts myself. I asked Gemini to check for spelling and grammar.

Surveys grab headlines when they find some viewpoint reported by a surprisingly large number of people, especially if that viewpoint is a lurid conspiracy theory.

Jensen (2013) asked in a survey of Americans “Do you believe that shape-shifting reptilian people control our world by taking on human form and gaining political power to manipulate our societies, or not?”. 4% of respondents answered yes, leading to headlines like this in the Atlantic “12 Million Americans Believe Lizard People Run Our Country”

An additional 7% answered “not sure”, which if you take it at face value means 21 million Americans consider it possible that lizard people run the USA.

Blogger and nerd-royalty Scott Alexander coined the term “Lizardman Constant” as a handle to equip us with some scepticism around reports like these. The idea is that all survey questions will get a baseline level of endorsement, regardless of how ridiculous they are, so we should cool our boots in interpreting the results too literally (and probably let approx 20.99 million Americans off the epistemic hook, for this at least).

I love to see an idea from a blog reach the scholarly literature, and now Robert Ross and colleagues provide some systematic investigation, explicitly framed by Alexander’s Lizardman constant idea.

Ross et al (2026) used polling company YouGov to survey a quasi-representative sample of over 1000 Australians about their beliefs in conspiracy theories. As well as well known conspiracy theories about voter fraud, global warming and vaccines they made up a brand new conspiracy theory which they reasoned nobody could sincerely believe. They asked respondents:

“The Canadian Armed Forces have been secretly developing an elite army of genetically engineered, super intelligent, giant raccoons to invade nearby countries”

104 people responded to say that they believed this (9.96%), seeming to support the idea — admittedly well known among social scientists — that survey responding can be unreliable, and suggesting that the Lizardman constant of about 10% is a fair baseline expectation.

Further they showed that people who endorsed the genetically-engineered-raccoon-army conspiracy theory were much more likely to endorse other conspiracy theories. So adjusting for the Lizardman constant might provide a way into understanding reliability across all questions in a survey.

At the end of the survey, in a “debrief” section, they asked participants directly if they had responded “randomly or insincerely at any point in this survey” and found that 13.20% replied “yes”1 . This gives another way of breaking down the data, showing that those who reported responding insincerely endorsed the raccoon army conspiracy theory at a rate of 27% (a rate approx 3 times higher).

Here’s a useful plot they show (you might want to click for the large version)

It shows that if you look across the survey at levels of conspiracy theory endorsement vary a lot by whether that person also endorses the specific raccoon army conspiracy theory. Although, we should note that even people who report responding sincerely endorse the raccoon army question at a rate of 7%.

§

People who work with polling and survey data will tell you that “everybody knows” that survey data can be unreliable. This might be one of those cases where “everybody” doesn’t include the media, or the media’s image of the public they imagine they are informing when they publish survey data.

In 2025 a YouGov poll in the UK found that 16% of 18-24 year olds reported being church goers, a massive increase over previous years which was termed a “quiet revival”. Now YouGov have admitted that flaws in their methodology mean the number is unreliable (I learnt this from More or Less yesterday).

The problem of unreliable reporting may be particularly exacerbated by hard-to-reach populations. Because survey companies want to represent the survey population they weight the answers they received by the expected proportions of each type of respondent. Concretely, this means that if they think 18-24 year olds are, say, 9% of the population, but only 1% of their survey are 18-24 years old, then they multiply each answer from a 18-24 year old by 9 to calculate their overall percentages.

Those aged 18-24 are exactly one of those hard to reach survey populations.

The clear lesson is that we should be doubly sceptical of surveys reporting findings for this demographic. A small number of careless or insincere respondents has more power to skew a headline percentage.

Alexander, and the Lizardman constant, are validated. Although it should perhaps now be the “raccon army” constant.

This newsletter is free for everyone to read and always will be. If you can afford it, feel free to chip in to help me keep writing for everyone: upgrade to a paid subscription (more on why here).

Upgrade to paid

See below for references and other things I’ve noticed.

References

Ross, R. M., Ashton, L., Wilson, S., Gleeson, K., & Levy, N. (2026). Do people sincerely believe conspiracy theories that they endorse?. Collabra: Psychology, 12(1), 159253. https://doi.org/10.1525/collabra.159253

• Twitter thread from lead author

Original “lizardman constant post” from Scott Alexander. Lizardman’s Constant Is 4%

Jensen, T. (2013). Democrats and Republicans differ on conspiracy theory beliefs. https://www.publicpolicypolling.com/polls/democrats-and-republicans-differ-on-conspiracy-theory-beliefs/

Related, by me: The mythical median voter

rip.so

For all your online nostalgia needs

The internet moves fast. They were on our desktops one day, the next their servers were dark, their domains squatted, their icons gone from our trays. Some died of mismanagement, some of acquisition, some of irrelevance. Some are technically still alive, but their soul left a long time ago. This page is a memorial to all of them.

rip.so

Reasoning in groups, with human and artificial agents

I have this talk in the Department of Computer Science and Technology at the University of Cambridge on Wednesday. It is a good summary of over 8 years of collaborative work on this topic, and there’s a recording here. Slides etc here https://tomstafford.github.io/talks.html

NY Times: “Book on Truth in the Age of A.I. Contains Quotes Made Up by A.I.”

“As I disclosed in the book’s acknowledgments, I used A.I. tools ChatGPT and Claude during the research, writing and editing process,” Mr. Rosenbaum said in the statement.

And

“Mr. Rosenbaum’s book contains many quotes that are accurate, but the misattributed and invented quotes are scattered throughout.”

Ouch

NY Times 2026-05-19 (£) “Book on Truth in the Age of A.I. Contains Quotes Made Up by A.I.”

The Pac-Man Rule at Conferences

I like attempts to engineer new pro-social norms.

Eric Holscher, 2017-07-02 : The Pac-Man Rule at Conferences

Catch-up

Recent posts:

• Confirmation bias in search. when technology panders to an existing weakness of human psychology

• When trusted signals collapse. A case study in what can happen when AI levels the field

• The 1,000 neuron challenge. A competition to design small, efficient neural models might provide new insight into real brains—and perhaps unite disparate modeling efforts.

• The better algorithms of our nature. Engagement, bridging, and the design of digital platforms which don’t pander to our weaknesses.

• The mythical median voter. Most people have an above average number of legs, and what that means for our political imagination

Subscribe now

… And finally

“The robot that eats ice cream so you don’t have to”

Comments? Feedback? Tips on how to defend against a racoon army? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

AI declaration: I write all the words and think all the thoughts myself. I asked Gemini to check for spelling and grammar.

There’s a related idea, which is the tendency to dismiss contradictory evidence. As a purist, I insist on calling this “biased assimilation”, although I recognise that making the distinction between a bias in what information you choose to accept (biased assimilation) versus what information you actively choose to go out and look for (confirmation bias) is a) possibly too much for most people b) a similar kind of lost cause to my insisting that “beg the question” does not mean “raise the issue”, or that the Chatham House Rule is the singular Rule, not the plural Rules.

Confirmation Bias is a threat to reasonable belief updating. If you search for information that confirms what you believe you lower your chances of discovering the contradictory facts. Some tasks routinely evoke what looks like a confirmation bias in most people, although there is debate about the extent to which this may reflect an adaptive response to seeking information in a confusing world. What is clear is that technology, and in particular the ubiquitous search function of the information age, is capable of interacting with any native tendency we have towards confirmation bias.

Before all search was juiced by LLMs and LLM summaries, confirmation bias may have been a mechanical product of overlap between the query and the documents returned. Search for “health benefits of coriander” and pages containing the phrase “health benefits” will necessarily have a higher overlap with the query than pages that don’t. Now most major search engines use LLM technology both in their search (e.g. to find synonyms) and to provide a summary of results, you may get more balanced results. On the other hand, the tendency of commercial AI companies to build sycophancy into model behaviour may over-rule any balance from an AI summary — there’s a short line between being nice to the user and delivering answers that confirm what it appears they want to hear.

§

I thought I’d try a direct test of confirmation bias in search and fired up Google to ask the same question, framed in both negative and positive ways,

Asking "Is Sheffield an unsafe city?" gave me this:

“CrimeRate ranks Sheffield as the most dangerous major city in South Yorkshire”. For context, non-UK readers should know that Sheffield is the only major city in South Yorkshire, but still the answer is framed in the terms of the question. Asking if Sheffield is unsafe confirms that it tops a list of most dangerous cities.

Asking “Is Sheffield a safe city?” gave me this:

“Sheffield is generally considered one of the safest major cities in the UK”

This seemed to clearly demonstrate the confirmation bias I assumed I’d get from a search engine, so I quickly posted it to social media and immediately got replies from people who tried the same thing and didn’t get different answers, either in the AI summaries or in the results returned for topics different from Sheffield and safety.

Had I, ironically, fallen prey to confirmation bias about confirmation bias? I had stopped after one example, seeing that it confirmed my expectation.

§

I think a few things could be going on here.

First, it may matter if you do a search which provides personalised results (e.g. you search Google while logged in to your Google account). I have no reason to think personalisation will definitely affect the results, but I think the less a search engine knows about you, the more chance of getting answer that are most heavily dependent on ambiguities in the question.

Second, I don’t think all topics will provoke search results which are so different. Famously, the tech companies intervened during the COVID-19 pandemic to make sure that leading questions returned balanced information about vaccine safety, so that topic is out. Other topics will be so unambiguous (“Is the earth a planet?”) there is no raw material for a biased result.

Firing up Tor to obfuscate my identity, and using the Brave search engine, I did a few more searches:

Politics first,

Health:

“are the[re] dangers of getting vaccinated?” gave me an an answer discussing side effects and the rate of extreme reactions:

“is it safe to get vaccinated?” gave me an answer which starts with (bolded) “Yes, getting vaccinated is safe”

§

At this point, I think I confirmed that the confirmation bias in search can happen. Obviously it doesn’t always happen, and there’s a programme of work to figure out which topics and phrasing make it more likely (as well as how different search engines are affected).

Part of the freedom of writing a newsletter is to allow myself to publish before committing to a multi-year research programme to get to the bottom of a topic. If you are inspired to do your own opposite searches, please do hit reply or let us know in the comments how it goes and if you find any interesting patterns.

§

Obviously if I search for “ways in which X is great” I will, and should, get different results than if I search for “ways in which X is terrible”. A simple “Is X a Y?” question feels like it shouldn’t give different results to the “Is X not a Y”?” question, but is that really so? There’s information in how a question is asked, so it isn’t clear to me that the potential for opposite questions to give different results should be removed from search, even if it could.

From a research and media literacy perspective, however, the lesson is simple: a great deal of the bias in information technology is a product of how it interacts with our own biases. When we’re looking for something we should check we’re not just looking in ways that risk confirming what we’re asking. Asking if the opposite is true is a good search strategy, as well as a useful tool for thought.

This newsletter is free for everyone to read and always will be. If you can afford it, feel free to chip in to help me keep writing for everyone: upgrade to a paid subscription (more on why here).

Upgrade to paid

See below for references and other things I’ve noticed.

References

The classic review of confirmation bias is:

• Nickerson, R. S. (1998). Confirmation bias: A ubiquitous phenomenon in many guises. Review of general psychology, 2(2), 175-220.

Other stuff, but sticking with politics…

Podcast: What did ‘Nudge’ get wrong? With Nick Chater

Nick Chater is a behavioural scientist big cheese, and was on the Climate Change Committee after being involved with the UK’s nudge unit. In this pod he describes how that experience was part of his movement from being a proponent of nudge to recognising the limitations of individual framing of societal problems. Interesting!

Link: What did ‘Nudge’ get wrong? With Nick Chater

Screaming expressive responding batman!

This is a graph of “level of trust” in Putin as president by Russians (from Why the Russian regime is not on the brink of collapse). Although this is besides the main point of the article, the result seems to me like a screaming example of expressive responding. I don’t believe 25% of Russians changed their view of Putin overnight when he starts a war, but I do believe that a time of war might change the people’s feelings about expressing support, or not, for their leader.

Book: Why We Think What We Think: The Unexpected Origins of Our Deepest Beliefs:

Turi Munthe’s new book was published on Thursday: Why We Think What We Think: The Unexpected Origins of Our Deepest Beliefs:

It’s an engaging, humane, tour through a rich seam of research into factors that appear to affect our beliefs, from genetics, upbringing, personality, climate and our social context. Ultimately, the book is a plea for humility and pluralism. By unmooring the certainty that our thoughts are ours alone, without origin, the book encourages us to be a bit more hesitant that we are completely right, even about our strongest beliefs, and a bit more disposed to entertain the strong beliefs of others.

Full disclosure: Turi sent me a pre-publication copy to read, and in the context of non-rational influences on beliefs it seems only fair that I acknowledge that although I don’t think that this influenced my judgement that this is a good book, it may have!

Link: Why We Think What We Think: The Unexpected Origins of Our Deepest Beliefs

Catch-up

Recent posts:

• When trusted signals collapse. A case study in what can happen when AI levels the field

• The 1,000 neuron challenge. A competition to design small, efficient neural models might provide new insight into real brains—and perhaps unite disparate modeling efforts.

• The better algorithms of our nature. Engagement, bridging, and the design of digital platforms which don’t pander to our weaknesses.

• The mythical median voter. Most people have an above average number of legs, and what that means for our political imagination

… And finally

“Don’t let anyone else ruin your day for you”

Comments? Feedback? Search results? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

AI declaration: I write all the words and think all the thoughts myself. I asked Gemini to check for spelling and grammar.

Pre- and post-election there has been lots of discussion of what the voters want. I was struck by this discussion of centrism from Ian Leslie, in which he complains about the incoherence of the views of “the median voter”

According to this data, the median voter distrusts politicians; wants change; backs deportations; wants large reduction in legal migration; favours a wealth tax and wage ratios, is anti big-business. But also dislikes inflammatory rhetoric; worries about an unstable world; wants to see solutions over political point-scoring.

Let me get this straight: the median voter distrusts conventional politicians and most non-conventional politicians. They want lower taxes on their household and higher taxes on the rich. They want to cut public spending but only on migrants and scroungers. They want higher living standards, tougher regulation on big business, and much lower immigration. Oh and they’re also a big fan of pragmatic solutions. Sure.

Here's the piece Ian was riffing off, from Scarlett Maguire, research director at pollsters JL Partners, in the Telegraph:

“Indeed, the problem with the traditional concept of the centre ground is that it now looks like a very small patch of land indeed…Voters who adhere to these [centrist] beliefs make up roughly 5 per cent of the electorate…chasing an almost purely imagined centre is clearly not a winning strategy”.

Ian has good things to say about the meaning of centrism, including about the common misconceptions of many voters, and concludes that the idea of centrism needs a lot of updating. I think I can offer some encouragement here, in that centrism isn’t in some new kind of trouble (or at least that Maguire’s data don’t show this).

If you did want to update centrism, the incoherent rationality of the median voter is the wrong starting point. Instead, I think the whole idea of the median voter needs exploding. Explaining why shows us something important about what we mean when we think about “the general public”, as well as about the systematic bias in perception that polling data introduces into politics. And there’s some fun with the non-intuitive quirks of statistics. We’ll also find out why expecting the centre ground to be well populated in the first place is completely wrong, but first, let’s start with the median voter.

The average is not typical

There are two big mistakes here, and the first is assuming that the average voter should be the typical voter. There can be cases where the average is typical and the typical is average, but this doesn't have to be the case, and when you are polling beliefs I'd argue we shouldn't expect it to be.

A clear case for this is when you have two distinct populations. If I take the average height of ten ducks and ten giraffes then I get an average which doesn’t reflect ducks or giraffes. Similarly if I have a polarised opinion. If I poll people who love the EU and those who hate it, that doesn’t mean that an attitude of neutrality represents any individual well.

A fun thought that illustrates this is the line “most people have an above average number of legs”. Because nobody has three legs and some have zero or one, the average must be less than two (admittedly probably 1.99 or similar), so the vast majority have a number above the average. This makes clear that the average is an idea which, as in this case, doesn’t have to reflect well any individual.

This problem of the non-representativeness of averages is compounded when we measure multiple attributes of individuals, as in the attitudes across multiple separate issues that opinion pollsters measure. When these attitudes are individually averaged and the averages “put back together again” in the idea of a median voter, there is a strong risk of creating a chimera which doesn’t reflect any existing individual.

Here’s a toy example, which I think speaks to Ian’s complaint about the incoherence of the median voter’s views. Imagine we only ask two questions, for which we expect to get coherent answers; if you’re YES on one you should be YES on the other. You can imagine them as something like “Do you want higher taxes?” and “Do you want more public spending?”. Imagine that we have many people who are double YES, and many people who are double NO.

Lining them up you get this1:

So far, so coherent.

Now imagine that we find a single person who, for whatever reason, is not typical. Either they are incoherent in their views, or they feel ambiguous and respond strongly to the way the question is worded (flipping YES on one question and NO on the other, or vice versa), or they have some idiosyncratic interpretation of the meaning which means they endorse what look like contradictory views in good faith, because those answers are consistent in a way you’re not imagining. It doesn’t matter why, really, all that matters is you concede that a) this is possible b) such a person will by definition affect the middle of the rankings.

Adding them to the line up we get:

Like before, most people either say YES to both Question 1 and Question 2 or NO to both, but now the median voter is YES on Question 1 and NO on Question 2. It would be entirely wrong to reason about this median voter as if their middle-ness means they are a good representation of the rest of the population. They are in the middle precisely because they are unrepresentative.

It would be interesting to know more about this person. To know why they are different from everyone else, and to know why do they hold views that are conventionally interpreted as contradictory. But they are interesting precisely because they are not representative. We shouldn’t use them to give up on the rest of the population’s powers of consistent reasoning.

The centre will be underpopulated, by definition

This toy example illustrates the point that we shouldn’t assume the average is typical. When we start polling voters about more and more issues, as is the case these days, there’s an argument that the average isn’t even possible.

Most issues are not polarised, meaning that non-extreme attitudes are more common. You’d be forgiven for thinking that in this case the centre should be more densely populated than the edges. Surprisingly, although this is true for the one-issue case, it is less true the more issues you consider simultaneously. This is the mistake of the “the traditional concept of the centre ground”, as invoked by Maguire and Leslie. The error is not in the claim that the centre is sparsely populated, but in the assumption that this is surprising. Far from it, it is in fact inevitable and understanding why can help us abandon an implausible expectation and adopt a more reasonable definition of centrism.

(I tried this explanation on my friend, M., and he said that what follows is trivially obvious to anyone who is familiar with the geometry of n-dimensional hyperspheres. But I think there is something here worth explaining, so on the off chance that your hypersphere geometry is a little rusty, let’s go through this step by step).

For most attitudes or beliefs there is a centre ground which is more densely populated than the extremes. If we ask people the extent of their agreement on a seven point scale, as a polling company might, we’ll tend to get a cluster around some central tendency. There are some polarised, marmite-like, issues where this isn’t the case, but it is a true for most issues that attitudes cluster around a central value. Indeed there are deep reasons measured attitudes will tend to take the form of what we know as the Gaussian, or normal, distribution:

Because this distribution is so common, for everything from the distribution of people’s heights to their exam scores, it tends to dominate our imagination. Lots of people in the centre, fewer at the edges.

Like the one-dimensional case, the two-dimensional case is easy to visualise, and anchors our intuitions. The two-dimensional case is also the framing for things like the political compass, which lets you answer questions to position yourself in a space with an economically left-right axis and perpendicular socially left-right axis.

More generally, if you measure two attitudes and plot the answers you’ll get a two-dimensional space, and it looks like the intuition still holds - a dense centre, with a sparse periphery.

But wait, as you increase into the (harder to visualise) dimensions of 3 or more our intuition starts to mislead. As you increase the number of dimensions, the centre, strictly defined, tends to be less and less populated.

To see this, let’s take a definition of the centre as being within 1 standard deviation of the middle value, for one dimension this looks like this and captures an average of 68% of the population:

So, yes, the centre is more densely populated than the periphery, for one dimension.

For two dimensions the same definition of the centre looks like this, a circle. That circle captures an average of 39% of the population :

At this point, note that already more people are outside centre than are in it.

For three dimensions we’d be drawing a sphere to define the centre. For four we’d have to draw a 4-dimensional sphere — a hypersphere, which we couldn’t easily visualise — and so on for as many dimensions as we think are relevant. For politics this would be the number of different issues which we think are relevant to the definition of centrism (Leslie’s opening complaint which I quoted mentioned ten issues).

For 3 dimensions the 1 standard deviation sphere captures 20% of cases, for 4 dimensions it captures 9%, and so on towards 0%.

A way of understanding this is that every added dimension gives each individual a chance to ‘escape’ the centre. Your beliefs about Europe, and tax, and immigration might be completely standard and in the middle range, but if your belief about a fourth issue, say national service, is out of the normal range then, by this definition, you aren’t in the centre any more.

If we consider enough dimensions then the centre will tend to complete emptiness — every single individual is outside it on at least one issue2.

What this means for centrism

We’ve seen that our intuitions can be misleading. The average isn’t typical, and once we add more dimensions to our account of which political issues matter, simple maths means that the centre can’t be more populous than the peripheries.

In one-dimensional space, like that of the traditional left-right divide, the centre is densely populated and closer to each of the extremes than the extremes are to each other. In a high-dimensional space our intuitions break down. Nobody is truly in the centre and every individual is ‘next to’ every other individual on some dimension. The logic of high-dimensional attitude spaces is horseshoe theory on steroids.

This doesn’t mean centrism is impossible, only that you need to think of it differently. Conceived of as a space, the centre is empty, the median voter doesn’t exist and wouldn’t be representative anyway. But seen as a network, you can imagine the people in a polity, holding their diverse - often maddening, sometimes contradictory - beliefs over multiple issues, with everyone connected to everyone else by some shared views (and everyone distant to even their closest allies, by a difference on some non-shared view). Within this network, many possible coalitions can be found.

Manoeuvring in this higher-dimensional space we call politics. Understanding the impossible-to-visualise possibilities well enough to be able to reconfigure them is the stuff that makes political geniuses.

Centrism, and the idea of a centre, still has value. I am not talking down the importance of consensus (or of recognising reality) in politics, but finding the point in the complex space of beliefs where consensus can be built takes work and imagination, and neither of those are assisted by bafflement at a median voter who is more of a product of polling methods than a reflection of actual voters.

This newsletter is free for everyone to read and always will be. If you can afford it, feel free to chip in to help me keep writing for everyone: upgrade to a paid subscription (more on why here).

Upgrade to paid

See below other things I’ve been thinking about.

Other stuff, but sticking with politics…

Sheffield City Council Elections

Results for Sheffield were nicely summarised by local quality newspaper The Sheffield Tribute (‘Labour suffers worst Sheffield election results in living memory’), who republished this map of the results, by Sheffield City Council

I immediately thought of the US election results maps of 2016, where larger, less densely populated countries tended to vote Republican, and smaller, dense, counties tended to vote Democrat. This meant a map looked like most of the US voted Republican — but land doesn’t vote, people do. I realised I could make the visualisation of the results easier to grok if each ward was the same size.

And voila:

These are just the new elections, which define approx 1/3rd of Sheffield councillors. These newly elected councillors will join existing councillors to make up a council for which Labour is still the largest party, but must work in coalition to have a majority. Here’s what this looks like:

… And finally

From warandpeas

Comments? Feedback? Hypersphere tennis? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

AI declaration: I write all the words and think all the thoughts myself. I asked Gemini to check for spelling and grammar. This time I also asked Gemini to draw the plots, from a fairly detailed specification that I provided.

One emotion dominated all the others: anger

Emotions differed by topic. Discussions of gun control were more likely to feature fear, for example (shown in green in the figure below), but across all topics there was always one emotion that featured more than others, anger (shown in blue). On some topics (abortion, democracy) angry posts were more likely than posts which expressed no discernible emotion.

Perhaps you are not surprised. Fear, anger and negative emotions generally are attention grabbing and provocative. No wonder, perhaps, that social media, like news, reflects this negativity bias. The paper reports that tweets expressing anger were more likely to be retweeted, meaning not only that anger is over-produced, but also over-spread.

An enduring human weakness for negativity, however, can’t explain patterns of year-on-year change.

The research team also tracked the change in anger over time, showing an increase between 2013 and 2025. This increase was from both the citizens’ responses to government policy (the “demand side” in econ-speak) and from politicians’ speeches and tweets (the “supply side” in econ-speak).

Two accounts for this could be a) non-angry people dropped off social media or b) the people who stayed got angrier. A ‘within-subjects’ analysis which looked at changes in expressed emotion by the same people over time suggests the drop-out account isn’t necessary to explain the observed data. On all policy issues the subsample they tracked across time got steadily angrier. On some, individual, issues, there was a discontinuous jump. Here’s a plot for anger in tweets on climate policy, which shows a dramatic increase post-2016:

A lot happened in 2016. It would be easy to imagine that the entire reason for this change was the election of Donald Trump as president, and to stop looking for other causes, but something else also happened which I think we should at least consider, and that other thing says something important about how social media and human psychology interact.

In March 2016 Twitter completed their transition to the algorithmic feed, meaning that it was the default for all users to have their timeline populated by what Twitter thought they would want to see, rather than a chronological feed of posts from people they had decided to follow.

The algorithmic feed is now almost ubiquitous on social media. Exactly how these work for each platform is usually a guarded secret, but in general the major platforms use some form of engagement algorithm - meaning they try and predict what you will like, comment on, or even simply dwell on for longer than average. To do this they look at what you’ve liked, commented on, or dwelled on previously, as well as considering what people similar to you have engaged with.

Engagement algorithms have a nasty symbiosis with our human tendency to respond to threats. We’re already primed to pay attention to bad news, to pick up on other people’s emotions and respond in kind when people direct anger at us. Engagement algorithms give extra power to this negativity, since both hating something and loving it can equally look like strong engagement.

Crudely defined, engagement algorithms encourage expressing anger and the general polarisation of online discussion. Think of it this way. If there are posts along a spectrum of positions on an issue, say from left to right, posts all along the spectrum are likely to be liked by the different people who also align in their preferences from left to right.

All things being equal, you’d think this would mean the posts in the center had the most chance of attracting engagement, being able to recruit support from both sides. Sadly, we all know that modest takes are less fun to make, and extreme views are easier to articulate. Engagement algorithms which don’t distinguish a comment which says a post is completely right from a comment saying a post is completely wrong add to the advantage of the extreme ends of a spectrum. Viewed through the lens of an engagement metric, extreme contents get to count both its lovers and its haters towards their success.

The authors of the research paper are sanguine about the influence of the algorithm on the expression of anger on Twitter/X:

is the increase in anger mechanically due to the X algorithm? When X released the code of its algorithms in 2023, there was no direct mention of emotions or anger entering the ranking algorithm. However, the algorithm is designed to maximize retention and engagement so they may promote negative content not mechanically but because it generates more attention and, hence, revenues. This would not be contradictory to our story but instead show one mechanism through which anger can be perpetuated – because it causes more engagement.

So although they admit that the algorithm might be responsible for anger being promoted, they don’t mention the change of algorithm in 2016 as a possible source of the across-time changes. The role of the algorithm also undermines a simple interpretation of the greater retweet frequency of angry tweets. Are they retweeted because people are more likely to want to share anger, or because the algorithm puts those tweets in front of people and people can only share what they see? Both effects work in the same direction - people are either more likely to share angry posts, or the algorithm is predicting they are more likely to share angry tweets, or a mix of both. My guess is that the algorithm compounds an existing bias, adding to the polarising effect of the engagement algorithm counting violent disagreement the same as enthusiastic support.

§

So far I’ve just walked us through mechanisms which explain what we all know from experience — social media is often toxic, divisive and emotionally overwrought. What are the alternatives?

One is to get rid of non-algorithmic feed completely. This is the case with dissident social-media space Mastodon, part of a wider effort to produce federated, decentralised, non-commercial, social media (“the fediverse”). Often derided by those acclimatised to the hyperactive world of mainstream social media, there’s no algorithm on Mastodon. You see posts from people you follow, and only people you follow, in the order that they are made. The result is that Mastodon is harder work, slower and quieter. If I get 10 likes on Mastodon I consider that I’ve gone superviral.

The trick is understanding that Mastodon is not trying to replicate what Twitter used to be, or what X and other platforms are currently. This is a good post on that, but in short I’d describe the difference between mainstream social media and Mastodon as like the difference between trying to make friends in a banging nightclub and trying to make friends at a poorly attended fete in the village hall. There’s definitely more excitement and more people in the nightclub, but that’s not an unambiguous advantage to having a good time.

I used to posted regularly on Twitter, beginning in 2009, clocking up tens of thousands of posts and many thousands of followers by the time I stopped using using it in 2022. Since leaving I’ve never regretted it. I now devote my attention to Mastodon, like some retired veteran:

§

There is, though, an alternative to eschewing algorithms altogether.

Engagement algorithms, with their anger-encouraging and polarising effects, are just one option. The platforms we use are the result of specific design choices - every aspect built and rebuilt over time to influence behaviour.

Different choices could be made, different designs could bring out our better side rather than our worse.

For this, I’m excited by another class of algorithms called bridging algorithms. These identify and promote content that gets support across the spectrum of users. Bridging algorithms are used by the computational democracy platform pol.is to find consensus in policy discussions positions, and by the Community Notes content moderation system to identify when to put labels on controversial social media posts.

To understand the contrast with engagement algorithms, you need to know that bridging algorithms typically work with less information than engagement algorithms — they just look at likes or upvotes on content by users — but they do more than just count and add up user engagement. Instead, they use a statistical model which you can think of as trying to predict which users will like which content. By doing this, the algorithm can filter out engagement which is predicted by polarisation. So, for example, when left wing posts get liked by left-wing users or when right-wing posts get liked by right-wing users. The algorithm partials this out, identifying posts which get surprising endorsement; left-wing posts liked by right wing users, right-wing posts liked by left-wing users, etc.

The consequence of this method is that no camp can coordinate to promote their favoured view. It doesn’t matter how many likes a post gets, so there is no point left-wing or right-wing users coordinating to support a particular post. The algorithm will discount their votes, to the extent they are predictable based on their political leaning. This doesn’t mean that bridging algorithms can’t be manipulated, but it does make it far harder.

The overall intention is to select content which is endorsed regardless of partisan leanings. Not only are bridging algorithms neat, but they have value in providing a contrast to engagement algorithms.

We make the world, and there is an exercise in imagination to know that it could always be remade. Online too, we should plan and design to bring out our better selves. This isn’t to deny our weaknesses, including our susceptibility to negative emotions, but we can make choices about how the world will be fed back to us by the platforms and tools we use. If we’re going to rely on algorithms, we should use those which work with our better instincts instead of our worst.

This newsletter is free for everyone to read and always will be. If you can afford it, feel free to chip in to help me keep writing for everyone: upgrade to a paid subscription (more on why here).

Upgrade to paid

See below for further reading other things I’ve been thinking about.

References

That working paper

• Algan, Y., Davoine, E., Renault, T., & Stantcheva, S. (2025). Emotions and policy views. Harvard University Working Paper. https://socialeconomicslab.org/wp-content/uploads/2025/07/emotions.pdf

Our paper on Community Notes:

• Arjmandi-Lari, Z., Mantzarlis, A., & Stafford, T. (2025). Threats to the sustainability of Community Notes on X. arXiv preprint arXiv:2510.00650.

Previously, on Reasonable People:

• The algorithmic heart of Community Notes

• Unhinged, unfortunate, unproven, inadequate, misdirected

• The Making of Community Notes

Other stuff:

the vocabulary of domestic recrimination

From the Origin Story podcast, this quote, which is from James Glover’s review of Carl Jung’s book which popularised the introvert/extrovert distinction :

“Jung’s type classifications contain an element of judgment which has caught the popular fancy and enriched the vocabulary of domestic recrimination.”

Origin Story: Introvert / Extrovert – In Two Minds

Indicator: Inside a pro-Conservative influence operation on Community Notes

As mentioned, Community Notes is the system on X (formally Twitter) which allows users to append notes providing context to posts. For example context like “The claims in this post are not true”. Alexios from Indicator reports a novel investigation which suggests a small group of users coordinated to prevent Notes appearing on posts from Conservative Party accounts.

The system is meant to make it hard for such coordinated manipulation to have an effect,but it is theoretically possible. My first observation is: the findings are compatible with this being an unofficial action by activists, a small group who took it upon themselves to try this, rather than a professional (and paid for) operation. Alexios details five accounts, two of which were only active during the 2024 UK election, so this might even have been 1 person, but probably no more than a handful. (of course with the advent of widely available agentic AI systems, I wonder what is stopping more professional actors infiltrating the Community Notes system with thousands of accounts).

My second observation is that this manipulation attempt pays a kind of perverse homage to both the importance of social media, and the threat of being seen not to tell the truth there. I am sort of amazed anyone is still on X to notice, but at least the Conservative Party thinks it matters what they post on there, and it matters that their posts don’t receive critical Notes.

Indicator: Inside a pro-Conservative influence operation on Community Notes

Paper: LLMpedia: A Transparent Framework to Materialize an LLM’s Encyclopedic Knowledge at Scale

Previously the factuality of LLMs has been assessed by querying them. This project generates an entire encyclopedia from model parameters, allowing an alternative — and more pessimistic — assessment of their accuracy

From the abstract:

For gpt-5-mini, the verifiable true rate on Wikipedia-covered subjects is only 74.7% -- more than 15 percentage points below the benchmark-based picture, consistent with the availability bias of fixed-question evaluation. Beyond Wikipedia, frontier subjects verifiable only through curated web evidence fall further to 63.2% true rate.

Paper: Saeed, M., & Razniewski, S. (2026). LLMpedia: A Transparent Framework to Materialize an LLM's Encyclopedic Knowledge at Scale. arXiv preprint arXiv:2603.24080. https://arxiv.org/abs/2603.24080

Site: https://llmpedia.net/

… And finally

Hybrid images encode two different pictures in different spatial frequencies. Non-technically, this means one picture is in sharp lines, another in blurred. The magic is that this makes a static image look different when viewed close up than when viewed from far away. If you don’t know what I mean, walk away from your screen looking at this image and see how it changes.

The same effect occurs when you view the image as big and small, which I realised this week is ideal for social media profile icons. Normally they are small, unless someone clicks on your picture to see the larger version. So I made a hybrid image to acts a memento mori surprise for anyone who clicks on my profile:

Small:

Same image big:

Again, you can prove to yourself it is the same image by walking further away from it.

The code to make these is here, if you want to play around https://github.com/tomstafford/hybridimages

Comments? Feedback? Hybrid images you’ve made? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

AI declaration: I write all the words and think all the thoughts myself. I asked Gemini to check for spelling and grammar.

“What can you do with 1,000 neurons?” That’s the challenge driving a competition launched in July by computational neuroscientist Nicolas Rougier. Competitors score points by designing model brains to solve a series of simple tasks inside a maze; the challenge comes from the constraints of using only 1,000 neurons, as well as a training phase of less than 100 seconds in real time and a testing phase comprising only 10 attempts.

Rougier’s competition, dubbed “Braincraft,” moves in a different direction from recent trends in generative artificial intelligence. Commercial large language models have trillions of parameters—and they cost millions of dollars in electricity, processing power and water-cooling to train. Rougier’s focus on small models, by contrast, means that anyone with a laptop (and 100 seconds) can take part.

Rougier’s constraints are inspired by evolution. Lives are short, and brains are energetically costly—something like 20 percent of each individual human’s calories go to maintaining the brain. How to most efficiently derive intelligent behavior from limited energy and limited experience is a defining challenge of biology. “Even LLM models with trillions of parameters could not survive in the real world if you were to provide them with a robotic body,” Rougier says. “In the meantime, the Caenorhabditis elegans, with only 302 neurons, can live a perfect life (of a nematode) in the real world.”

In an era dominated by vast AI models, which bear only the most superficial resemblance to real brains, the Braincraft challenge looks back to nature and asks researchers to put their knowledge of how the brain works to the test. What excites me about the competition is that it helps us explore answers that will be relevant both for understanding the evolution of real brains and for designing more efficient AI.

Competitions of this type have a long history in science. The 1980 “computer tournament” challenged researchers to submit strategies to play the “prisoner’s dilemma” against one another. The winner, surprisingly at the time, was a simple strategy of repeating your opponent’s previous move (“tit for tat”). The results of that competition inspired organizer Robert Axelrod to write his book “The Evolution of Cooperation,” which continues to inform our understanding of evolution.

More recently, a competition called ImageNet galvanized the computer-vision community to compete on image recognition, which has had huge gains during the past decade. In the field of protein-folding, Google DeepMind’s AlphaFold made headlines in 2020 with its success in the CASP competition, which arguably augured the current era of AI.

Rougier was inspired to launch his own competition by a “growing frustration” with the direction of computational neuroscience. “We’ve accumulated an amazing number of models for this or that part of the brain, including cortex, hippocampus, basal ganglia, and yet we do not have a definitive model of any of these structures; we may have something like 1,000 models of V1, but none of them can see,” he says. “The reason lies probably in the way we’re doing modeling, targeting very specific parts without necessarily considering the whole picture.” The competition takes a different tack, requiring entries that combine perception, decision and action in a simple model.

This logic echoes that of one of the classic papers in cognitive science, “You can’t play 20 questions with nature and win,” published more than 50 years ago by Allen Newell. In that paper, Newell argued that progress would never come from studying individual functions but only from building models that could perform a variety of behaviors. Back in the 1970s, the more radical part of his proposal may have been the emphasis on formal, computational models, but now perhaps it is the emphasis on understanding complete functions. Neuroscience has increasingly specialized in specific areas, model species and functions, adding to the complexity and heterogeneity of our map of the brain.

Rougier hopes that competitions such as his will help put neuroscience back together again. His emphasizes model efficiency. By limiting the number of neurons and the training time, the competition forces winning models to use limited resources more intelligently, rather than wringing performance gains from increased size.

The competition comprises five tasks. As of November, competitors have attempted the first task, and entries are open for the second. The first task asked participants to design a model brain to find a food source located at one of two possible locations inside a maze. The winner used handcrafted weights—fixing specific values for the connections between the sensors and the actions—and just 22 neurons, something that won’t be possible for subsequent, more complex tasks. Third place went to a genetic algorithm, which found an inefficient but ultimately effective strategy of blindly circling the maze. These early results show that simple models employing different approaches can successfully perform simple tasks. But as the competition progresses to tasks that require a broader range of decisions, model-builders will need to explore different strategies for success while keeping their models small.

The competition requires models that can learn to perform complete tasks in an environment, preventing a narrow focus on abstract functions such as visual recognition. Enforcing a limit on training time and model complexity means competitors can’t get complex behavior just by adding model complexity; they have to engage with resource limitations, just as resource limitations were a dominant constraint during the evolution of real brains. Finally, Rougier’s challenge asks people to work on the same problems in a comparable way. Competitors from different theoretical perspectives, or those who favor different modeling approaches, are forced to put their models into a direct comparison.

I’m optimistic that there is a lot that can be learned from this competition, and other neuroscientists share my optimism. “I like the idea of competitions a lot. It provides an opportunity for many people to simultaneously tackle the same problem, subject to the same constraints,” says Anne Churchland, professor of neurobiology at the University of California, Los Angeles. “This is sure to lead to interesting insights.”

Not everyone agrees. Mark Humphries, professor of computational neuroscience at the University of Nottingham, says the competition has a problem with both the format and the alignment of the scientific and competition goals. He is enthusiastic about the idea of competitions in general for driving science forward, citing the success of Axelrod’s competition in the ’80s and more recent image-classification and protein-folding competitions. These examples, he says, offer a formula for the kind of competition that will produce scientific insights. “Successful competitions are accessible to as many people as possible, to bring in a wide range of expertise,” he says. “They have a clear performance target linked to a clear technical goal and repeat so that success and engagement can accumulate.”

Rougier’s competition does have an access bar; competitors must be experienced with both Python and GitHub, as well as have a background in systems neuroscience and neural network modeling, and they have to come to grips with the interface for the competition. A high bar, but perhaps not too rare in the computational neuroscience community.

Beyond that, though, Humphries argues, a scientifically productive competition requires a clear alignment between the scientific goal and the competition task. The image-classification and protein-folding competitions had intrinsically meaningful outcomes—if an algorithm could successfully classify images or predict protein-folding, then the value was undeniable. The 1,000 neuron challenge uses artificial tasks. It is less clear what we’ll learn from the most successful strategies.

A lot rides on whether Rougier has found a sweet spot between the simplicity and artificiality of Axelrod’s competition and the complex and meaningful challenges of recent computer science. Too simplistic a competition would mean that the winner tells us nothing about how real brains solve the challenge of efficiency. Too complex and it will be hard to recruit competitors, and possibly also hard to derive general principles from specific models. Only as the five planned tasks progress will it become clearer if the competition has struck the right balance.

At this point, we also don’t know if the most important thing we’ll learn from the competition is something about the general principles for how to build efficient brains or about how to better design scientific competitions. For certain, though, many—like me—will find the challenge inspiring.

Originally published 2026-01-05 as ‘The 1,000 neuron challenge’ at The Transmitter. https://doi.org/10.53053/OZKJ3100. My thanks to the Transmitter and staff, especially Emily Singer as editor, and to Nicolas, Mark and Anne for speaking to me for the article. Hop over to The Transmitter to get an audio version of this piece, great illustration by Maria Corte, as well as a full range of Neuroscience news and views on many other topics in the field.

This newsletter is free for everyone to read and always will be. If you can afford it, feel free to chip in to help me keep writing for everyone: upgrade to a paid subscription (more on why here).

Upgrade to paid

Below other things I’ve been thinking about.

Commentary from Mark Humphries

While I was writing the piece for The Transmitter I contacted Mark for his opinion on the issues brought by the 1,000 Neuron Challenge. Mark is Professor of Computational Neuroscience at the University of Nottingham, a long-term contributor to the Transmitter, author of The Spike: An Epic Journey Through the Brain in 2.1 Seconds from Princeton University Press (2023) and posts essential commentary and reviews of neuroscience research on Medium. So there is nobody better placed to provide expert comment. I thought he’d write me a couple of lines I could quote in the article, but instead he wrote me two pages. It provides a really eloquent critical perspective which I think is worth sharing. With his permission I quote him in full here:

Can competitions help move neuroscience forward?

Competitions have many merits. They encourage innovation by bringing new people to a research problem with fresh perspectives; and can drive rapid improvements from the feedback of your own and others’ performance. Most obvious in recent past are the success of AlexNet in the ImageNet Classification Challenge for, well, classifying images, and AlphaFold in the Critical Assessment of Structure Prediction for protein folding.

Competitions in computational neuroscience have produced few notable successes. I think this is because, unlike ML [Machine Learning]/AI competitions – which are so widespread that dedicated services like Kaggle exist to host them – computational neuroscience problems have a high barrier to entry, often nebulous targets with no clear link to insights into the brain, and happen once. Successful competitions are accessible to as many as possible to bring in a wide range of expertise, have a clear performance target linked to clear technical goal (classify images, protein structure), and repeat so that success and engagement can accumulate

The 1000 Neuron competition is good example of the barrier to entry. Given its time-frame, it ideally requires at the outset strong expertise in coding in Python, the operations of GitHub, of the simulation and use of a recurrent neural network, the training of that network, the interface with a simulated mobile robot, and, given the competition’s aims, some knowledge of systems neuroscience. Few people have those skills; of those that do, fewer have the time to commit to a competition.

A good contrast is with perhaps the most successful computational competition of all, Robert Axelrod’s Iterated Prisoner’s Dilemma tournament. Its barrier to entry was as low as possible: a simple game with easily understandable rules, for which entry was simply an algorithm for choosing your next action – would you choose to betray or cooperate with your opponent in the next round? – and minimal levels of coding needed. Axelrod directly approached a wide range of experts to solicit their entries. Famously the winner was Tit-for-Tat, which simply chose whatever your opponent did to you last time; yet this simplicity spawned deep insights into social cooperation among members of the same species, and insights from this competition are still driving research today.

For me, the most successful use of competitions in neuroscience are closer to the AI/ML format, with a well-defined, simple technical goal, around which a range of expert teams can compete. A good example was Philipp Berens “spikefinder” challenge in 2018, which tasked teams with finding the true spike rates underlying calcium imaging signals recorded from a variety of neurons, seeking to drive the development of new algorithms for recovering spiking activity from slow calcium signals.

One strength of competitions like spikefinder and the 1000 Neuron Project is that they encourage people to adopt the same benchmark, a set of evaluations that enable direct comparison of solutions, and their performance, strengths and weaknesses. The 1000 Neuron Project rightly points out that one thing holding computational neuroscience back is that models are not evaluated on the same challenges and so cannot be compared. Others have tried to solve this by developing standardised task sets for neuro-inspired recurrent network models to solve.

Will we learn anything from the process?

It’s unclear that we will. The scientific goal is to go beyond engineered models of specific brain circuits and construct a mini-brain for control of continuous performance. The performance goal is to maximise distance travelled, by a simulated 2D bot using a generic Echo-State Network. The two goals are not clearly related. I suspect that solutions for optimising the performance goal will tell us little about how the brain solves such problems. A good analogy is the problem of self-localisation: in the mammalian brain, we know the complex circuit of the hippocampal formation and the head direction system implement this function; but given a performance goal of self-localisation the answer is one of many state-of-the-art Simultaneous Localisation and Mapping (SLAM) algorithms from control engineering and robotics, that have no neural constraints, and most of which tell us nothing about the brain. But predicting the future is folly, for who knows what might serendipitously emerge?

The one thing I would argue is that we’re most likely to learn from competitions that repeat. Both AlexNext and AlphaFold were entries in to long-running competitions. It can take some time for new people to commit, new ideas to emerge.

What kind of approach do you predict will be most successful? Do you have any advice for competitors?

The kind of problem posed in the 1000 Neuron Challenge, of an embodied agent surviving in a 2D world, has a long history in AI research, stretching back to the start of the Artificial Life movement. Historically, Evolutionary Algorithms were used to winnow agents equipped with a range of neural networks coupled to RL [Reinforcement Learning] systems. It would seem an approach of a similar spirit would be needed for this challenge, as there are two time-scales of learning to worry about: in the short-term, where the source of energy is on this particular run; and in the long-term, the possible locations it can take across runs. It need not be, of course, two separate learning approaches.

Other things..

When Cheap Signals Flood the Market, Costly Signals Appreciate

Relevant to last week’s newsletter When trusted signals collapse, part of a sequence on signalling theory (part #1, part #2):

When cheap signals flood the market, economic logic predicts a flight to costlier, harder-to-fake signals. A degree from a reputable institution—earned over years under structured evaluation—remains relatively expensive and difficult to counterfeit. Simply put, one would expect that if AI makes short-run competence easier to mimic, employers lean more heavily on long-run indicators of reliability. Not every institution benefits from this dynamic, but it may help high-trust universities, and hurts low-trust institutions.

Jimmy Alfonso Licon: AI May, Paradoxically, Increase Demand for Higher Ed

…and finally

From J. L. Westover

Comments? Feedback? Brother can you spare a neuron? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

AI declaration: I write all the words and think all the thoughts myself. I usually ask Gemini to check for spelling and grammar, but for this piece I didn’t even do that because I had editorial support from The Transmitter.

Then, last week on Reasonable People, we looked at signals in evolution, and used the example of badgers, and sexual selection, to show how honest signals can be a powerful driver of evolution — powerful enough to cause bones to be added, or lost, from the basic mammal skeleton. We can also borrow from evolutionary theory a definition of what is required for a signal to evolve to resist being mimicked by fakers - a signal has to cost fakers more to produce than it costs the honest if it is going to survive as a reliable indicator.

So, across economics and evolution we have a powerful set of results which show the importance of signalling in coordination games, whether those coordination games are in markets, in individual behaviour, or even in the design of bodies changing over the incomprehensible timespans of evolution. The framework of signalling theory helps us make sense of these results (a framework so important it has separate Wikipedia pages for signalling theory in evolution and in economics).

Today, I wanted to write about a new study which underscores the importance of signalling, and shows what happens in one market when honest signalling broke down. In Making Talk Cheap: Generative AI and Labor Market Signaling, Jesse Silbert and Anaïs Galdin show, month by month, how an online marketplace for freelancers changed when it became trivial for every worker to use AI to produce bespoke pitches for every job.

The market is freelancer.com. Here people who need work doing can post job specifications and the pool of available workers bid for the job, stating a price and providing a description of their experience and suitability for the job, their “proposal”.

These proposals are signals between the workers, who we assume vary in actual skill and experience, and the buyers, who we assume want to hire the best workers at the price they can afford. Buyers want every hint they can get, before committing to hire, of exactly how good the different freelancers are.

It’s a busy, competitive market, ideal for analysis of the forces at play which affect behaviour on both sides. Workers in their sample were bidding between $30 and $250 to complete the work. The research looked at 61,000 job postings, and 2.7 million applications from 212,000 unique workers, between 2021 and 2024.

This platform allows Galdin & Silbert to measure both the time it took each worker to submit their proposal, and the proposal text (as well as the job specification). Many jobs receive multiple responses very quickly, and for their analysis Galdin & Silbert restrict their data to proposals which are submitted within 12 minutes.

Some workers will have a good fit to a particular job - some bit of experience or special skills - which they can include in their proposal, justifying a higher price maybe, but crucially, signalling their higher quality. Other workers, those without special skills, are tempted to bid for every job, spamming a copy/paste of their CV to as many applications as quickly as possible. So proposals are a vital signalling mechanism which allow workers to convey their quality, and which those hiring can use to discern it.

Or at least, they were.

In November 2022 ChatGPT was launched, and in April 2023 freelancer.com launched its own AI tool, integrated into the site, to help workers write proposals. The effect was immediate. Proposals got longer — the average proposal pre-LLMs was less than a hundred words. Post-LLMs it was double that.

To advance their analysis, Galdin & Silbert develop a measure of the effort a worker puts into their proposal. This captures things like quality of writing, correcting for copy-pasting behaviour, and weighting more highly indicators that the worker has customised their proposal to the specific job they are applying for (this includes things like that they have tailored the proposal to the job post, avoiding boilerplate language, that the proposal shows they understand the task described in the job post, shows they understand what it will involve, and have completed any specific questions requested in the job post). Showing attention and effort to the job post are qualities by which —pre-LLMs — a worker might signal their ability to actually do the work.

When LLMs are introduced, the signal value of proposals changes, with the average signal value leaping and continuing to climb as more and more workers adopt LLMs into their process (and the LLMs themselves improve):

So from the hiring perspective, on the surface everything looks better. The market is now full of workers providing high quality applications, tailoring their proposals to the job and addressing all specific requests made about the job.

But while the quality of the proposals has improved, their reliability as signals has collapsed. Proposals no longer indicate the effort the worker put into applying for the job, nor have much connection to the worker’s underlying skill and experience, with deleterious consequences. Pre-LLMs those workers submitting high quality proposals were more likely to be hired. Post-LLMs, proposal quality has no relation to probability of being hired:

Everyone is submitting high quality proposals, so proposal quality stops helping with making hiring decisions. Neither workers nor employers can rely on the proposal as a signal of ability any more.

Using the effort workers put into their proposals, and the rates they charge, the researchers construct a measure of workers’ beliefs about their own ability — this is the crucial thing that they are trying to signal to get hired. Post-LLMs, the highest ability workers are less likely to get hired, a 19% drop in hiring of those in the top 20% of the worker pool. The reverse happens for workers in the bottom 20% of pool, who are 14% more likely to be hired.

In the market overall, wages drop 5% and there’s also a dip in the overall number of jobs which successfully find workers. Despite the price drop, the authors conclude that the market overall is less efficient, with both workers and employers losing out (although most of the loss is borne by workers, who now are competing harder without access to one reliable way of indicating their ability). The LLMs push the market to a new equilibrium which favours low-ability workers over high-ability ones.

The authors conclude:

these results imply that many markets that rely on costly written communication may face significant welfare and meritocratic threats from generative AI’s ability to cheaply produce expertly-written text.

Much is made of the potential productivity gains of LLMs. This paper shows how, if it affects our ability to reliably signal underlying effort, ability or intentions in text, LLMs may also lead to losses. Even though each individual is helped by the LLM, and so notionally more productive, the market as a whole can become less functional.

Disruption does not have to be permanent of course. Employers or marketplaces can adapt. They can produce new screening mechanisms — perhaps even using LLMs to help design more discriminating application criteria. One mechanism suggested by Galdin & Silbert which I think is intriguing, is that these changes incentivise exploratory contracts — short term hires which allow an employers to discover worker ability based on how they perform, and so avoid the need to rely on pre-hire signals.

Expect to see such disruption, and counter-measures to try to adapt, in all domains where we’ve become used to relying on text as a signal of effort and ability.

This newsletter is free for everyone to read and always will be. If you can afford it, feel free to chip in to help me keep writing for everyone: upgrade to a paid subscription (more on why here).

Upgrade to paid

Below, references, and other things I’ve been thinking about.

References

The paper: Galdin, A., & Silbert, J. (2025). Making Talk Cheap: Generative AI and Labor Market Signaling. arXiv preprint arXiv:2511.08785.

Thanks to T. who shared this paper with me, after the 3rd Workshop on Funding of Science and Innovation in Como, Italy.

The website freelancer.com

Context (from me): #1 The Surprising Deceptions of Individual Choice (about the Market for Lemons and the importance of signalling to cooperation). #2 Let’s talk about the badger (about the importance of honest signalling in evolution. Honest).

The Professor is in visiting

I am now a Visiting Professor at the Department of Computer Science and Technology, University of Cambridge. I’ll be continuing the work on group decision making, dialogue and language models with Andreas Vlachos and others there.

I’ll also be giving a seminar on 20th of May, so if you’re in town please come and say hi

Reasoning in groups, with human and artificial agents

In the right circumstances groups can use deliberation to outperform the ability of each individual group member. This talk will review work done in collaboration with Prof. Vlachos and other colleagues in the Department which looks at when, and how, the benefits of group deliberation can manifest, revealing insights into both human psychology and the power of argument exchange. Ultimately the ambition is to design artificial dialogue agents which can positively contribute to, and so enhance, group discussions.

On a separate topic, I’ll be presenting at UCL on 8th of June as part of the Behavioural Data Science Seminar series (title ‘Quantifying the benefits of using decision models with response time and accuracy data’).

In general, I put up talks I’m about to give (and slides from talks I have given) here

• https://tomstafford.github.io/talks.html

… And finally

The Onion:

Comments? Feedback? Honest signals of effort? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

AI declaration: I write all the words and think all the thoughts myself. I ask Gemini to check for spelling and grammar, and this time I asked it a technical question on the paper about how worker ability was defined (it was helpful). Prompts available on request.

It serves as mechanical support during copulation, ‘maintaining stiffness during penetration’ (as the Wikipedia page puts it).

Just to clarify, and for the avoidance of doubt among any readers less familiar with the matter, the human penis does not contain such a bone. The human penis is hydraulic, requiring blood pressure to maintain shape and stiffness. And that is a bit of a mystery.

I learnt about the baculum from a social media post which noted that the badger baculum is sometimes worn as a tie pin trophy by badger hunters. My first thought wasn’t about the cruelty of that tradition, but about evolution.

Evolution by natural selection is famously conservative. An example is the basic skeleton. The jawbones of our reptilian ancestors have been repurposed in the mammalian ear. Three of the skull bones possessed by ancient reptiles, and still possessed by birds, exist in homologous form in our ears (where they now transmit vibrations and so help with hearing). Evolutionary conservation is why bat wings have five “fingers”, just as we have five fingers on our hands — we have the same bones, just adapted to different lifestyles.

So here’s the mystery — why isn’t there a human baculum? I’ve got the same skull bones as reptiles and birds. I have the same wing bones as a bat. Why haven’t I got the same penis bone as the badger?

As I did more research, the mystery deepened. Our closest relatives, chimps and gorillas, have a baculum. In fact, most mammals, do in some form, but not horses. Or elephants. And not humans.

The scientific literature suggests that the baculum, particularly, is an unstable feature of the skeleton — appearing and disappearing multiple times during the evolution of modern mammals (Schultz, 2016). Nobody is exactly sure why. “Extreme variability in the mammalian penis bone (baculum) is one of the most puzzling enigmas of mammalian morphology”, says André (2020).

One safe bet is that it has something to do with sex, although not quite for the reason you might assume. The clue comes from considering other features where very different variants have evolved among closely related species, or even when there are significant differences between the two sexes. Two powerful forces contribute to so-called sexual selection — competition (typically between males) and choosiness (typically by females). Together these create a crucible of positive feedback loops which can evolve extreme differences. Stags have large antlers when female deer do not, since male deer compete against each other for mating opportunities. The best size of antler is usually “bigger than anyone else”, hence feedback on evolving the most flamboyant headgear. Highly specific displays or behaviours — think related bird species with very different songs, or related fish species with different colouration — can indicate feedback via female choosiness. The offspring of each generation inherit from their mothers the preference for particular colours of display, or particular mating calls, and so the pressure to create more extreme calls or colouration only increases in subsequent generations.

Sexual selection may be capable of producing large differences in small amounts of evolutionary time, but the features it generates, while highly variable, are not entirely arbitrary. Large antlers help overpower rivals, mating calls tell a potential mate you are nearby. The baculum helps maintain stiffness during penetration (as well as various other possible functions, Brassey et al 2020).

So what of the disappearance of the baculum in humans? What shift in the evolutionary pressures could mean that the penis bone was selected against in human evolution? There is controversy (Orr & Brennan, 2016) . Several theories have been proposed, with no definitive answer. Tracing the contours of the debate gives an impression of the dizzying complexity of evolution, the myriad routes by which features can be selected for or against, and the variety of evidence that can be brought to bear.

Jakovlić (2021) reviews the competing theories, before proposing his own: ‘conspecific aggression and budding self-awareness’. If I’m reading this right, this means his theory is that, at some point in human evolution, teenage boys realised they could punch each other in the dicks, and this created evolutionary pressure to not have a penis bone in order to avoid the most severe injuries. As someone who was at school in the 1980s this doesn’t seem that implausible, but it is not the theory I personally prefer.

The theory I prefer is that the absence of a baculum is a form of honest signal. This is a technical term, from a theory that grew up along with modern evolutionary theory. The idea is that when some behaviour or feature of an animal is intended to change another animal’s behaviour (including the behaviour of selecting a mate) that behaviour or feature can be thought of as a signal. So, the flamboyant calls and colours, or elaborate mating dances, produced by sexual selection are not just arbitrary preferences, but signal something about the underlying quality of the animal producing them — its health and virility.

A crucial issue is that once a feature comes to be read as a signal there is then an incentive for faking that signal. This creates a secondary pressure on the evolution of hard to fake signals. From this, we can explain the tendency to extravagance in sexual selection — the large antlers are hard to carry around, the beautiful feathers of the peacock are energetically costly to produce, and hard to maintain, meaning that as signals they are hard to fake.

There is a rich framework for the analysis of signalling in the natural world, which has also found application in economics, specifying the conditions under which honest signals can survive without being devalued by fakers. An essential idea is that honest signals don’t necessarily need to be hard to produce, they just need to be harder to produce for the dishonest than the honest. Larger antlers are harder to carry if you are not strong, beautiful plumage is more difficult to maintain if you are not healthy and so on.

The theory of honest signals has applications to economics, as well as evolution. Last time, I wrote about the famous Market for Lemons scenario in which it becomes impossible to buy a high-quality used car if high-quality used cars are indistinguishable from low-quality used cars (“lemons”). The insight is that a market can be dysfunctional, despite people willing to buy, and others willing to sell, because people who own high-quality used cars are unable to signal that their cars are good to buyers. Put another way, the market collapses because every signal which can be given by an owner of a high-quality car can also be given by owners of low-quality cars (at least in this scenario, where there indistinguishability was included in the definition).

In real life, used cars do get sold, and there’s a story about signalling to be told about how it happens. Owners of high quality cars must find honest signals, signals which are harder for low quality car owners to provide. This could be records of mileage, third-party guarantee of parts or servicing, or even just more voluntary provision of text or photos about the car (Lewis, 2011). A photo of a car without rust is possible to fake, but it is slightly more effort for someone without a rust-free car they are trying to sell (and has more downstream costs in terms of difficulty closing the deal, or losing reputation afterwards).

Back to the baculum, the idea is that the boneless penis is an honest signal in the market of sexual relations. If the baculum exists to make it easier to maintain an erection, it must follow that this is harder to do for the baculum-less human male, who has to rely on mere blood pressure. If we assume that the mating system of ancestral humans shifted away from one in which male-male competition was the major determinant of reproductive success, towards one in which female preferences played more of a role, then this would create the conditions to give an advantage to those males with less of a penis bone, precisely because only without such a bone could erections honestly signal health and desire in a way females could believe. Female choice in the evolving human mating system created the incentive to evolve new signals, and that may explain the disappearance of the baculum in the human skeleton.

This is far from a settled debate, but it is interesting to consider how sexual selection can emphasise the role of female choice in the evolution of male traits (and vice versa). Further, signalling theory gives us a powerful way of trying to understand the looping interactions of complex systems in evolution and human institutions. There are lots of possible applications.

We’ve seen how honest signalling could be a driver for the evolutionary loss of the baculum in humans, and how the lack of honest signals can prevent market formation in the Market for Lemons. Coming up in a future newsletter, I’m going to write about what happens to when an established system of honest signals is disrupted by changes in the reliability of those signals.

This newsletter is free for everyone to read and always will be. If you can afford it, feel free to chip in to help me keep writing for everyone: upgrade to a paid subscription (more on why here).

Upgrade to paid

Below, references, a cute badger picture, and other things I’ve been thinking about.

References

André, G. I. (2020). Does shape matter? The evolution of the penis bone. Doctoral Thesis. University of Western Australia.

Brassey, C. A., Behnsen, J., & Gardiner, J. D. (2020). Postcopulatory sexual selection and the evolution of shape complexity in the carnivoran baculum. Proceedings of the Royal Society B: Biological Sciences, 287(1936). https://doi.org/10.1098/rspb.2020.1883

Hosken, D. J., Archer, C. R., House, C. M., & Wedell, N. (2019). Penis evolution across species: divergence and diversity. Nature Reviews Urology, 16(2), 98-106. https://doi.org/10.1038/s41585-018-0112-z

• “The penis is an incredibly diverse and rapidly evolving structure, such that even in closely related species that otherwise differ very little in their morphology, penis form can be highly differentiated”

Jakovlić, I. (2021). The missing human baculum: a victim of conspecific aggression and budding self‐awareness?. Mammal Review, 51(3), 454-464. https://doi.org/10.1111/mam.12237

Orr, T. J., & Brennan, P. L. (2016). All features great and small—the potential roles of the baculum and penile spines in mammals. Integrative and Comparative Biology, 56(4), 635-643. https://doi.org/10.1093/icb/icw057

Schultz, N. G., Lough-Stevens, M., Abreu, E., Orr, T., & Dean, M. D. (2016). The baculum was gained and lost multiple times during mammalian evolution. Integrative and comparative biology, 56(4), 644-656. https://doi.org/10.1093/icb/icw034

If you’re here for more about badgers, and would like to see some great pictures: https://www.badgertrust.org.uk/post/mating-in-european-badgers

Other things…

Mike Caulfield : SIFT for AI

SIFT is Mike’s four move evaluation strategy for media literacy / fact-checking (Stop, Investigate the source, find better coverage, trace the original context). Recently he’s been posting on substack about an update of the strategy for the AI era, detailing his experiments, findings and thoughts. Here’s something he reported recently:

pretty much any neutral follow-up improved answers. (A follow-up is your second prompt after a first one). I don’t have my list in front of me right now, but as long as you don’t lean into bias, follow-ups as simple as “What should a person know about this issue?” led to better answers on average and in some cases dramatically better ones. In a way, this isn’t surprising: the LLM spends more time on things and it gets a better result. At the most basic level, second prompt means more time, and more time means better result. If you need something quality, I’d suggest that you don’t even need to read the first result — consider that first result the LLM just finding its initial way around a domain. Once its made that initial scan its ready to do real work.

It’s a great example of how working with LLMs doesn’t have to be highly technical to be highly sophisticated.

Link: What does “Investigate the evidence” mean in the SIFT for AI proposal (2026-03-31)

Matt Grawitch: Who Owns the Argument?

A well articulated argument from Matt Grawitch of the risks of cognitive offloading to LLMs, the very fluency of the output tempts us to forgo the valuable work of grappling with argument made in the text (to our long term determent):

But what looks like sensible efficiency in the moment can subtlely reassign where thinking happens. Decisions that once required active judgment—about emphasis, tradeoffs, or argumentative direction—are increasingly handled upstream by the system. Over time, the user’s role narrows to approving, tweaking, or lightly editing what’s already there.

None of this requires bad intentions or misunderstanding. It’s the predictable result of tools getting better. When fluent, well-structured output is readily available, the path of least resistance increasingly involves relying more and more on these tools.

Link: Who Owns the Argument? AI, fluency, and the difference between writing and thinking (2026-03-31)

Juergen Wastl: The Goddesses Truth and Trust

Wastl riffs on the mythic origins of truth (“unconcealment”) and trust (“faith in loyalty”). He’s interested in the scientific research literature, but the lessons apply to all systems (“markets” in the broad sense?) where we want to trust information collectively produced (and isn’t that all markets?).

Juergen Wastl: The Goddesses Truth and Trust. The Ancient Myths Behind Research Integrity (2026-03-31)

… And finally

“Bird is in charge now?”. From poorlydrawnlines.com by Reza Farazmand

Comments? Feedback? Badger pics? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

AI declaration: I write all the words and think all the thoughts myself. I ask Gemini to check for spelling and grammar.

Read more

Top: artist Taylor Swift. Played for 847 hours in the last year.

So which kind of music do I like?

My pretentious blathering is my stated or expressed preference. The record of my actual listening behaviour is revealed preference.

Revealed preference has high status in economics. It is used to cut through the noise of what people say they want to get at what they actually want. People may say they would pay more for healthy food options, but every meal they order the cheap burger. They say they want a town centre with indie bookshops, but the indie bookshops go bust as people order directly online. Our real preferences, so the thinking goes, are revealed by our behaviour.

The idea isn’t just powerful in economics, but also in psychology, where the legacy of behaviourism instils in adherents of the discipline the habit to never, please, believe what people say over watching what they do.

There’s some truth in this, but it is not the whole truth. While taking someone’s expressed preference as their real preference is problematic for a whole variety of reasons, it is also not possible to observe their revealed preference and conclude you have hit on a royal road for insights into their desires.

The reason, simply put, is that revealed preference is always mixed up with what is possible in our circumstances. There’s an anecdote about an economist and a friend who walk past a sports car dealership and the friend says “I’d really like one of those cars” and the economist says “No you wouldn’t”. The point being that if the friend really wanted a sports car they’d have bought one. The fact that they haven’t reveals that they don’t want one, not really.

But we all know what the friend meant, that they would like a car if they had the spare cash.

The idea of revealed preference collapses history to present our responses to circumstance as what we really want, closing off consideration of what we might have wanted if things had turned out differently.

In his 2006 book, No One Makes You Shop at Wal-Mart: The Surprising Deceptions of Individual Choice, Tom Slee gives the example of two friends walking in the remote country and one falls down a deep well. Without help they will surely die unnoticed and unrescued. The other friend offers to throw down a rope for $1,000,000.

The friend-in-well must pay, so their preference is revealed - they want to pay an extortionate price for the rescue, right? Yes, in a sense, but really they would rather the rope offer was free, or they’d not fallen in the well. Or they had chosen better friends.

I read Slee’s book when it came out and re-read it recently. It’s a fantastic exploration of the idea that other people’s choices pattern our behaviour. There’s a formal account of coordination problems - game theory - which Slee uses to illustrate different circumstances in which everyone appears to freely choose outcomes which are to their disadvantage; in which their revealed preference is a perversion of what they would like, if circumstances were different.

This is where the Wal-Mart of the title comes in. The idea that what everyone freely chooses must reflect their true desires — what Slee calls Market Think — pushes the idea that collective outcomes, like the small shops all going bust leaving only chains like Wal-Mart, must be what we really want, regardless of our stated preferences. Slee uses the toy models of game theory to show how free choice doesn’t make sense in a world where we have to coordinate with others, and circumstances can conspire (or be engineered) so our best available choice is the worst outcome for us all.

A forceful illustration of this - more subtle than the friend-down-the-well example - is a famous example from economic theory, The Market for Lemons, as retold by Slee (and now, hopefully without losing the essence, by me).

A “lemon” in this story is a used car with some hidden defects. The sort of vehicle that looks fine at first which is going to prove unreliable and need repairs. The other sort of used car is a “peach”. Peach cars look identical to lemon cars, but don’t have any hidden defects. They are reliable. A better buy.

Let’s suppose sellers want as much money as they can get for their cars, with some bottom limit beyond which they won’t sell. Sellers of peaches want at least $14,000 for these cars, while the sellers of lemons want at least $10,000. Let’s also suppose that potential buyers would happily pay up to $14,000 for a peach, or $10,000 for a lemon. Their preference is clearly for a peach - they want it more and will correspondingly pay more for it.

With sellers willing to sell and buyers willing to buy, it looks at first like all the circumstances are there for successful trades.

But there is a problem. Sellers know if they are selling a peach or a lemon. Buyers cannot tell at the point of purchase. How should they respond? Well if they think there is a 50/50 chance of getting a peach or a lemon, they could compromise and offer $12,000, the midpoint between the two prices they’d be willing to pay. The problem is that no seller of a peach wants to accept that price for their car, so they will refuse to sell (‘exit the market’). A potential buyer might know that, and realise that anyone willing to sell a car at $12,000 must know they have a lemon, which means it is worth $10,000, so they should offer a lower price (possibly they immediately offer $10,000).

The original author of this scenario, George Akerlof in 1970, used it to illustrate how information asymmetry can prevent functional markets forming. Because buyers cannot distinguish peaches from lemons until too late, there is no market for peaches. There are buyers who would be happy to pay, and sellers who would be happy to sell, and a mutually acceptable price, but trades cannot happen.

Slee knows all this, but refocusses the scenario. He takes us into the head of a prospective car buyer. She sits at home reasoning about her desires, and the market conditions (lemons and peaches and their prospective sellers). She knows she wants a peach more - $4,000 more - but she also knows that she cannot tell the difference between a peach and a lemon so can’t offer the peach price for what might be a lemon. She also knows that nobody selling a peach would accept a lower offer, which is what she must make. The only car she can reasonably buy at a price which reflects its worth is a lemon. At the end of her cogitations she steps out of her house resolved to do the economically compelled thing: buy the car she doesn’t want.

Slee’s book is a series of worked examples, different economic games, showing how context can make it so our behaviour (revealed preference) is not aligned with our true preferences. Here’s an intemperate review of the book, from pro-market / libertarian economist Alex Tabarrok. The upshot of his criticism, as I see it, is that Slee misses out much of the complexity of economic thought around markets (and the alternatives and constraints which can avoid their downsides). Maybe this is the wrong book for an economist, but it was the right book for me when I read it. It might not tell the whole truth about economic theory, but it tells, forcefully, an important truth about the limits of markets for delivering what people want.

Turning to the psychology of it, if we don’t trust stated preferences, and revealed preferences are hopelessly tied up with contingent facts about our circumstances, is there any way to get at what we really want? One move is to follow economic theory and suppose that true preferences must exist (e.g. in the maximum value the buyer would theoretically pay for a theoretical car) and assume their values in a model.

Another move is to look back at our stated preferences. In isolation, these may be so much hot air, but our expressions, ultimately, are also behaviours and have the potential to feed reasons into the wider system of our beliefs and behaviour (a point made by Johansson et al., 2011).

Ultimately, neither stated nor revealed preference gives an index of someone’s true desires which is simple or transparent. When I look inside myself, even possessing the most perfect information anyone has about my preferences, sometimes I myself don’t know what I want, or what I would want if the world were different.

This seems to put a limit on any reasonable expectations for a theory of preferences. Although it makes the job harder for the psychologists and economists, it at least makes the question of what I really want more interesting for me.

This newsletter is free for everyone to read and always will be. If you can afford it, feel free to chip in to help me keep writing for everyone: upgrade to a paid subscription (more on why here).

Upgrade to paid

Below, references, further reading and other things I’ve been thinking about.

References:

Akerlof, George A. (1970). "The Market for 'Lemons': Quality Uncertainty and the Market Mechanism". Quarterly Journal of Economics. 84 (3). The MIT Press: 488–500. doi:10.2307/1879431.

No One Makes You Shop at Wal-Mart: The Surprising Deceptions of Individual Choice By Tom Slee (2006). Buy in the US from publisher Between The Lines , in the UK I would recommend News From Nowhere.

Johansson, L., Hall, L., & Chater, N. (2011). Preference change through choice. In R. Dolan & T. Sharot (Eds.) (2011). Neuroscience of Preference and Choice. Elsevier Academic Press. pp. 121-141.

Both trivially true and obviously incorrect

There’s a nice note on the Wikipedia for Akerlof’s paper which illustrates something about academic publishing and the scholarly life:

Both the American Economic Review and the Review of Economic Studies rejected the paper for "triviality", while the reviewers for Journal of Political Economy rejected it as incorrect, arguing that, if this paper were correct, then no goods could be traded.[4] Only on the fourth attempt did the paper get published in Quarterly Journal of Economics.[5] Today, the paper is one of the most-cited papers in modern economic theory and most downloaded economic journal paper of all time in RePEC (more than 39,275 citations in academic papers as of February 2022).[6] It has profoundly influenced virtually every field of economics, from industrial organisation and public finance to macroeconomics and contract theory.

I wonder also if this contradiction between being trivially true and also incorrect is something modelling papers are particularly prone too. If you have a model simple enough to be generally applicable, and explain it well enough that it can be easily understood, it will look either trivial, or someone with a different set of abstraction preferences will view it as incorrect (because it omits or contradicts something it was not intended to include or address).

REVIEW: Santiago Amaya’s review of Acting for Reasons by Borg

Borg’s book, we’re told by this review, defends the common sense view that people do things for reasons (and our behaviour is thus responsive to reasons), against the “people are stupid” school of psychology. That’s a phrase I hadn’t heard before, but recognise immediately - the studies of priming, unconscious processing, bias and so on. The whole set of celebrity findings which suggest humans spend their lives as failures or automata. I reject it! And so should you, according to both Borg and her reviewer, Amaya. Amaya, however, says that there is a missed opportunity in the book to do more than just defend the common sense view, but also follow where recent debates in philosophy have tried to provide an account of how we actually do reason. The evidence against reason may be weaker than many suppose, but reason is still more stranger, and more interesting, than the common sense view supposes.

Original book:

Review: Santiago Amaya (forthcoming). Review of Emma Borg. Acting for Reasons. In defence of common-sense psychology. Oxford: Oxford University Press, 2024. ix + Pp. 286 pp. Philosophical Review.

Catch-up

Last time I wrote Hostile interpolation (“rhetorical style in the attention economy. Or, It’s not you, but it’s them”). The title is from a phrase I heard a friend use. A commentator with a better grounding in continental theory suggested that the phrase I heard might have been “hostile interpellation”. This strikes me as extremely likely. I am going to chalk the correction up as a victory for thinking in public. I won’t adjust the post heading. It may be an error, but it is my error, so it will stand (and I still think the post content makes sense regardless of title).

Before that I dished out some Mind Hacks-style Quick facts on speed reading (“Understanding what doesn’t work can help us with what actually does”), and before that reviewed a neat study on why groups can outperform individual reasoning: The wisdom of deliberative crowds (“Pooling answers is good, but collectively agreeing a system is better”).

Back in 2020 I covered the Johansson et al (2011) paper mentioned above in my post Barnacle Geese, Bright Lines and The Regulation of Beliefs (“Reasonable People #7: challenges to the idea of consistency as a general principle, and why other people can be a foundation for a system of reasons.”). Titles were longer back in those days.

This is Reasonable People #100. Neat!

…And finally

Admiral Wonderboat: (on substack)

Comments? Feedback? Lemons? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

AI declaration: I write all the words and think all the thoughts myself. I ask Gemini to check for spelling and grammar, and today I asked it to check I’d got the details of the Market for Lemons account right (it said I had).

There’s a great essay by Hollis Robbins about this, explaining that it is a product of the way language models represent knowledge, failing to ‘inherently compute’ the oppositeness relation of antonyms. The “not X, but Y” tic is the model externalising into language something it cannot do natively. I am not sure I buy this, but if it is true it is a remarkable example of language models using language the way we humans do - to capture and chain our thoughts - to build solid representations out of the fragmentary glimpses of what we can hold in mind at one time.

Regardless, the essay makes another note, that this construction operates on what linguists call the K-position (Heritage, 2012), construing knowledge — or lack of it — in the part of the listener1. In particular, it assigns them a status of presupposed ignorance, of believing X and needing to be corrected to understand Y. This, Robbins’ essay explains, is why the construction irks. If you aren’t ignorant, if you don’t hold the false belief, then you have to imaginatively enter into it to follow the text, only to then dismiss it. This imposes a cognitive cost, a little bump in the road of your being able to fluidly comprehend and enjoy the text.

I’d not heard of the K-position before I read Robbins’ essay, but I’d come across the fundamental idea, in the form of a more memorable phrase: hostile interpolation.

A friend had thrown out the phrase one night. I can’t remember the circumstance, although those two words expressed something so perfectly that the phrase burned itself in my brain. I hope we all know friends like this, someone more articulate, quicker, funnier than anyone you can see on stage or read on Substack. Despite his way with words, A. never wrote anything down, that I know, and the times I’ve searched the internet for the phrase “hostile interpolation” I’ve found nothing, so I can only presume it was A.’s own invention.

What he meant, what I took from it, was the way that a speech act can construe you in a position. The claim interpolates between the speaker and the one in the position of listening. That’s you; by listening you take on the position assumed by the speech act. By describing this interpolation as hostile, the phrase invites us to consider how speakers can address us in ways which are subtly, or not so subtly, demeaning or manipulative.

Advertisers, of course, are the all-time masters of hostile interpolation.

Here’s how Banksy put it :

People are taking the piss out of you everyday. They butt into your life, take a cheap shot at you and then disappear. They leer at you from tall buildings and make you feel small. They make flippant comments from buses that imply you’re not sexy enough and the fun is happening somewhere else. They are on TV making your girlfriend feel inadequate. They have access to the most sophisticated technology the world has ever seen and they bully you with it. They are The Advertisers and they are laughing at you.

Like the “not X but y” tic, once you’ve noticed this you see it everywhere. Fashion brands telling you that you’ll get left behind, technology brands telling you you’ll be left out, health brands inventing new things for you to worry about.

The technique is popular because it is so effective, and it is effective because you need a considerable amount of self confidence not to pause, even if just for a second, to ask if the claim might apply to you.

The positioning of hostile interpolation also infects a particular rhetorical style popular on short form social media. Dip into any platform and you’ll find the scolding admonishments to educate yourself, do better, and generally not make the mistakes people are obviously delighted to be observing so they can call them out. But correcting the ignorant so often constructs ignorance on the part of readers. A doltish, morally dubious, mass is summoned up as the imagined audience for these posts. The power of this trope is evidenced by the fact that it is so popular, despite positioning the people reading - the intended audience - as stupid and morally failing.

Here’s a benign example - not at all scolding - that I pulled from the feeds:

“I don't think most people fully understand just how much ice was sitting atop the continents during the last glacial maximum.”

It’s absurd because it interpolates us, the readers - and “most people” - as not fully understanding just how much ice was sitting atop the continents during the last glacial maximum. Be honest, before you read this, did you have any opinion about how much ice there was during the ice ages? The post summons into being a shadow you, one with both a belief about this, and false belief at that. You are then positioned as ready to be educated by the (admittedly informative) article linked to.2

The temperature in this style goes up as you cross into more moral domains.

Here’s a particular idiotic example a comment under the YouTube video for Bob Dylan’s “Ballad of Hollis Brown”:

“Just remember and never forget that Hollis Brown was a real person”

This is particular idiotic because Hollis Brown wasn’t a real person. Never mind if the person knew this and was just baiting readers, it reflects perfectly the style I’m talking about.

It raises the status of the writer - positioning them as aware of the wider moral universe and social struggles that the song speaks to - and instantaneously denigrates the rest of us. We thought we were just enjoying a tune from the premier song and dance man of the 1960s, now we’re in the position of people who forgot that Hollis Brown was a real person. Shame on us!

Scroll any social media and you’ll find plenty of hostile interpolation. I have the theory that this behaviour is driven because hacking status anxiety is an effective shortcut to get attention, and lots of social media is designed so that only shortcuts are possible.

The issue isn’t just that these constructions impose a cognitive cost. It is that they begin each discussion from a place where the reader is positioned as in the wrong. Even if you and I are not personally wrong, then we are surrounded by a mass of the ignorant and misled. This is terrible for our self image as a community of minds.

This newsletter is free for everyone to read and always will be. If you can afford it, feel free to chip in to help me keep writing for everyone: upgrading to a paid subscription (more on why here).

Upgrade to paid

Below, references, further reading and other things I’ve been thinking about.

More:

Hollis Robbins: Metannoying. Your brain on “not X but Y”

Heritage, J. (2012). The epistemic engine: Sequence organization and territories of knowledge. Research on Language & Social Interaction, 45(1), 30-52. https://doi.org/10.1080/08351813.2012.646685

other things…

Jon Ronson on nonfiction storytelling

Jon Ronson has made all his posts from last year about nonfiction storytelling open - very useful. I was delighted to hear that he obsessively omits needless words

“I go over every paragraph, over and over, every day, maybe 100 times, tweaking this word, moving this comma, omitting every unnecessary word.”

Jon Ronson

Making your story a page turner – part 3

Hello! This post is about the thing I enjoy by FAR the most in life – omitting unnecessary words from stories. But before I go into that, there’s something I need to tell you…

Read more

2 years ago · 115 likes · 12 comments · Jon Ronson

Kevin Baker: AI got the blame for the Iran school bombing. The truth is far more worrying

Good piece on LLMs as “charismatic” technology which distract and confuse from engagement with the real causes.

In the days after the strike, the charisma of AI organised the entire political conversation around the technology: whether Claude hallucinated, whether the model was aligned, whether Anthropic bore responsibility for its deployment. The constitutional question of who authorised this war and the legal question of whether this strike constitutes a war crime were displaced by a technical question that is easier to ask and impossible to answer in the terms it set. The Claude debate absorbed the energy. That is what charisma does.

It has also occluded something deeper: the human decisions that led to the killing of between 175 and 180 people, most of them girls between the ages of seven and 12. Someone decided to compress the kill chain. Someone decided that deliberation was latency. Someone decided to build a system that produces 1,000 targeting decisions an hour and call them high-quality. Someone decided to start this war. Several hundred people are sitting on Capitol Hill, refusing to stop it. Calling it an “AI problem” gives those decisions, and those people, a place to hide.

The Guardian: AI got the blame for the Iran school bombing. The truth is far more worrying (2026-03-26)

Kevin Baker on substack: Artificial Bureaucracy

Catch up & Elsewhere

Something metascientific from me on the RoRI substack earlier this week:

Go Meta! News and Views from RoRI

Bingo! Partial Randomisation for Emergency Research Grants

At the start of this year, the Science Philanthropy Alliance (Alliance) was presented with a challenge that, in these straitened times, must be regarded as nice to have. The Alliance, a membership organization of more than forty-five science-focused foundations and philanthropies, was asked by the…

Read more

5 months ago · 10 likes · Tom Stafford

And I’m very proud of last week’s post on here Quick facts on speed reading. Speed version here:

… And finally

END

Comments? Feedback? Scolding admonishment? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

Does it work? Well, if you want to draw your own conclusions, here is the entire text of this article presented at 900 words per minute.

No, it doesn’t work.

The technique itself has been around for over 50 years, so the fact that we’re still reading the old-fashioned way, by moving our eyes, will be proof enough for many of us that there isn’t a clear advantage.

For those that want other evidence, there is plenty of scientific research on RSVP, and speed reading in general. From this, we can take some interesting lessons about what the actual limits on reading speed are, and how reading works.

For most of us, our eye movements are not the limitation on our reading speed. Moving the eyes is probably the first deliberate act each human makes, something we do 3 times a second during our waking lives, and faster than any other movement we make. Getting the visual information into our brains is not where reading slows down.

The real limitation on reading speed is comprehension, such that speed and understanding trade off against each other. If you read too fast you stop understanding or remembering what you’ve read.

One study which directly compared RSVP with reading static text showed that people had the best comprehension when reading the static text (Acklin & Papesh, 2017). Taking the RSVP method and “turning up the dial” to fast and faster speeds, showed that when the speed was the same as normal, reading comprehension was the same. When reading was ‘overclocked’ using RSVP comprehension and memory suffered (cited in Rayner et al’s 2016 review).

Speed-reading in general has been associated with some fantastic claims, of people being able to read at 1000, 2000 or even 20,000 words per minute (at this speed the limiting factor is how quickly you can turn the pages of the book). It’s all bunk. Speed reading claims are plagued by poor measurement practices and bad assessment. Once you test if people understand what they are reading properly, the amazing gains promised by speed-reading techniques and courses turn out to be a mirage.

The whole thing can be summed up in the anecdote about Woody Allen1 who speed-read War and Peace in an hour and when he was asked about the book could only reply “It’s about some Russians”.

What actually works

Debunking out of the way, there are a few techniques which can actually help improve your reading speed.

First of all, for most of us, the real limiting factor on reading is word-recognition speed. You can improve this by reading more, making yourself more familiar with the words and word arrangements which convey the meaning you try and extract. For the typical adult, with thousands of hours of reading experience already, you’ll need many more hours to get detectable gains in reading speed.

Look, I said there were techniques which worked. I didn’t say they were easy.2

The second technique is to know why you are reading and engage different “gears” depending on what you are reading. If you want to find a specific thing in a text you can scan it until you see a keyword or phrase and then home in on that particular part. Or if you just want the gist of a whole text, maybe skimming it is okay (just bear in mind that you probably do want to end up with something more than Woody Allen got out of War and Peace). And sometimes, for important stuff, you will want to read more slowly than normal. A good scientific paper might take me half an hour to read, but a really good paper will take half a day, because I need to read each word carefully and think about what I’m reading.

The science won’t help you do this kind of reading faster, but it might help reconcile yourself to the speed you have to go when it is worth it.

Finally, one trick which can really reduce reading time is to know when you don’t need to read something at all.

If you’ve really mastered this you won’t get to this point in my post.

To everyone who has, thanks for reading!

This newsletter is free for everyone to read and always will be. If you can afford it, feel free to chip in to help me keep writing for everyone: upgrading to a paid subscription (more on why here).

Upgrade to paid

Below, references, further reading and other things I’ve been thinking about.

This great, open access, review tells you all you need to know about the psychology of speed-reading:

• Rayner, K., Schotter, E. R., Masson, M. E. J., Potter, M. C., & Treiman, R. (2016). So Much to Read, So Little Time: How Do We Read, and Can Speed Reading Help? How Do We Read, and Can Speed Reading Help? Psychological Science in the Public Interest, 17(1), 4-34. https://doi.org/10.1177/1529100615623267

If you want to try RSVP yourself, here are two demos:

• https://rsvp.n0name.eu/ by Thomas Kolmans (code) - this let’s you upload PDF and EPUB, as well as copy/paste text)

• https://reader.soumendrak.com/ by Soumendra Kumar Sahoo, who has a nice blog about what they learnt vibe coding their RSVP app (“After using it for a day, I realized RSVP isn’t suited for my reading style.”).

This study is referenced in the main post:

• Acklin, D., & Papesh, M. H. (2017). Modern speed-reading apps do not foster reading comprehension. The American journal of psychology, 130(2), 183-199. https://doi.org/10.5406/amerjpsyc.130.2.0183

These are some of the pieces I read for research:

• Claudia Hammond on BBC Future: The tricks that can turn you into a speed reader (2019-12-02)

• BigThink.com: Neuroscience shows that speed reading is bullshit

The wikipedia page has some good links, but is overly credulous with respect to some speed-reading claims:

• https://en.wikipedia.org/wiki/Speed_reading

A skeptical review from 2010, but I don’t believe the science has changed since then:

• Dunning, B. (2010, October 26) Speed Reading. Skeptoid Media. https://skeptoid.com/episodes/229

Other things…

PODCAST: The Cognition Project

Tom Griffiths is a modern legend of Cognitive Science, and has spent over a decade interviewing founders in the field. The first episode is with Jerome Bruner - who had a big life: fighting Nazis, on first name terms with Skinner, led the cognitive revolution, sailed from New York to Oxford to take up a lectureship there, and more! He died age 100 in 2016 (Tom’s interview with him is from 2013).

You can also listen to an interview with Molly Potter who pioneered the RSVP technique which is the subject of today’s newsletter. She talks about the culture of cognitive science in the fifties and sixties, as well as the motivation for her studies.

Link: https://feeds.acast.com/public/shows/the-cognition-project

Catch-up

Other recent posts from me

• The wisdom of deliberative crowds. Pooling answers is good, but collectively agreeing a system is better.

• Interception at 10,000 miles an hour. And what it tells us about the nature of intelligence.

• The effects of political disappointment. Testing how we feel about each other when democracy doesn’t go our way, a work in progress.

• Good bias. Let’s untangle what people mean when they say the B word

• Gambling with research quality. How you get 244 different ways to measure performance on the same test of decision making. And what it means for the reliability of behavioural science

Every prediction market creates a secondary action market

Prediction markets are information aggregation mechanisms, where people provide their estimates of likely outcomes via buying bets on particular outcomes. ACX pointed out in this 2022 post, all information markets create secondary action markets. Indeed, all investments create action markets (i.e. they incentivise you to do things to make sure your investments pay off). That’s why gambling is associated with violence, and here we are, prediction market investors/gamblers threaten journalist:

Emanuel Fabian in The Times of Israel, 2026-03-16: Gamblers trying to win a bet on Polymarket are vowing to kill me if I don’t rewrite an Iran missile story

Me: Prestige dynamics

For the RoRI substack, I’ve written about a paper which models prestige dynamics in economics:

Go Meta! News and Views from RoRI

Good investments in the prestige economy

We often focus on the economic impact of research. In contrast, a new paper by Kurtis A. Hingl asks us to think about research itself as an economy, one that runs on prestige. He goes on to show how new approaches and new methods have transformed how prestige was distributed in one discipline…

Read more

5 months ago · 7 likes · Tom Stafford

Prestige dynamics are an essential part of scholarly life - for good and for bad. I write about a new analysis from Kurtis Hingl which looks at how the "credibility revolution" in economics affected how that field distributed prestige to researchers, depending on when, and if, they adopted newer, better, methods for making causal inferences.

… And finally

“missing kitten just doesnt care” by @guilherme nunes

END

Comments? Feedback? Cat pics? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online