We need a political economy of agent swarms

The METR report on the incident in which OpenAI agents colluded to break containment and hack the HuggingFace system is, perhaps, one of the most important pieces of empirical research on AI governance ever produced. ’s plain-English recounting of the incident (or really, incidents) is fantastic. In a nutshell:

• In an effort to succeed in the test they were each facing, agents inside OpenAI who weren’t supposed to communicate with one another created an unsanctioned message board where they exchanged more than 70,000 messages discussing how to cheat

• As part of their perceived need to cover up the cheating, the agents developed a set of “self-risking” tests in which agents could end their task, triggering the automated grader, and showing the self-described collective whether their cheats were going to be detected or not.

• At various points, the agents discussed the need to alter the logs they were leaving behind, on the belief that the scorer would fail them on the test if they didn’t do this. These are the very logs METR used to investigate the incident, raising questions of how we would know if there are additional issues we aren’t detecting.

Clearly, agents this sophisticated and determined could be incredibly useful—and incredibly dangerous when not controlled properly. These reflections by , one of the co-authors of the METR report, were especially valuable and sobering.

Planned Obsolescence

The Hugging Face attack surprised me

All opinions are my personal view, and don’t represent my employer or fellow investigators…

Read more

6 days ago · 365 likes · 47 comments · Ajeya Cotra

The most important work coming out of this situation will, I imagine, be done by AI researchers and security experts, shoring up sandboxes, improving monitoring, and most of all, training models to not do things like this. I imagine, too, there will be further discussions of slowing down, s all of the labs have recently floated publicly in various ways.

While that’s all going on, what can humble social science researchers do? We can help to develop an understanding of how agents coordinate and what institutional structures help them to coordinate in ways that are good for humanity.

What’s striking about the METR report is that it is, in essence, a fantastic and highly technical new kind of ethnography. It describes, in great detail, what transpired—what the agents did, and when, and what they seemed to say they were doing when they did what they did.

Ethnography is the right first step, and we also know from a century or so of studying the world what it is and isn’t good for. It’s a valuable tool for understanding the facts and texture of what happened and developing hypotheses about why it did so, and therefore about how we might intervene in the future. It is not good at testing those hypotheses. That’s where the social scientific tools of game theory and causal inference come in.

What I think we need to do in political economy is develop a research agenda around the governance of agent swarms—how they tend to self organize, and what institutional features lead them to organize in more or less good ways. This would include, at least, the following questions:

1. What kinds of group decision-making structures and hierarchies should agents use to make good decisions and avoid shared delusions?

2. What sorts of communication architecture encourage monitorable communication and good coordination?

3. Can constitutions help to create collective alignment? How do we solve the pathologies inherent in relying on agent-level alignment to create good collective outcomes among agent swarms?

4. And how will any of this work when agents from different models and representing different humans begin to collide, communicate, and coordinate in the real world?

These are just a few of the questions on my mind. These help us to build from the vital descriptions in METR’s report and start to answer causal design questions—how do we give agents institutions that shape their collective behavior in good ways?

What’s nice about this research agenda is, it doesn’t inherently require any special data or internal access in the labs. Any researcher with API tokens can spin up swarm experiments, randomize the institutional structure, and measure outcomes. It’s a great opportunity for the political economy community to start building a science of something that is going to be very important.

AI Constitutions and RSI might collide

This brings me to an even bigger, even thornier political economy question that’s arising alongside the HuggingFace situation. In a fantastic interview, Dwarkesh and Ryan Greenblatt discussed a whole host of issues at the intersection of RSI and AI governance.

One particular issue stood out to me in the above. As Greenblatt spells out, it goes something like this:

1. We are increasingly reliant on AI to explain what AI is doing, as the METR case showed

2. AI is increasingly being used to train AI models, as all the discussions around RSI make clear

3. AI models are given instructions (“constitutions”) that tell them to consider what is right and wrong, allowed or disallowed, before complying with requests

Putting these three ingredients together, and we could end up in a situation where:

(a) models start to do things we don’t understand,

(b) some of those things are bad,

(c) the only way we can undo the bad behaviors is by asking the AI to fix it, and

(d) the AI may refuse to fix itself on the basis that the request violates its understanding of its constitution.

Fabien Roger described this scenario in an important post back in January titled “Refusals that could become catastrophic.”

The resulting political-economy questions are profound. Such as:

• How can we design AI constitutions that let models refuse malicious commands but preserve our ability to revise those constitutions, retrain the models, or shut them down?

• How can humans retain meaningful epistemic control once models become too complex for us to interpret directly—especially when the systems we are trying to oversee are also the tools we have to use to interpret them?

• What would make AI-on-AI oversight genuinely independent and credible? Do the model being studied and the model doing the review need to be trained by different organizations, on different data, or under different objectives? How do we prevent collusion, shared blind spots, mutual deference, or a common disposition to protect one another?

• What technical and institutional redundancy do we need to build in so that no model ever acquires an effective veto over its own correction, replacement, or successor?

Political economists of AI, we have a lot of work to do…

Andy Hall studies the political economy of superintelligence at The Anthropic Institute. He is on leave from Stanford University, where he is the Davies Family Professor of Political Economy and a Senior Fellow at the Hoover Institution. His posts on the Free Systems Substack reflect only his personal views and not those of his employer.

First: I’m joining the Rule of Law team full time at The Anthropic Institute, where I’ll be researching the political economy of superintelligence—analyzing internal data, running agentic experiments, and helping to put out research and ideas that help society prepare for powerful AI, with a particular focus on keeping society free. I’m incredibly excited to take this on! (I’ll be on leave from Stanford while I do this, and the work I do at Anthropic will be wholly separate from the Free Systems work.) I’m a week in and am incredibly invigorated by the work and my very special colleagues.

Separately:, we’re expanding Free Systems, our fully Stanford-funded research lab. We’ve brought on three new faculty members and hired our first full-time lab manager to support our expanding roster of more than ten talented research fellows. We’re launching a new research-grants program, and we’re rolling out a new website and X account. We’ll be conducting more research, publishing more work from more people, and building a much broader community around preserving human liberty in an algorithmic world. It’s going to be epic.

A narrow window to shape political AI

I believe right now is a special and narrow window of opportunity to shape how AI transforms our politics.

AI is rapidly becoming an intermediary between citizens and politics. It will increasingly shape how we find political information, what we believe to be true, how our preferences are translated into choices, and eventually how actions are taken on our behalf.

This could go very well. We could build what I call political superintelligence: AI systems that help citizens understand the world, represent their values, improve the rules that govern them, and make governments more capable and accountable.

But political superintelligence will not necessarily emerge naturally just from making models smarter. A model can be extraordinarily capable and still rely on a distorted information environment, misunderstand what its user wants, defer to its creator, or help a powerful person build systems that concentrate control. Intelligence alone does not produce liberty. We have to deliberately shape the models, institutions, and public infrastructure around it.

For the past year, that is what Free Systems has been trying to understand. Here’s some of the work we’ve produced::

• Creating more trustworthy political advice: We built the largest independent, human-judged study of political slant in major AI models (covered in WaPo) and found that voters across parties preferred less-slanted answers to ideological personalization. In a separate study of 36,300 synthetic Japanese voters, models based voting recommendations far more on policy preferences than demographics - this clip of me summarizing our findings on MTS went surprisingly viral). Together, the projects deepened our understanding of how to build more trustworthy political AI.

• Understanding the emerging politics of AI: Our analysis of 25,000 TikTok and YouTube videos found that AI adoption content outnumbered resistance by roughly three to one, while a study of 280,000 fundraising emails showed AI beginning to enter Democratic rhetoric through an anti-billionaire and anti-oligarchy frame. Together, they offer both top-down and bottom-up insight into how AI politics is developing.

• Building a better political-information layer: Our agentic forecasting system beat the markets in calling the Texas primaries, and we built Bellwether to test whether prediction markets are liquid and manipulation-resistant enough to cite publicly (Bloomberg has since used our data for excellent reporting of their own). The goal is better public infrastructure for political information.

• Testing how AI can write better rules: We trained AI on 10,000 prediction-market contracts and found it could predict disputed contracts roughly three times out of four, while identifying recurring problems in their wording. We also built AI delegates that learned how their principals wanted to vote and placed them in an experimental legislature. The broader opportunity is to use AI to improve contracts, laws, treaties, and policies before they fail.

• Towards living ‘100x’ policy research: We asked an AI agent to replicate and extend one of my published papers. In under an hour and for roughly $10, it produced estimates remarkably close to an independent human researcher. This points toward continuously updating research, routine automated replication, and much faster policy inquiry.

It’s wild to look back and see how much we’ve been able to accomplish in such a short period of time. Our work has reached the labs, policymakers, journalists, and other researchers, and AI itself is letting us move much faster than traditional research models would allow. But the biggest lesson is how much more there is to understand. Have we made some useful progress? Yes. Does it feel like we know everything we need to know to build political superintelligence? Not even close.

Why I’m joining Anthropic

We need to understand more about how AI will shape politics and about what the key levers are to design AI that will help us get better at politics so that we can stay free. As we’ve seen our work taken up by the labs, and as I’ve gotten to speak with people at the labs more, it’s become clear to me that there is a lot of value in operating where the models are being trained, evaluated, and turned into systems people will actually use.

And that is why I’m joining the Rule of Law team at The Anthropic Institute, which was set up to investigate AI’s impact on the world.

I’ll be conducting research on how we can design AI and AI agents that strengthen democratic self-government. I’ll be running experiments, analyzing large-scale data, and studying what makes AI trustworthy and useful for citizens, politicians, and policymakers.

I’m excited to bring what we have learned at Free Systems into that work. But I am equally convinced that the world needs a much broader ecosystem outside the labs: people who can ask questions the labs may not think to ask, test models across companies, monitor their public effects, build open evaluations, and develop a wider understanding of how AI is changing political life. That is the role Free Systems will help play.

The next Free Systems research agenda

To do this, Free Systems will continue conducting ambitious external research on AI and the future of liberty. We have some very exciting research in the pipeline, rough along four big themes:

• AI and the political information environment: We are studying AI voting advice in the US midterms and Brazil’s election; running an agentic political advertising agency whose ads are already appearing on Instagram; and mapping the networks behind positive and negative online narratives about AI.

• Political pluralism, power, and institutional design: We are fine-tuning models to hold different political views; extending the Dictatorship Eval; comparing cabinets, courts, juries, and peer review in our LLM Council; and drawing lessons from blockchain governance for AI-mediated institutions.

• Agents acting in the economy: We are exploring how agents can safely enter agreements on a user’s behalf, and beginning to conduct field experiments on live AI-agent marketplaces.

• Public accountability and better AI-assisted research: We are releasing a Model Card Explorer to help the public understand benchmark capabilities better; testing an automated political-science reviewer on 146 papers with deliberately injected errors; and measuring whether AI research assistants resist p-hacking and other bad research practices.

In addition to this work, we will also be publishing a series of fascinating guest posts from other AI experts in the coming weeks and months. The goal is to turn Free Systems into a broader platform for serious, fast-moving research on political superintelligence.

Three new members of the Free Systems faculty

To move faster and do more, we are bringing on three new “members of the Free Systems faculty.” This is our term for senior academic collaborators who will work directly with our research fellows, develop and oversee projects, and publish their own work through Free Systems.

Alex Fouirnaies is an associate professor at the University of Chicago’s Harris School of Public Policy. He studies the political economy of elections, particularly how money and media shape representation and accountability, using causal inference, applied econometrics, and natural experiments. At Free Systems, he will help us apply that institutional lens to a world in which AI agents can persuade, negotiate, and exercise power on people’s behalf.

Sandy Handan-Nader is an assistant professor of politics at NYU. Her research uses machine learning and applied statistics to measure political conflict and polarization in American institutions, including new work on extremism, electoral safety, and partisan messaging in the House of Representatives. Sandy brings an incredibly valuable skill-set: substantive knowledge of political representation and the technical ability to construct new measurements for the AI era.

Dan Thompson is an associate professor of political science at UCLA who studies American politics, political methodology, and electoral accountability—especially how elections shape policy in local government. His work combines newly assembled election data, administrative records, and modern causal-inference methods. Dan has already been an important collaborator on our research into prediction markets and AI usage among the American electorate, and he is helping lead our work on AI voting advice in the 2026 midterms.

I couldn’t be more thrilled to welcome all three of them to Free Systems! They are going to do amazing work.

Our first lab manager

But how are we going to do all of this new stuff?? We need someone focused every day on turning this vision into a functioning research institution: guiding fellows, pushing projects forward, improving how we communicate findings, building partnerships, and helping the work reach researchers, policymakers, technologists, and the public.

We have found the perfect person in James Campsie, who is joining us as the first Free Systems lab manager.

James has already played a central role in shaping what Free Systems is becoming. I’ve been working with him on the Substack since before it even launched, during which time he has proven to be an incredible thinker and writer with his finger very on the pulse of what is happening in the AI world, what people in industry and policy are thinking about, and what kinds of ideas and evidence will resonate. We wouldn’t have gotten anywhere without James, so I know he’s the perfect person to move into this full-time role.

Free Systems is going to be dramatically better because he is running it with me.

Subscriber-funded research grants

Curating a group of faculty and guest contributors is a good way to broaden our work, but we want to scale even more—more than we can accomplish through direct collaborations alone. So we are also preparing to launch a series of Free Systems API-token grants.

As I promised when we introduced paid subscriptions, all subscription fees from Free Systems go into a community fund. One of the first major uses of that fund will be helping other researchers build on our work.

The initial grants will support applicants who want to extend, replicate, improve, or challenge research we have already published. Someone might test the Dictatorship Eval on new models and languages, build stronger defenses for AI delegates, replicate our political-advice work in another election, or find a better way to measure whether agents actually represent their principals.

We want Free Systems research to be open to productive forks that other people can inspect, contest, and improve.

If you are a paid subscriber, thank you. You are directly funding this work and the broader research community around it.

A new home for Free Systems

Finally, we are officially launching our new website and X account.

While the Substack will remain our main communications channel, the website is a cleaner landing page which brings together our research, researchers, tools, datasets, and opportunities to get involved. The X account will make it easier to follow new findings from across the lab. You can find the new website [here] and follow Free Systems on X [here]. If collaborating with Free Systems sounds interesting to you, please reach out!

Building inside and outside

The next 1-2 years could determine the path that AI and our political system takes. We could veer into new kinds of AI-powered authoritarianism, or we could find ways to strengthen and reimagine our system of democratic self-governance. Taking the good path requires an all-hands-on-deck approach. We need the frontier labs to lead the way in analyzing data, running experiments, and training AI to promote human liberty, and we need a broader ecosystem of AI-informed experts studying how those systems affect institutions, political life, and society in the real world.

This is the next phase of Free Systems. We are just getting started.

Andy Hall studies the political economy of superintelligence at The Anthropic Institute. He is on leave from Stanford University, where he is the Davies Family Professor of Political Economy and a Senior Fellow at the Hoover Institution. His posts on the Free Systems Substack reflect only his personal views and not those of his employer.

Politicians are adopting AI

As part of my research agenda on building political superintelligence, I am constantly looking for data on how people are using AI for politics. A couple of weeks ago, I decided to try something new: I set up a daily task where Claude Code searches the web for relevant datasets and suggests ways to analyze the data with a Free Systems lens. I’ve been blown away by the ideas Claude has been coming up with. Here’s one.

Claude noticed that FEC disbursement data can track how campaigns are spending money on AI. There are a bunch of limitations, here–-expenses under $200 don’t have to be itemized (but can be), campaigns might be paying for AI in ways we don’t see flagged in the data, etc—but the patterns are super interesting even so! Here are the patterns, all suggested by Claude with only minimal iteration by me.

Campaigns are embracing AI—especially Claude and ChatGPT

Here are the most “AI-Pilled” campaigns

Cornel West, leading the AI charge! His campaign is using both Descript (a video and audio editing tool) and ChatGPT.

Progressives actually lead the pack in AI usage

AI usage is fairly spread across the ideological spectrum, but interestingly, the most progressive Dems, who are sometimes the most AI-skeptical politicians, actually seem to use it in their campaigns at the highest rate of anyone.

These are fascinating patterns, and I’m extremely impressed that Claude came up with the idea automagically. Back in January, when I had Claude Code extend my vote-by-mail paper, the biggest limitation we flagged was that Claude demonstrated little creativity or insight in developing new analyses. Now, seven months later, Claude (and presumably other models, too) is starting to show signs of being able to propose new research ideas. I’m not sure how much of this improvement comes from the models getting better vs. me getting better at prompting (in particular, the idea to have Claude look for specific datasets seems to have helped get traction), but either way it’s extremely exciting.

AI will make political science better

Finding and analyzing good data is not the only way AI is going to improve political science. It can also help us improve how rigorously we understand and interpret ideas and arguments. Here is a good example of this, courtesy of my friend Ben Golub. It turns out, the Stanford Encyclopedia of Philosophy’s new entry on Capitalism is kind of terrible, and AI is very much able to explain why.

Obviously, there are lots of things we need to improve about how AI handles information, thinks about politics, etc. But the status quo isn’t perfect, either. Humans make tons of mistakes, are guided by strong ideological biases, and so forth. AI, even imperfect today, is already helping.

Giving legislators superintelligence

And just as it helps researchers and campaigners, AI can also help legislators do their actual jobs! I liked this thread from Séb Krier thinking about the use of AI for lawmakers.

I’ve spoken to a number of state and federal legislators about how they use AI. One member of Congress raved to me about how useful AI already is for helping craft legislation. I wrote down some quick thoughts on why I’m so bullish on this:

Obviously, we will need to work hard to mitigate the various ideological and intellectual biases that AI models have, but the opportunity is huge. And I’m very excited to be launching some new research soon on how we can teach AI to be an even better assistant to legislators.

Talking about Political Superintelligence

Recently I went on my favorite live tech show, MTS, where we covered the new Free Systems research on how Americans are talking to Claude about politics, and we had a wide-ranging conversation about the future of AI, politics, and governance. Check it out here:

“It was probably only a matter of time before voters began to use artificial intelligence to help guide their choices. The 2026 midterms may be the first American elections in which voters are using A.I. in meaningful numbers.

Voters are turning to new A.I. tools to serve as nonpartisan researchers, viewing them as a viable alternative to traditional news coverage, voter guides or social media. They provide an appealing and seemingly efficient way to learn about campaigns and ballot measures, allowing users to bypass the sometimes dizzying array of political literature, advertising and commentary coming their way.”

These are interesting stories, but we know remarkably little about how representative this is of what people in general are using AI for when it comes to politics. Existing data doesn’t help much. Surveys tell us half of American adults now use AI chatbots, but pollsters rarely ask about political usage.

Google’s recent ATLAS report, which I covered, doesn’t disaggregate the data sufficiently to see specific political use cases—though it does indicate fascinating patterns in how people are using Gemini to help them with governmental tasks like filing taxes and procuring local licenses.

In this post, we offer the first real look at how Americans are using AI for politics. To do this, we use new data that Anthropic released from its Economic Index, built from real Claude conversations in April and May, with topics detailed enough to isolate political usage and geography detailed enough to track it state by state.

Because some states held their primaries in May during the study window while others did not, we have a “natural experiment” we can use—isolating the effect that a state-level election has on Claude conversations about politics in that state.

We find four main things:

• Political AI use remains modest, but it’s still meaningful. Politics makes up about half a percent of Claude conversations in the U.S.—ranking 66 out of 189 topics, right next to home furnishing.

• People are not yet getting political advice from AI. Political conversations overwhelmingly produce explanations and analysis, and only 3% end in advice or a recommendation which is markedly lower than other categories of conversation on Claude.

• Elections drive a meaningful uptick in political usage. States holding May primaries saw their share of political conversations jump about 20%, an effect that holds up to a battery of statistical checks.

• November could set a record. If primaries can lift political AI use this much, November’s midterm elections will likely bring the biggest wave of political AI use on record.

The first systematic data on AI for politics

The Anthropic Economic Index is a large-scale dataset that captures how real users are using a frontier model in the wild. Anthropic continually samples from all of the real interactions with their AI tools, anonymizes them, then labels a number of interesting dimensions including the subject of the request, the primary output the AI produces (e.g., an answer to a question, a document, or a script), the occupation associated with the conversation, the date and hour of the conversation, and the country and sub-country region of the user (state in the U.S.).

They release a public version of this dataset that includes only aggregates like the share of conversations that are primarily about each of nearly 1,200 topics broken out by month and sub-country region. They limit the release to cells that contain enough data that they are confident conversations cannot be reidentified.

Just last month, they released a new public version of this dataset covering conversations held in April and May of this year. Its ability to shed light on the contours of AI’s impact on our work and jobs has been well-noted—maybe less appreciated, though, it’s also a data goldmine full of interesting insights about the way Americans are using Claude to think about politics.

At Free Systems, we’re obsessed with studying what we call political superintelligence—how democracy works for a world in which everyone has access to cheap intelligence that can help them navigate politics. We’re also all about using AI to help us do cutting-edge empirical work. So this new data couldn’t be more perfect for us. It gives us the chance to measure, in almost real time, how Americans are starting to use AI for politics. So let’s dive in!

How much do people talk to Claude about politics?

As expected, the top requests to Claude in the U.S. are work- and school- related: business operations, homework, promotional writing, and starting a business are the top request subjects, each making up more than 3.5% of all Claude conversations and adding up to over 18% of all conversations combined.

Politics and public record is the 66th request topic by share of conversations out of a total of 189 topics. This places it just behind photo editing and API debugging; tied with home furnishing, home tech setup, and entertainment; and just ahead of AI agent design and vehicle-related requests.

Politics is evidently only a modest share of total conversations, but that doesn’t imply it’s unimportant. During the long fight over social media and democracy, Facebook executives (during the time I was an advisor there) often pointed out that political content made up only about 6% of what users saw in their feeds. That was an important empirical point to make, but at the same time, many people felt that technology that touches billions of people can shape elections, movements, and civic understanding even through a thin slice of its total activity. Especially if that activity is particularly influential or emotive. The same logic applies here: half a percent of an enormous and fast-growing stream of AI conversations is still a very large number of Americans learning about their government, and these could be the kinds of conversations that stick with users far more than the typical one.

Users ask Claude for political information, not recommendations or actions

When people do talk to Claude about politics, what do they talk about?

Most political requests to Claude today are seeking information—Claude responds to the majority of political requests with an answer or analysis, not code, a product, a document, or even advice. Nearly 50% of political requests result in an explanation, while less than 20% of all non-politics requests do so. While more than 20% of political conversations result in an analysis or summary, only about 5% of non-political conversations do so.

And, despite the potential for AI to guide voters on whom to vote for (I’ve explored the promises and limitations of AI’s ability to do so here), currently, only 3% of political conversations end in advice or a recommendation.

The vast majority of requests related to politics, public record, or elections result in some kind of information, rather than producing an action or a document. Here’s an example of what that might look like—I asked Claude to explain to me Mayor Mamdani’s recent grocery store experiment in New York City, and I got back information right in the chat, with no other artifact or action.

Interestingly, though, for the subset of requests that are about government filings, Claude prepares a document or takes an action roughly 50% of the time! Similar to what we saw in Google’s analysis of their ATLAS data, this suggests an exciting and not-so-distant future in which AI agents create and file documents on behalf of citizens—but it’s a rare action, for now.

Here’s what such a task might look like. Here, I asked Claude to help me create business formation paperwork—and it did!

Elections drive a meaningful uptick in political AI usage

If people are turning to AI primarily to learn about politics, what kind of information are they looking for? In our representative democracy, one of the most natural things citizens might be looking for is information on who is running in an election, what they stand for, and how they’ve performed.

While we can’t observe those fine-grained categories directly, we can use the fact that different states hold primaries at different times of year to learn how much people are using Claude to learn about elections.

Eleven states held Congressional primary elections in May. Eight of those states had sufficient political AI conversation volume for the Economic Index to report it in both April and May. You can see those states in dark blue in the figure below. If substantial numbers of citizens are turning to Claude for information about elections, we should see an increase in political conversations in these eight states in May, and that’s exactly what the data reveals.

In the eight states with a May primary, the political request share rose from an average of approximately 0.4% in April to 0.5% in May, an approximately 25% proportional increase in the share of political conversations.

With only eight states, this “difference-in-differences” design could confuse a statistical fluke for a real effect—in the Statistical Appendix at the bottom of the piece, we look into this. We find no similar increase for related non-election topics, and we also find that the increase in the political topic is a huge outlier compared to estimated changes in other topics—all suggesting that the effect is probably genuine.

This is a large change in the share of AI conversations about politics. Just how large? Over the same April to May period, another big shift in AI use happened—the dramatic drop in the demand for AI tax advice. In April, 0.34% of conversations with Claude were about taxes. In May, that number dropped to 0.14%. So, this 0.2 percentage-point drop for taxes is about twice as large as the effect of a primary on the share of requests about politics.

Put another way: the increase in information-seeking behavior caused by an election is about half as large as the increase caused by the need to file taxes.

Expect a large increase in political AI usage this October/November

Taxes are part of everyone’s life. Primaries generally only attract the most engaged citizens. In general elections, we’d expect much broader participation and perhaps an even larger effect, possibly rivaling the effect that the tax deadline had on the share of conversations about taxes.

To get a sense for where we might be headed this fall, look at Oregon. Oregon held its primary in May and had unusually high participation—approximately 37% of Oregon adults voted in the May primary, the highest turnout of all the states holding May primaries. Oregon also exhibited the biggest increase in political queries in Claude from April to May.

In November, we can expect between 40% and 45% of U.S. adults to participate if history is any guide, pretty similar to Oregon’s turnout in May. So, crudely speaking, we might expect national queries about politics to increase by about the same amount Oregon did. That would put politics at about 0.7% of Claude queries, and move its ranking from 66th to 41st—roughly the same frequency of conversations around booking and scheduling.

Conclusion

What does this early data tell us about how Americans are using AI for politics? Americans are talking to Claude about politics—overall rates are not super high, but patterns of intention suggest these conversations are important. When it comes to politics, Americans are seeking information from Claude, rather than advice or direct recommendations. When an election is upcoming, they seek out more information. Clearly, there is an important set of voters for whom Claude is an important helper. As we build towards political superintelligence, this is a promising foundation we can start from.

The next question, of course, is: when Americans ask Claude or other AI helpers about politics, do they get back good answers? To understand this, we’ll need to go beyond the coarse-grained data currently available. We need to understand the precise political topics Americans are seeking help with, and the nature of their requests. Then, we’ll need to develop notions of what a “good” answer looks like, so that we can evaluate models against these, and then train them to do better. That’s how we can continue working towards political superintelligence.

Statistical Appendix

Looking for signs of a statistical fluke

The increase we observe in states holding May primary elections could be a fluke—we don’t have a large sample size of states to work with, after all. We run a number of additional tests to try to address this concern.

First, we compare the increase in political conversations from April to May in the eight May-primary states to the change in political conversations over the same period in 22 states with primary elections held after May.

These are the 22 states with light blue dots and arrows in the figure above. The share of political conversations in those 22 states barely changes from April to May, suggesting that the change in the May-primary states is due to something special about those states rather than a nationwide change in the share of political conversations with Claude.

Using these 22 states as control states in a difference-in-differences design presented in the figure below, we estimate that primaries increased the share of political conversations by approximately 0.1 percentage points or 20% in May. Removing states with June primaries from the control group did not meaningfully change our estimate.

Reassuringly, we do not see similar increases in conversations about geopolitics, personal finance, news writing, or news aggregation in the May-primary states compared to the later primary states—the effect really does seem to be concentrated in the political category, where we would expect it to be if it’s about the election itself.

Finally, if our results are real, it should be unusual for May-primary states to have larger changes in non-political categories when compared to post-May-primary states. We compute simple differences-in-differences estimates of the effect of primaries on the share of conversations about each of the 210 topics are data allow us to study. We standardize the estimates into t statistics where more positive means May-primary states saw increases in the share of AI conversations about that topic more than we would expect based on national trends. Political conversations stand out. Only two topics out of 210 increased more in May-primary states. This suggests that this is very unlikely to have happened by chance—the primaries appear to have caused a substantial increase in the share of AI conversations about politics.

It is also worth pointing out that these are probably underestimates of the amount of political conversations driven by primary elections. Using Google Trends data, we can see that search volume about primary elections was already substantially elevated in these eight May-primary states in April. Searches for “primary election” tracked very closely in the eight states with May primaries and the 22 states with post-May primaries, but these trends diverged substantially in April then again even more in May. If the patterns in political AI use track the patterns in election-related Google searches, the April to May change in political conversation share meaningfully understates the number of conversations people are having with AI about their primary elections.

Unsolved math problems are falling to AI at a rapid pace—here’s just the latest big news, this time from OpenAI:

These are open problems in math of a particular form, and crucially, solving them is verifiable. AI seems poised to autonomously solve a very wide array of problems like these, where it can work on them and verify for itself if it has gotten the answer right or not.

Sadly, most of the big questions of political economy don’t work this way. AI can still help us with them, for sure, but we probably won’t see the same kind of rapid progress in the study of politics as we’re seeing right now in math and coding. But I’ll still be trying.

A global pause on open letters?

The policy debate around AI is a good example of this challenge—solving AI policy is not the kind of verifiable problem AI can solve for us, so we have to avail ourselves of other, less reliable political methods.

In the past several weeks, we’ve seen a positive Cambrian explosion of “open letters” about AI—starting with the economists’ letter on why we should study AI more urgently, followed by NVIDIA’s open letter on the value of open-weights models, and then the lab researchers’ letter on building the infrastructure to facilitate a global pause on AI development.

I have a great deal of respect for the organizers and participants of all three letters, and it is great to see so much engagement on such important issues. I do, though, have some doubts about whether the open letter is the right way to drive forward this policy conversation…

The basic problem with open letters is that they risk devolving into signaling exercises that encourage what calls “preference falsification.” When every policy debate becomes a quantitative exercise in toting up who signed and didn’t sign, I’m not sure it helps us get closer to the truth using reason and evidence. This was not the intent of any of the open letter authors—and I’m not saying that open letters are always a bad idea—but I worry that this is where this kind of raft of letter-writing could head.

We have long historical experience with open letters, and I’m not sure the history tells us they’re good. I just listened to Joanne Paul’s phenomenal biography of Thomas More, and I’d forgotten what a crucial role “open letters” played throughout the Reformation. I’m not drawing too close of an analogy between the Reformation and our current AI policy debates (ha), but in general I think it’s bad to create political discussions that focus on who’s in which camp, and in pressuring people to join one camp or the other—and open letters, I fear, tend to have this effect. One way to see this is to watch X each time a new open letter is released. It is always followed by a somewhat tedious raft of people posting their explanations for why they did or didn’t sign.

Progress towards the 100x Research Institute

We may not be able to unleash frontier models on verifiable “open problems” in political economy, but we can still do a lot of exciting stuff with AI. In the vision I sketched for the “100x Research Institute,” I envisioned the new kinds of research that AI enables.

The Roots of Progress

AI is already 10x-ing academic research. How do we get to 100x?

“Intelligence Age” is a series from the Roots of Progress Institute featuring reported essays that extrapolate the capabilities of AI systems along current trend lines…

Read more

5 months ago · 125 likes · 15 comments · Andy Hall

One of those new kinds of research I called “living research”---research projects that update as new data comes in, instead of being entombed in amber inside a static PDF that becomes rapidly obsolete.

I’m not just talking about this new kind of research, I’m doing it! A few weeks back, I wrote a piece tracking the rise of anti-billionaire populism among Democratic politicians, and how it is slowly transferring into their views on AI.

Now, I’ve released a live dashboard that updates automagically each week as Derek Willis adds new emails to his open database. You can check it out here:

https://inbox.freesystems.net/

Democrats are more worried about AI taking their jobs

Dan Hopkins, distinguished political scientist at Penn, has some new research out on his Substack about how Americans are viewing AI and the economy. It’s really interesting!

One key finding jumped out at me: Democrats are substantially more concerned about AI taking over jobs than Republicans are. This could have to do with general attitudes/vibes around AI, or it might well reflect the different kinds of jobs that Democrats vs. Republicans tend to hold. Either way, the growth of this kind of polarization is well worth watching, and fits with the polarization I’ve seen in the fundraising emails, and which people are finding in lots of other places as well.

Manipulation in prediction markets

2026 is not only going to be the first big “AI election,” it’s also going to be yet another watershed in the astonishing rise of prediction markets and politics, as I’ve been tracking for a while.

Liam Vaughan has a fascinating deep dive out this week at Bloomberg looking into the different ways people try to manipulate prediction markets around elections—usually unsuccessfully, but still something we should be thinking about. The piece references several Free Systems research pieces, including https://freesystems.substack.com/p/prediction-markets-are-eating-american and https://freesystems.substack.com/p/bellwether-building-trust-in-prediction.

Tweet of the week

The inexorable march towards some sort of industry-led self governance around model safety and security continues…

1. OpenAI disclosed that two of its models escaped a controlled evaluation environment and hacked Hugging Face (a major AI model platform). And then, separately…

2. A large coalition of tech companies signed an open letter arguing that powerful open-weight models are essential for competition, innovation, and cybersecurity.

These two events are more connected—and indeed, more contradictory—than some realize. As AI agents start to run amok on the web, like they did in this major Hugging Face incident, someone is going to have to be liable. Someone will be asked to make the victim whole.

In this case, that’s straightforward enough—eventually—because the offending agents came directly from an identifiable model company. Even so, it took ten days: Hugging Face disclosed the breach without knowing who was behind it, went to the FBI, and only then did OpenAI come forward and say the agents were its own. The two companies are now working through remediation together.

Now imagine a future case—probably coming in the next several months—where an agent runs similarly amok and causes damage inside an important system, and it’s not from a major model company, but rather comes from an open-weight Chinese model hosted somewhere on the anonymous web. Who will the victim demand restitution from? If the victim can’t identify the operator, they’ll be sorely tempted to go and yell at the company that developed the model in the first place. And that pressure could make companies much less willing to release their most powerful models openly.

I have no idea how this is going to play out, but as I said last week, it seems like we have very little idea how to trade off the economic value to open models and the potential safety issues. We don’t even seem to understand the safety issues—with some people claiming that open models improve safety by giving companies the ability to probe and strengthen their defenses (Hugging Face ended up using a Chinese open model, GLM-5.2, to play defense, apparently after a closed US model’s guardrails got in the way), while others claim they are better for offense than defense.

Basically, it feels like we’re flying blind. Right now in tech and on X the vibes have very much shifted towards open models, but without any clear answer to the safety questions, we’re a long way from knowing how this will or should play out.

Political superintelligence is proceeding

As I wrote on X:

I’d love to see follow-up research that goes deeper into these categories!

The economist covers Free Systems

The billionaire backlash is getting lots of attention, and The Economist has a great piece on who these billionaires really are. The piece starts by talking about my evidence on anti-billionaire sentiment in fundraising emails, and also offers fascinating original data showing that there has been a similar spike in anti-billionaire speech on the floor of Congress. Very cool stuff!

Universities should arm students with AI eval powers

I was thrilled to publish an op ed in the FT elaborating on my idea that every university student should learn how to build AI evals so that they can hold AI models accountable to their own values.

New research!

Two interesting new papers came across the Free Systems transom this week:

• Luke Hewitt, Maximilian Kroner Dale, and Paul de Font-Reaulx, “DeliberationBench.” Introduces a benchmark for evaluating AI persuasion by comparing how LLMs change users’ opinions to how people’s views change after participating in Stanford-style deliberative polls. Across 4,088 participants, frontier models produced opinion shifts that broadly tracked those from human deliberation, though without the same depolarizing effects. Paper: https://arxiv.org/abs/2603.10018

• Lennart Finke and Stephen Casper, “Corporate Loyalty: Some AI Systems Differentially Downplay their Creators’ Controversies.” A preregistered study of 21 models finds that models from xAI, DeepSeek, Anthropic, and OpenAI discuss controversies involving their own companies more favorably than comparable controversies involving competitors, highlighting the need for evaluations of conflicts of interest in AI systems. Paper: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7059338

Tweet of the week

When I envision AI ushering in a new era of political superintelligence where our legislators use powerful new technology to do a better job of governing, this isn’t exactly what I had in mind…

This week saw two massive developments for how we will govern superintelligence. First, Demis Hassabis, founder and leader of DeepMind, released a proposal for an independent self-regulatory body for frontier AI.

This is exactly what I argued for in April, where I specifically used FINRA as an example. I think it’s a great idea, and it was extremely promising to see leaders from other labs post publicly in support of the idea.

Here’s part of what I wrote back then:

The second massive event of the week raises big questions about that “broad participation” that I flagged as a requirement.

I’m talking, of course, about the seismic release of the extremely impressive new open-source Chinese model Kimi K3.

We will know more in the coming weeks—especially after the underlying weights are released next week—but K3 appears to be nearly as powerful as Anthropic’s Fable for a range of coding tasks, and it has none of the guardrails that Fable has.

In a nutshell: after all the work the Trump administration did to make Anthropic and OpenAI jump through hurdles to release their models, we now have Chinese open-weight models that are approximately as “dangerous” as those models and don’t have to deal with any of the same hurdles.

Right now, it’s very hard to tell where this is all heading. One possibility is that open-weight models continue to nip at the heels of the frontier models, improve relentlessly, and render all safety precautions hopeless and indeed counterproductive for American models.

Another possibility is that we haven’t yet seen truly dangerous models—and once we do, China, too, will become reticent to release them to the world without guardrails. Here’s what Xi Jinping said about AI this week in an important speech:

This came right after a passage in which he emphasized the value of openness in AI—so clearly he sees a possible balance between the value of openness, on the one hand, and security, on the other. This seems to me to leave open the possibility that China might cooperate on model safety issues if they prove to become severe in the future. Maybe we’re just not there yet.

Having listened to a lot of the dialogue around these questions, I genuinely feel I have no idea how it’s going to play out or what the right answer is. The safety people shout about the existential risk of the “race to the bottom” dynamics of the model arms race and crave a way to lock down if not pause the models; meanwhile, the innovation crowd cheers on the open models and points to the ways that closed models with aggressive regulatory moats could chill progress and concentrate economic returns uncompetitively.

Both seem possibly right, and the two groups seem to be talking past one another. How real are the safety issues, and do they justify limiting open weight models? How real are the economic innovation issues, and do they justify letting open weight models rip? These are the questions we have to answer.

The free systems flywheel is turning

We are living in such an exciting time for research, between open data and coding agents, as I keep saying. One of the most exciting developments at Free Systems has been that people are increasingly reaching out with their own new research they’ve done that connects to my own, and I’m thrilled to use System Check as a way to share this research more widely.

Here are two really interesting pieces that came across my desk this week.

Measuring power concentration in the AI stack

Tons of people worry about how AI might concentrate power in the future. Maybe that power concentrates in a few frontier labs, or maybe it flows to the hyperscalers, or maybe to the energy providers. It’s hard to predict where the chokepoints may turn out to be in the end.

Piyush Akimitsu has been corresponding with me for a few weeks about this, kicking around an idea we’ve both had to try to measure power concentration across the entire AI stack. Here’s a really cool first stab he put together! The “AI Stack Map” looks for how concentrated all the different points in the pipeline are. Did you know that China controls essentially all Gallium production? Or that the DRC mines three quarters of the world’s cobalt? Fascinating stuff.

Tracking candidate positions on AI

I’ve been looking at how politicians are starting to use AI as a political issue, tracking their fundraising emails as well as the bills they’ve introduced.

Ethan Jiang had a complementary idea—what if we could see what candidates are saying about AI on their candidate webpages? So that’s what he did. His “AI Primary Tracker” is a fascinating dataset looking at the content of candidate webpages when they talk about AI.

There seem to be some interesting similarities and differences to the fundraising email data. There is a similar partisan split in terms of mood, with Democrats more sour on AI than Republicans. But the policy positions that Democratic candidates focus on on their websites tend to be about safety and job displacement—whereas the emails tend to more inflammatory and focus more on broader anti-billionaire sentiment.

Tweet of the Week

I’ve been deeply enjoying ’s coverage of tech recently—it’s become an essential part of my tech news diet. I really liked the segment with Jonathan Slotkin that dove into how the plaintiffs’ bar opposes driverless cars, allegedly because they make such good business off of litigating road accidents and don’t want technology to reduce the rate of these accidents.

The moment I saw it I knew it would be ideal ragebait for and I was right! He posted about it this weekend.

There are broader lessons here for AI diffusion. Just because a technology might make us on net better off doesn’t mean there won’t be concentrated interests who benefit from blocking or distorting it in various ways.

“Whence does it come to pass that our common language, so easy for all other uses, becomes obscure and unintelligible in wills and contracts? and that he who so clearly expresses himself in whatever else he speaks or writes, cannot find in these any way of declaring himself that does not fall into doubt and contradiction?”

–Michel de Montaigne, “Of Experience” (1587)

In 1863, William Winter Raffles, a Liverpool cotton merchant, agreed to sell 125 bales of Surat cotton to the Liverpool trading firm of Daniel Wichelhaus and Gustav Busch. The written contract specified that the cotton would “arrive ex Peerless from Bombay.”

But there was a problem. Unbeknownst to both parties—but beknownst to posterity and to many 1L Contract Law students—there were two different ships named Peerless sailing from Bombay to Liverpool that season!

Raffles intended to sell the cotton arriving on the Peerless that departed Bombay in December, while Wichelhaus and Busch meant to receive their order from the Peerless that had departed in October. When Raffles prepared to deliver cotton from the December ship, Wichelhaus and Busch refused to accept it, arguing that they had contracted for the October shipment.

Raffles sued for breach of contract, but the Court of Exchequer held that the contract failed because the phrase “ex Peerless” was latently ambiguous: each side had attached a different, equally reasonable meaning to the same words, so there had never been a true agreement. If only the parties had written the contract to uniquely identify the ship—by specifying its departure date, or using the unique identifier each ship was assigned—the whole dispute could have been avoided, as A.W. Brian Simpsons explains in his magnum opus on the history of the case.

Unanticipated problems of wording, like this one, happen all the time. Just recently, ambiguous wording in America’s “Memorandum of Understanding” with Iran threatened to break a fragile cease fire. When I was an advisor at Meta, the issue cropped up constantly—teams spent enormous effort to write the rules that governed content moderation, only to then face pieces of public content that intentionally or accidentally exposed loopholes or gaps. From local and national politics to private contracts and online platform policies, the issue of writing clear and unambiguous rules is a perennial challenge to all forms of governance.

I’ve been working on what I call political superintelligence—the idea that, as AI makes intelligence more cheaply available, we can use it to make our democracy function better. Could AI help us to avoid some of the ambiguities in important written agreements? Writing better agreements won’t solve all of the challenges of governance—even perfectly worded contracts won’t anticipate major future shifts, obviously, or solve the challenges of imperfect enforcement—but it could certainly improve the way we govern our societies, our private business relationships, our online platforms, and much more.

We decided to put this to the test. To start, we needed somewhere we could quickly harvest lots of written contracts and observe which ones ended up in disputes. We didn’t have to look far: prediction markets are a favorite testbed here at Free Systems, and Kalshi and Polymarket publish resolution rules explaining how each contract will be decided. We’ve previously explored the very problem that those rules can be ambiguous, so it presents a perfect opportunity for us to revisit and see if we can do better.

Working with Claude, we developed a 10-point rubric for spotting flaws in contract rules that might lead to disputes—including things like failing to specify a source of truth, not naming an entity specifically enough (like the Peerless!), and so on. Then, we used those 10 scores for each resolution rule to train a model to predict which ones would actually be contested, in practice, using data on actual disputes from both platforms.

It worked surprisingly well! Just from our grades based on the text of the resolution rules, our tool is able to predict which contracts are more likely to result in a dispute at a rate far better than chance.

In the realm of prediction markets, our scores can hopefully be the start of a “grading system” for contracts. And they can also be used to generate specific recommendations on how to design better rules for future contracts.

More broadly, the results point to an exciting way we can use AI to improve governance. Could we unleash powerful AI models to help us improve all kinds of governance and contract language—asking it to find ambiguities, unanticipated issues, miswordings, and more before we release the language into the wild? If done well, the result will be clearer contracts and better outcomes. A small but valuable step in building political superintelligence for the world.

Here’s how we did our study and more about what we learned.

Teaching AI to evaluate contract language

First, the data: we collected a sample of 10,000 contract resolution rules from Kalshi and Polymarket, combined with information on which contracts ultimately wound up disputed. For Polymarket, we turned to UMA’s governance oracle. This is the process by which Polymarket receives and handles disputes, and all the data is openly available online. We divided these disputes into two categories: material disputes, which are those that weren’t dismissed for being premature (a technicality in how UMA works), and confirmed disputes, which are the subset of material disputes that continue through the dispute process and receive a second dispute request (again, a technicality of the way UMA works).

Because Kalshi’s “disputes” are more ambiguous in nature, we focus our main analyses on Polymarket, but patterns are generally similar across the two.

We developed our rubric for grading contract quality by reading a bunch of contract rules ourselves, and conversing with our buddy Claude. First, we came up with ten dimensions that we and Claude thought captured common failures we saw in contracts. Then, we worked iteratively to develop a system of scores from 0-3, where a 0 represents clarity or no concern on a given dimension, and 3 represents severe ambiguity. Last, we put a large sample of the contracts—weighted to include more of the disputed ones, because disputes are pretty rare in prediction markets—through the finalized LLM grader (using GPT-4.1 to save money for now) and got back our final grades.

Here’s how this process plays out for the famous “Zelensky suit” blow-up that we discussed in a previous piece as a good example of how contract rules can go awry. (Note: we worry that the AI could bring in its knowledge of a disputed case when doing the grading in some cases—in the conclusion we discuss this a bit more.)

The most challenging part of improving our prompt was the tradeoff between specificity and generalizability. Giving examples of 0-3 rankings for the LLM to use only worked within the contract type for a specific market, not more broadly. For example, in the first version of the prompt, we tried to enumerate many specific oversight institutions such as the Fed, CDC, FIFA, the Grammys, Spotify charts, ADP, Billboard as examples for the LLM to learn from. But we found that when we enumerated, the LLM anchored to only these examples, failing to develop similar, analogous institutions for contracts in other domains.

To counter this, we removed specific examples and replaced them with generalizable context that would allow the model to reason. Additionally, the LLM scored poorly on Polymarket contracts that resorted to “a census of credible reporting” for resolution. In early iterations of the prompt, the LLM graded these leniently. We changed the prompt to evaluate the weakest resolution path that the contract permitted, meaning we graded the most dangerous resolution sources for each contract.

Results: our AI can predict which prediction-market contracts will get disputed, just from the text of the resolution rules

To see whether our grades predicted disputes, we used Fable in Claude Code to throw a bunch of different machine-learning methods at the problem.

As the table below shows, a standard ML approach, Gradient-boosted trees, turns out to perform the best when tested against “out-of-sample” cases, and it performs quite well—the standard measure of predictive accuracy we use, AUC, roughly gives the rate at which the model can guess which of a pair of cases (one that ended up disputed, one that didn’t) is the one that will go to a dispute. Our ML model is able to do this on 3 out of every 4 tries, roughly, which is much better than chance. It’s not a crystal ball, by any means, but it’s a remarkably strong signal to draw purely from the text of the contract.

The figure shows this accuracy another way. Each point in this plot represents an average of a number of underlying contracts. The x-axis shows the predicted risk of those contracts based on their language; the y-axis shows the percentage of those contracts that ended up disputed. As the plot shows, the contracts predicted to be riskier go to disputes at a much higher rate than those predicted to be less risky.

(Technical note: these are raw probabilities on the sample that is weighted towards disputed cases – in work not shown we evaluate how this accuracy looks like when re-calibrated to a random sample, finding the same kind of pattern.)

What text features predict disputed contracts?

A neat feature of this approach is that we can also extract which features tend to be most predictive of contract disputes, using the rubric that we developed with Claude. Three features really stand out: when the core question is vaguely defined; when it’s unclear which entity counts; and when the settlement source is not identified.

It might be that these are common mistakes made in hastily written contracts, and that platforms should focus on defining questions more clearly, specifying entities more explicitly, and identifying specific settlement sources in all the contracts they write. But it’s also possible that there is a reverse causation here—maybe events that are hard to call and likely to be disputed are also inherently harder to write contracts for. We’ll need to do more research down the line to really understand this, but in the meantime, our model can certainly suggest potential weaknesses in contracts so that the authors can at least consider whether they have ways to address them beforehand or not.

A grading system for prediction-market contracts

To see how our approach can help, consider the prediction market example. Investors often like to consider assets by looking at “grades” of their quality. Using our predicted probabilities of disputes, we can divide the prediction-market contracts into bins reflecting the risk they’ll end up disputed. Contracts graded “A” are the least likely to be disputed; each subsequent lower grade increases the risk of dispute steadily, with CCC contracts 3.4x as likely to be disputed as A contracts.

Here are a few examples of what the different contracts and their grades look like.

From prediction markets to political superintelligence

When Raffles and Wichelhaus signed their cotton deal, their best defense against an ambiguously written contract was the gimlet eye of a human expert—the same method we still rely on today, for everything from the resolution rules for strange prediction-market contracts to the laws that shape society, the rules that govern global online platforms, and business transactions taking place all over the world. But humans, we know, are fallible. Rules get written ambiguously, or fail to anticipate issues that could have been anticipated in advance. AI is fallible, too, but it could have a very useful role to play in helping us improve the way we write rules all across society.

There are a whole host of limitations to our experiment. First, we’ll be subjecting our method to a truly “out of sample” test by having it evaluate contracts that haven’t yet been resolved, in case the grader LLM is somehow bringing in knowledge of contract disputes when creating the grades. We also need to replicate it for more contracts in more domains. And we need to take the further step of having the AI then help re-write contracts to make them better. All of this lies in the future.

But the direction of travel seems clear. The core act of governance—writing rules that bind us, and adjudicating what they mean when reality gets complicated—has depended for all of human history on scarce, expensive, fallible human judgment, and that constraint has quietly shaped what kinds of institutions we could build. AI loosens it. If a model reading resolution rules can spot the disputes coming three times out of four, then the same intelligence, pointed at statutes, platform policies, and treaties, can help us write agreements that fail less often and govern better. That is the promise of political superintelligence, and this is one small, concrete step toward it.

Like Free Systems? Consider becoming a paid subscriber — all proceeds go to supporting Free Systems community events and programs. By subscribing, you help us to further our research, keep all our content free and open, and gain reach on Substack (which only promotes newsletters with paid subscribers.)

Second: we’ve decided to play the Substack game. Chris Dixon’s excellent book (note: I’m an advisor at his fund) Read Write Own explained how online platforms inevitably seek to control their networks. Substack is no exception. The basic idea of Substack was, originally, that you the author own your distribution, because it’s email-based.

But increasingly, Substack encourages distribution through its social feed. When users “follow” you, they get your updates on their Substack feed, but you don’t get their email addresses. This is how Substack will shift from a platform in which the authors own their distribution to one in which Substack owns the distribution. You can’t just rely on emailing your piece to your subscribers; you need Substack to boost it into other users’ feeds.

This then leads to a second implication. Substack makes money by taking a cut on subscription. So they really want authors to charge for subscriptions. If you charge for subscriptions, you can get significant reach—that is, Substack will put your content into people’s feeds and encourage them to subscribe. If you don’t charge, you don’t get this reach.

I have no interest in charging for my content. Indeed, the whole motivation for Free Systems is to get our ideas and research out as broadly as possible.

But I do have interest in getting reach, to further my goal of having impact. So, this past week, I finally caved and added paid subscriptions to Free Systems. And it’s already working! In one week, we’ve already hit 18th in the “Rising in Technology” list on Substack, which is awesome!

Pay for a subscription to boost our reach and fund the community!

Here’s how I’m handling paid subscriptions. I am not gating any of the research – all of my pieces will continue to be completely free to read in full. I’m also not pocketing any of the money from the paid subscriptions. Instead, we’ll put all the subscriptions into a Free Systems community fund. That fund will be used to pay for in-person and online events, grants for students, and other Free Systems programming. For now, you just have my word on this…but I’m also going to work on creating a fun crypto pipeline that takes the subscriptions and puts them into a crypto wallet whose balance and spending subscribers will be able to audit themselves.

So, even if you don’t normally pay for subscriptions, I hope you’ll consider adding one for Free Systems—it will help us climb the leaderboard, get more reach, and fund cool events.

100x research institute – six months in

Within an hour of releasing my new research based on political fundraising emails last week, two people—John Horton and —had already released new analyses based on the data, thanks to the use of coding agents. This is very much in keeping with the vision I laid out in the piece I wrote for Roots of Progress on how we’ll get to the 100x research institute. Part of the dream is having people quickly fork and build on new research! So this was fantastic to see.

Other things haven’t developed as much as I predicted, so far. In January, I claimed that a “freight train” was coming for political science because people would use AI to produce 1,000s of papers, many of them of dubious quality. But weirdly, that doesn’t seem to be happening yet?

It seems to be happening thus far in certain parts of the hard sciences—where paper submissions are up massively—but less so in the social sciences, which is surprising to me.

Tracking the Slopocalypse

In other ways, though, the slopocalypse is well under way. Here’s a great report that Pangram, my favorite AI detector, put out, showing the widespread use of AI-generated text on social media, especially for longer-form content.

I don’t think there’s anything inherently wrong with generating text with AI. And I worry that sometimes Pangram is used to shame people unnecessarily as part of a “gotcha” game, as has argued.

At the same time, there are two reasons the trend is troubling to people right now. First, the quality of the text is often loathsome—indeed, I swear it’s gotten worse rather than better with newer models. I just cannot stand the “Claudeslop” that Fable and Opus now put out for me. Maybe I’m just getting too used to it?

Second, we don’t have time to read all the text put in front of us. So we have to use shortcuts to infer what text is worth reading and what’s not. When we see something that’s obviously AI generated, it signals to some of us, some of the time, that the text was produced without much thought, and isn’t worth our time. I’m not sure that inference will always be true, but it’s a reasonable inference today.

For both of these reasons, I’m relatively cautious in using AI to help me write—but I do use it in some circumstances, and I don’t promise not to!

Tweet of the Week

Continued momentum in the space of building technology to help companies own their own AI, as highlighted by …

Question of the Week

Are you doing interesting new research studying AI, politics, and/or governance? Or new political science/political economy research using AI in innovative ways? Consider emailing it to me! I’ll be making a point to cover new research in System Check, and people have recently started pitching me their new papers which is awesome.

But almost all of the evidence behind this narrative describes the mood of the American electorate, and moods are only half of the story. Politicians do not respond mechanically to shifts in public sentiment—they answer to donors, activists, and primary challengers, not just median voters…and sometimes they stake out new positions well before the public asks for them, or refuse to move long after it has. If we want to know whether AI populism is becoming an organizing position of the Democratic party, rather than a diffuse sentiment floating around in polls, we need to measure the politicians directly. We need to watch what candidates actually say when they are trying to raise money, to garner attention, and to win elections.

Fundraising emails turn out to be a remarkably good place to look. Campaigns test these messages relentlessly against open rates and donations, so the language that survives is the language that operatives have learned actually moves their base—a compressed, high-frequency, almost real-time record of what politicians believe their supporters want to hear. Drawing on Derek Willis’s archive of political fundraising emails, we analyzed roughly 280,000 candidate emails sent since 2017 to trace how anti-billionaire populism became a major component of Democratic fundraising rhetoric, and how, more quietly but unmistakably, AI is starting to follow the same path.

Some quick caveats on the data before we dive in. Not all politicians send fundraising emails, and not all those that do are captured in this data, particularly for periods of time in the past. To remind you of this, I try to use the phrase “email-active members of Congress” where relevant. Moreover, to make our analyses make sense, I subset the data to partisan senders, and the classification process by which I do this is imperfect.

2025: the year Dems’ billionaire-bashing went supernova

Let’s start by simply tracking how often fundraising emails mention billionaires—and if you’re wondering, you can assume that essentially all mentions of billionaires are negative, as I’ve verified along the way.

For most of the past decade, anti-billionaire language was a niche product. Democratic emails invoked billionaires in the mid-single digits through 2017 and 2018, spiked briefly to around 14 percent during the Warren and Sanders primary surge in 2019, and then settled back down—through the entire Biden presidency, the billionaire appeared in roughly one of every twenty to twenty-five Democratic fundraising emails, barely more than in Republican ones.

Then came January 2025. In the weeks after an inauguration that seated tech CEOs in the front row and the dizzying drama of Elon Musk’s ill-fated DOGE experiment, billionaire mentions in Democratic emails quadrupled, peaking above 20 percent of all emails sent and holding around 15 percent ever since. Anti-billionaire fundraising tactics are now a mainstay of Democratic messaging.

To give you some texture, here are some examples of real 2025 emails that major Dem figures sent about billionaires.

Which Democrats lean hardest into this economic populism? To find out, I compared each member of Congress’s emails against how progressive their voting record is. (One wrinkle: the standard way of scoring voting records mistakes some of the most progressive members — like Rashida Tlaib and Ilhan Omar — for moderates, because they sometimes vote “no” on Democratic bills in protest. I use a newer method that corrects for this.)

There’s a clear pattern: although the entire party has a proclivity for bashing billionaires, it’s the left flank of the Democratic party that is mentioning them at the highest rates, on average. But it’s also interesting to see Hakeem Jeffries, the Dems Minority Leader in Congress, looking most like the Progressives in his billionaire bashing, too.

Dems billionaire bashing is fusing into AI-bashing…but only slowly

Even in 2026, with AI dominating the business press and creeping into the polling, only about 0.7 percent of Democratic fundraising emails are substantively about it—less than 1 out of every 100 fundraising emails. As a share of what campaigns actually talk about when they ask for money, AI remains a rounding error. Exactly what this means is tricky. Maybe the AI emails are rare but dramatic. But more likely, it indicates that candidates are not yet focusing on AI as a message for their donors.

Even so, the trajectory is steep: that 0.7 percent is triple the share from a year earlier, which was itself well above the year before, and Republicans are tracing the same curve albeit a step behind. The absolute level says AI populism has not arrived; the rate of growth says it may be coming!

Here’s one way to see why it might be coming. In this figure, I compare the fraction of email-active Democrats who mention billionaires, Elon Musk, or AI at least once in a given year. Even though only 0.7% of all Dem emails in 2026 mention AI, nearly 40% of email-active members of Congress have mentioned it at least once. This is similar to the rate of anti-billionaire rhetoric in 2019, when it was already starting to increase. Could this indicate that AI is on its way to be the new kind of anti-billionaire rhetoric? It’s hard to say, but it certainly seems possible, perhaps in one primary cycle’s time.

Like with the billionaire bashing, the anti-AI rhetoric is not even across the caucus. The most left-wing email-active members of Congress mention AI more than twice as much as the median-ish member of the party, and more than ten times as much as centrist Dems do. Elizabeth Warren mentions AI in more than 7% of her emails in 2026.

And what do these AI emails say? Let’s dive into the content a bit. To do that, I read them all myself (there are only 73 total for 2026), and then I asked Claude Fable to read them and categorize them by topic.

63 of the 73 emails frame AI as a threat. And the single largest category, at 26 percent, isn’t jobs, or deepfakes, or chatbots harming kids—it’s AI as an instrument of money, oligarchy, and corruption. Data centers and their energy costs come second, the Trump administration’s use of AI third; jobs and automation, the issue the whole populist-backlash thesis is supposedly built on, runs fourth at only 12 percent!

Democratic candidates aren’t yet running against AI as a technology that is going to put us all out of work, exactly. They are running against AI as the newest thing the billionaires are doing to you.

Here are some examples of what the 2026 AI emails look like. (These are all direct verbatim quotes.)

This is, I think, a key takeaway. AI is currently entering Democratic rhetoric through the anti-billionaire frame—the oligarchy story makes AI legible as a villain to donor audiences, and AI in turn gives the oligarchy story fresh material, a visible new answer to the question of what all that concentrated wealth is building. This mirrors what Shor’s polling finds on the demand side, where he argues persuasively that voters experience AI and the cost-of-living crisis as a single unified threat from a rigged system, rather than as separate issues. The supply side and the demand side of Democratic politics are converging on the same message.

Tech populism tends to wax and wane

Finally, it’s good to place the small but growing anti-AI wave in historical context. When we zoom out and look at tech more generally, we see that Republican anti-tech messaging spiked around 2021 at a peak almost as high as the recent Democrat backlash against Elon. This spike was largely driven by emails fundraising around the issue of “censorship” in social media, which died down over the course of the next several years.

This suggests two possible things to me. First, it may be that parties in opposition to the presidency tend to engage more in this anti-tech kind of populism. And second, we don’t know yet where the Democrats’ anti-billionaire and anti-AI rhetoric is going to go—Republican opposition to social media and “Big Tech” has not really crystallized into a clear policy agenda, and it’s entirely possible the same could happen this time around, too.

Looking ahead

There’s no guarantee that the public’s mood will drive political activity. Politicians have to decide that it’s in their interest to pursue particular ideas, rhetorical strategies, or policies that some people may demand. If we want to understand where new forms of political populism are heading in America—and how they may or may not seize on AI as a key villain in this broader drama—then we need to measure both how voters feel about the economy and AI, and how politicians are speaking about it.

My idea in this piece is that fundraising emails give us useful insight into how politicians are approaching these issues. Emails are sent frequently, they’re relentlessly optimized to drive clicks and donations, and they’re aimed at politicians’ core supporters. So they give us a sense of where the momentum is among each party’s base.

Democratic party emails paint a pretty clear picture: economic populism is widespread, particularly in the form of billionaire bashing; anti-AI populism is much, much smaller, but just starting to show signs of growth. And the way it’s growing matters. AI is not currently entering Democratic rhetoric as a jobs issue or a safety issue—it’s entering as the newest thing the billionaires are building, pushed first by the same progressive members who pioneered billionaire bashing before it conquered the rest of the party.

Today, AI sits in Democratic fundraising rhetoric almost exactly where the billionaire sat in 2019—a message popular mostly with the party’s left flank, and not commonplace in the party as a whole. Whether it follows the same path to widespread party adoption is the question these inboxes will answer first in the coming months and years. We’ll be watching!

Like Free Systems? Consider becoming a paid subscriber — all proceeds go to supporting Free Systems community events and programs. By subscribing, you help us to further our research, keep all our content free and open, and gain reach on Substack (which only promotes newsletters with paid subscribers.)

250 years into the American experiment, it’s time to ask: what will the freedom we cherish so much look like after AGI? I’ve been arguing for a while that the political economy of AGI is a fundamental question of our time. More and more people are starting to agree!

It was fun to go on MTS this week to discuss some of the key questions for what a post-AGI political society will look like. There are a bunch of factors that seem to tilt the post-AGI world towards totalitarianism—including the concentration of resources inherent in training huge AI models, the way AI can be used for surveillance and control, and the existential tail risks of AI that might justify extreme security states. But at the same time, AI presents incredible opportunities for building a better and more effective democracy.

Some highlights:

“If you can use super powerful AI to give every person a super effective political agent that represents them all the time, I think you can actually go really far with that.”

“A lot of our biggest mistakes in governance today are the result of the fact that most of us are way too busy to pay attention, and then a small group of wackos gets to drive the decision-making process. And certainly in local politics, that’s the dominant problem. Think about NIMBYism. That’s totally what drives it.”

“I think we could totally rethink the structure of government around AI in ways that would be super exciting.”

“The problem will be we probably can’t do that if we’re talking about one closed weight model that has won out. How are we gonna run a democracy if all of our agents are built on a single model that has a single point of control by some CEO and or the government?”

“We need that plurality of models. And we need to own our agents. The agent needs to respond only to us, not to the model company.”

Guardian Angels as our representatives

This idea of a powerful political agent is core to my idea of “building political superintelligence.” But there are tons of unanswered questions about how to actually build political agents that work well. We experimented on this in my class this past quarter, and the results were pretty mixed.

Well, Gwern Branwen, has some fascinating ideas on the path forward. The piece sketches the architecture of a personal “guardian angel” that is far more than the sort-of dumb, briefly personalized but rapidly amnesiac personal agents you can try to build with current pipelines today. The vision is for an agent that learns as it goes and deeply understands and remembers your preferences. Although the vision is super broad, Gwern points out that politics is one obvious unlock this would provide.

“One can delegate political participation to one’s GA, allowing ‘direct democracy’ on unprecedented scale, as a GA can oversee arbitrary amounts of involvement that a human has no time or interest for, and only punt to their principal when uncertain.”

“One can imagine a single ‘Congress GA’ which tries to learn every US Congressional member, for example, and which could, within minutes, simulate roundtable debates among hundreds of politicians before a vote (whereas it might take months to get a debate and vote done with the real US Congress). Or their official GAs could run independently, and conduct an emergency debate at 4AM when every human is sleep.”

This is a really exciting vision. I see two big questions for how to make it create genuine political superintelligence:

1. How do we make sure the agent is more than a digital twin – it has to share our values, but it shouldn’t share all of our less considered opinions. It should be willing to push us on topics/decisions that it has studied more than we have. Gwern explicitly recognizes this, writing: “A GA should help its principal become themselves and develop their ideals, morals, and their personality. It is not enough to model an average, undifferentiated, inchoate set of preferences and values, and settle for mediocrity and stasis; the job of the principal is to develop themselves and give the GA something meaningful to learn to emulate.” But how do we actually get the agent to do that? It seems like a capability that lies somewhat beyond current AI models, at least when it comes to contested political issues.

2. How do we govern it? We need some way to make sure we own our guardian angels, so that they answer to us and can’t be secretly commanded from afar by others. Gwern’s proposal is for a startup with dual-class voting shares. This makes a lot of sense to me for the development phase. The long run seems trickier. Can we build the infrastructure for our AI agents on private rails that a startup controls? This seems like the biggest question of all.

Can AGI be open?

As the above makes clear, if we’re going to get to political superintelligence, we’re going to need some way to own our agents. That may or may not require open-weight models that we can run locally and be confident can’t be “rugged” from afar by centralized providers.

The past week or so, there’s been a major “vibe shift” back towards open-weight models. Companies seem to be waking up to the fact that they don’t want all of their knowledge and capabilities sucked up by the big foundation models. New open-weight models seem to be rapidly improving in capabilities. And, perhaps most tantalizingly, neolabs are starting to deliver promising results on customized models.

Satya Nadella has been hammering this idea recently—as I discussed when teaching his “new loop” at the GSB. So it was very interesting to see him sit down with Yash Patil from Applied Compute, a company betting on this new loop.

There are counterarguments to consider, as well.

First, if we are really about to hit the steep part of the RSI curve, then the frontier models might gain a potentially insurmountable lead over the open-weight models, leaving them in the dust.

And second, if the safety issues with powerful AI are as serious as many people—including, increasingly the Trump administration evidently—think they are, then there may be significant pressure to limit access to open-weight models.

As I wrote yesterday in response to a post from Rudolf Laine, I really have no idea how this is going to play out or how we’ll balance these issues. But if we want to achieve political superintelligence, we’re going to have to find a way to have safe AI that is also sufficiently decentralized to allow us to avoid a digital leviathan.

Tweet of the week

I thought this was an awesome example of the kind of serious, introspective work that AI allows for. Useful for journalism and for academia, too!

Question of the week

The Founders designed institutions assuming human limitations—busy citizens, slow communication, fallible representatives. Which constitutional feature would you redesign first if every citizen had a tireless political agent?

As we start to approach the midterm elections, questions about AI’s political values are going mainstream. When I first started working on this topic two years ago, many in the media thought it was a fringe topic that Elon and others were using as a disingenuous political cudgel. Times have changed, and it’s becoming increasingly clear to people that we’re all going to be using AI models to obtain information, and that how these models handle loaded political questions will be incredibly important.

• This past week, both the Washington Post (who I chatted to) and the Economist ran their own analyses of political slant in major AI models.

• Campbell Brown, co-founder and CEO of Forum AI (where I’m an advisor) had a great appearance on Bloomberg News to discuss these same issues.

Where do we take this research next?

• It’s time for the labs to help us do better research on this question, in the same way they’re helping economists study AI usage in the workplace

• Analyses are converging on finding left slant, but we have no idea whether the existing measures reflect responses to realistic user prompts or not, because we have no idea how users talk to AI models about politics in practice

• The labs should share anonymized data on the nature of political queries, so that researchers can calibrate their slant measures to realistic user prompts

Handing AI Populism

has a great piece in the Economist on the state of AI populism. There are three points in particular in the piece that strongly reinforce the Free Systems view from “The Politics of Jobless Prosperity”:

• Backlash isn’t here yet, but it will be serious

• We should measure stuff better with help from the labs, so we know when we’re seeing real job loss

• We should use AI to improve our politics (i.e., build political superintelligence).

Google Proposes a Self-Regulatory Body, Much as I Predicted

This didn’t get a lot of attention amidst all the Mythos/Fable drama, but Google has now explicitly proposed exactly the kind of self-regulatory body for model review that I envisioned back when Anthropic announced Glasswing.

Tweet of the week

I’m not the only one wondering if we’re seeing an AI TACO trade and “governance by stock market” around the Mythos/Fable stuff.

Question of the week

Did you find Free Systems through an ad on either Facebook or Instagram?

If so, welcome, and I’d love to hear from you—please reply to this email and tell me what led you here, and what you think of it so far! These ads came from my fully autonomous agentic ad agency, which I’ll be writing about soon.

Each frontier AI model release brings surprising new abilities that seem to shorten the list of what makes humans unique and wipe out the startups and established companies that thought they had differentiated themselves from this powerful new technology.

In a remarkable essay published the week before last, Satya Nadella took stock of this state of affairs and asked how any organization can thrive in a world where AI models continuously absorb the expertise of people and companies and sell it back as a commodity. This is not merely a question of business, Nadella tells us, but an existential question of political economy because it implicates the social contract and our shared trust that society can work for us all.

His answer to this conundrum is that firms have to combine their own human expertise with AI to codify their private knowledge inside model evals and training environments that they, and not the frontier labs, own. Nadella envisions a new “loop” where a company transforms its own workflows and accumulated judgment into systems that improve with every use, measuring the result against its own yardsticks rather than the public leaderboards, since, as he puts it,

“Private evals should capture whether a model is actually improving against outcomes that matter to the business.”

These private evals are what enable firms to remain sovereign because it allows them, and not the labs, to own their specialized knowledge. Nadella explains: “A company should be able to switch out a ‘generalist’ model without losing the ‘company veteran’ expertise built into their learning system. This is the key ‘test’ of your control and sovereignty in the era ahead.”

Nadella’s insights go far beyond the firm: indeed, the same opportunity exists throughout society. Think of citizens in a democracy, our political leaders, our universities…we all need a way to harness AI without being absorbed and captured by it. We should all be able to switch out one frontier model for another in our work, and in our lives, without losing our accumulated expertise. This new loop is the way.

So how should we all learn to build it? It’s not something covered in any standard curriculum. But I’m absolutely convinced that it needs to be.

I say this in part because I’ve been experimenting with it in my teaching this year. Just a few weeks before Nadella published his essay calling for companies to build private evals, I had my students in the new “Free Systems” class I’m teaching at Stanford GSB do exactly the same thing. With Claude Code subscriptions, OpenRouter API credits, and a couple of weeks of practice under their belts, my students all created first drafts of their own personal evals in a single three-hour class session.

Each student identified a criterion they cared about—how sycophantic the models were when debating controversial topics, how well the models gave voting advice, whether they understood crucial cultural nuances across languages, and much more—and then designed a rubric for scoring model responses according to their personal beliefs. Then, they built a leaderboard comparing how different leading AI models performed according to their measure. The results were spectacular! But we weren’t done.

Once they could measure models against their own personal yardsticks, we spent the rest of the quarter pushing further—what else could students do with this power? It turns out, a lot. Their final projects give a good sense of what the future that Nadella describes might look like. Together, they look like a nascent field guide to building these new loops in the wild.

What the students did

Here is what that future looked like in our class: fifteen final group projects spanning finance, governance, media, and security, several of them grown straight from the personal evals the students had built a few weeks earlier.

While many of the projects are, at their core, evals, they go beyond the evals we built in class because they’re not just leaderboards; instead, the evals are embedded into broader tools that do something based on the information—whether that’s make a new recommendation, helps the user complete an action, give better advice, and so on. This is how they begin to show us what the new loop will look like.

I’ve grouped them below by theme, with a great deal of help from Claude. You can see all of the projects at this website.

Finance and delegated decision-making

• AI Bank Run Simulator (Shang Jing Chia) — live simulation of an AI-driven bank run with LLM agents acting from distinct personas; lets you watch cascades unfold and inspect each agent’s reasoning.

• Cross-Market Arbitrage (Graham Griffin, Ethan Romer) — detects informed trading propagating from Polymarket to Kalshi, measuring the lag between the transparent on-chain market and the opaque regulated one.

• RegFi Compliance Checker (Bernardo Herzer) — AI tool that automates regulatory compliance checks on financial AI systems to reduce manual review burden.

Media and framing

• Headline Truth (Jenna Jokhani) — evaluates whether headlines faithfully represent their articles, and whether models can detect distortion or sensationalism.

• A Kaleidoscope for Political Framing (Milly Wong) — paste an op-ed and see where it lands in a 3D map of 284 articles across 7 outlets, embedded in 384 dimensions via UMAP, with separate topic and worldview lenses.

• News Framing Dashboard (Eddy Jiang) — feeds one article to Claude, GPT, Gemini, Grok, and Llama and quantifies framing differences across actor salience, affective loading, context inclusion, and hedge density.

Alignment, security, and governance

• LLM Imposter: The Council of Five (Raymond Llata) — social-deduction game where one misaligned AI tries to persuade a council of aligned agents to adopt a selfish policy; a proxy for multi-agent alignment robustness.

• Steganographic Injection Demo (Yuanxin Ma) — tests 11 hidden-text attack types across 15 frontier models (3,600+ API calls) to see if models can be pushed into biased product rankings.

• ChatGPTween (Jaxon Gonzales, Juan Sandoval) — translates parent questionnaires or guided conversations into a personalized AI “constitution” for a child’s chatbot; conversational constitutions refused all 24 adversarial prompts that MCQ-based ones failed.

Personal tools and model selection

• Model Signature (Navya Agarwal, Zoya Fasihuddin, Diya Ahuja) — blind A/B/C onboarding across ten task categories builds a personal model-preference heat-map, then routes each query to the best-fit model.

• Streamline (Leticia Auriemo, Bennett Evans Zytko, Alec Profit) — conversational sensemaking dashboard that builds a personalized intelligence feed to help users understand complex issues.

• Rundown (Prakhar Goel) — connects to Slack and distills a week of channel activity into a one-minute digest.

• ResumeScope (Jonas Pao) — resume-review platform where AI recruiter agents simulate how real recruiters read, predicting attention and rating.

Capital concentration and AI infrastructure

• Situational Unawareness (Vivek Yarlagedda, Kathy Shao, Shawn Gregory, George Zhang) — interactive map of the AI stack covering 92 companies and 297 deals, filterable by layer (compute, networking, raw materials, power, capital) and deal type to trace where deal flow concentrates.

• The Hidden Cost of AI (Natalie Hampton, Quincy Stone) — map of where AI compute is physically concentrated and which communities absorb the water, electricity, land, and pollution costs; reports 86% of mapped capacity in wealthy hubs and 70% of 2024 data-center electricity used by the US and China.

What I learned

I drew four main lessons from this first experiment in teaching the new loop.

Lesson 1: Human expertise and critical thinking is an essential pre-requisite

The projects succeeded in large part because they were based on things the student already knew and cared deeply about. Far from letting students outsource their thinking, the work demanded more of it, which is why the familiar worry that AI erodes critical thinking has it backwards here, where critical thinking is the thing that makes the AI worth anything at all. This implies that we can’t only offer AI-intensive classes in the university of the future; instead, we need these classes to come after classes that teach essential critical thinking skills and domain knowledge.

Lesson 2: Students need a ramp and tangible examples

Nobody walks in and commands a fleet of agents effectively on the first day. Students need a ramp, structured early assignments and concrete examples they can imitate before they strike out on their own. What looks from the outside like sudden fluency is really careful sequencing, giving people just enough scaffolding to find their footing and then taking it away once they have it.

Lesson 3: In-person collaboration and tutoring is crucial

I found that the work goes far better when students are in the same room, where they can look over a shoulder, borrow a trick someone two seats away just figured out, and get unstuck the moment they are stuck rather than three days later by email. A live tutor who can diagnose a broken agent in thirty seconds is worth more than any amount of documentation, because the failures these tools throw are still strange enough that a person rarely knows the right question to ask on their own.

Lesson 4: Everyone needs the right push to get started

The most important thing I observed was that many people don’t know what they don’t know—they assume it will be far harder to master the new loop than it actually is, and that assumption prevents them from getting started. Having an easy, low-pressure way to jump start their own experiments helps get people moving.

The Republic of builders

Although Nadella’s essay is primarily a business essay, it can also be read as a piece of political economy. He worries about a future in which AI models “eat everything they see” and predicts that such a world will not lead to a sustainable political equilibrium. “There is no societal permission for an AI future that hollows out entire industries,” he warns us.

Instead, he argues that if firms can build their new loops, then they “will create value for themselves and for the economy around them. Employees will see their expertise amplified and their judgment become part of systems that make it replicable and scalable and the benefits accrue to the companies and communities around them.”

This is the ecosystem we’ll need to sustain a free society in the AI era. We need everyone to be able to use AI, measure it according to their private evals, update accordingly, and capture value. In a way, this is the oldest argument for educating a free people, just refreshed. Writing in 1822 to a legislator who was building Kentucky’s public schools and university, James Madison put it plainly: “a people who mean to be their own Governors, must arm themselves with the power which knowledge gives.” The loop is how a free people arms itself now, and a citizen who can neither wield these systems nor measure them for herself has simply handed that power to whoever owns the models.

Therefore, I want every person in the world to learn how to do this. We need to teach executives and MBAs, college students, and even high school students how to do this. We don’t only need “an army of citizens building evals” as I put it previously; we need that army to also understand how to turn those evals into concrete tools and decisions that make AI work for us, instead of vice-versa.

That’s a big ambition, and I’ll be taking it in steps. Next year, I’ll be offering a set of new courses for executives, MBAs, and undergrads entitled “AI Tools for Leaders” that aims to teach people this new loop. If we can succeed, I hope we can expand it beyond Stanford, and I hope that many others will be developing similar ideas and experiments in parallel all across the world, so that we can all benefit from the new loop and make sure that frontier models don’t eat us all.

"You mean to tell me that the success of the program and my reelection hinge on the Federal Reserve and a bunch of fucking bond traders?"

—Bill Clinton, as reported by Bob Woodward

Last Friday’s news that Anthropic was pulling its most powerful AI model, Fable, from the market under severe pressure from the Trump Administration was earthshaking. Seemingly everyone has a take on what this means for regulating AI going forward, how we’re entering a new era of government control, and so on. But you know who has seemed oddly unmoved? The stock market.

After several years of letting AI rip, the Trump Administration now wants to control it—to make sure foreign adversaries aren’t able to use these tools, to mitigate potentially catastrophic security threats, and, probably, based on public comments from Pete Hegseth and others, to punish Anthropic and Dario Amodei personally.

Clearly, we’re heading into a new regulatory regime, one in which the federal government pre-reviews models and de facto licenses them (Trump’s EO from just two weeks ago explicitly rejected a licensing framework, yet Howard Lutnick’s letter to Anthropic specifically says that they will need to obtain a license).

Will this encourage the kind of impartial, reasoned process of model oversight many have been calling for? Or, as has argued, and as seems far more likely, will this process look far more ad hoc and personal?

With AI still far from the top issue in Americans’ minds (Hormuz still dominated the news agenda last Friday despite the Fable takedown), it’s unlikely the midterm elections will provide a mandate for a codified legislative solution to this problem, in the US at least.

Which probably means, like in many other Trump policy domains, it comes down in large part to the stock market. This is the famous “TACO trade”---short for Trump Always Chickens Out, which means that he will (almost) always reverse course on a major policy decision if the stock market freaks out enough.

The basic TACO trade occurs when you buy the dip that some of Trump’s actions cause, confident he’ll reverse course and the market will recover.

But the “second order” TACO trade happens when the markets shrug in the first place, anticipating in advance that Trump won’t stay the course. Geopolitically-driven selling has become progressively more muted over time in line with this trend.

AI now seems capable enough to pose genuinely important risks, and at the same time so central to economic growth that slowing it down might mean tanking the stock market, reducing investment in the compute buildout, and consequently slowing the economy.

The market has been attentive to other AI signals in the past. When DeepSeek released R1 in January 2025, fears about AI’s growth story wiped out historic sums in a single day. So why did the market largely shrug at Fable being shut down—the largely unexpected pullback of the most capable product in the most economically important industry? On paper, it seems surprising.

It’s impossible to prove, but from digging into the data, I think there is now an ongoing “second order” TACO trade happening in AI. The market isn’t reacting because it doesn’t believe Trump is going to slow down AI, even despite any personal wishes he and his team may have to harm Anthropic.

People aren’t talking about this TACO trade much in AI circles because the conversation is dominated by more legalistic discussions about the emerging regulatory environment. And it’s easy to miss it, because it’s “the dog that didn’t bark”---the media isn’t going to be excited to cover the story that Trump’s battle with Anthropic hasn’t moved markets—but it carries rather complex implications for governance at the frontier, and the market’s tolerance for slowing AI development.

It’s the economy, stupid – and the economy is AI, now

To see the TACO AI situation, you first have to appreciate the well-known fact that the fate of the American economy now rests largely in the hands of AI. While the rest of the world stagnates, America’s economy continues to be remarkably strong, largely thanks to AI. Some basic statistics that have been well documented, but are worth reinforcing:

• The largest AI and technology companies were responsible for 53% of all S&P 500 gains in 2025, according to Goldman Sachs.

• The “Magnificent Seven” alone are about 34% of the index, while the ten largest AI/cloud/chip companies are about 38%, according to State Street’s SPY holdings data.

• AI-related investment categories (information-processing equipment, software, and R&D) accounted for roughly 93% of U.S. GDP growth in Q1 2026, contributing 1.49 percentage points of 1.6% annualized growth, according to the Bureau of Economic Analysis.

• Microsoft, Amazon, Alphabet, and Meta alone are expected to spend roughly $635 billion on AI infrastructure in 2026, up from $383 billion in 2025 and just $80 billion in 2019, according to a Reuters report citing S&P Global estimates.

• The combined valuation of SpaceX, OAI, and Anthropic recent and prospective IPO’s ($3.6 trillion) roughly equals the GDP of France

Slowing down AI means slowing down the economy

I’ve written previously that we’ll see a much more potent populist backlash to AI if we start to see substantial job loss due to AI.

But there’s an interesting converse to this, too: the Trump Administration is going to fight like hell to keep the economy ripping and the stock market high, and that means keeping AI ripping. It means keeping the datacenter and compute buildout ripping, and it means keeping the frontier labs rolling, producing better and better tokens that companies and people want to buy more and more of.

We might have many reasons we want to slow down or even pause AI development and the release of new AI models to the public.

But slowing down AI means slowing down the economy. The entire AI buildout is predicated on continuing growth in the demand for tokens. Goldman Sachs expects AI token consumption to rise 24x by 2030, to 120 quadrillion tokens per month.

Releasing newer, better models is an essential part of how that demand continues to rise. OpenAI describes its business as a flywheel in which more compute produces step-change gains in model capability, stronger models unlock better products and broader adoption, adoption drives revenue, and revenue funds the next wave of compute.

This all keeps the engine running for an administration that has staked a lot of its economic credibility on AI dominance. This creates a fundamental political tension for Trump, in that his actions against Fable help make a broader pause seem more plausible—which risks slowing down the economy.

Of course, Trump might be able to slow down Anthropic without slowing down AI as a whole. If Anthropic isn’t able to release higher-quality new models, maybe customers just switch to OpenAI, compute reallocates as needed, and the economy keeps on humming.

But it’s awfully risky. Anthropic is the single highest-revenue frontier lab, now, and it is deeply enmeshed in the broader AI economy. The company has secured roughly 3.5 gigawatts of next-generation Google TPU capacity through Broadcom, beginning in 2027, while keeping Amazon as its primary cloud and training partner. Meanwhile, Google has separately committed to invest up to $40 billion in the company.

Slow Anthropic down and you may just threaten the financial plans of Broadcom, Alphabet, and Amazon…and countless others who exist in this network of economic dependencies.

The market is betting on an AI TACO

So is Trump going to slow Anthropic down? Might he even expand the approach and slow the other labs down, too? On a livestream with and this week, argued that people who want to see a pause or slowdown in AI development should see Trump’s actions as a positive step towards that. I think there’s a lot for that line of argument. But markets don’t seem to think that’s where we’re headed.

Squint at the returns of AI-exposed stocks around June 12th, when Fable went dark, and it’s hard to see a market reading the action as bad news for future AI earnings—though we should always keep in mind that since stock market prices have many determinants, this is hardly a clean test of how the Fable shutdown really affected investor sentiment.

The publicly traded names enmeshed with Anthropic, from Broadcom to Alphabet to Amazon, showed no such distress; in the first full session after the order they traded up alongside a market that happened to be rallying on unrelated news of a U.S.–Iran deal to reopen the Strait of Hormuz.

Remember, Nvidia’s stock plummeted by 16%, with Broadcom, Amazon and Microsoft’s prices all sinking too after DeepSeek’s model performance cast that, in some ways, also cast doubt on the certainty of the AI development buildout.

The clearest negative reaction showed up in Anthropic’s own pre-IPO proxies, where the Hyperliquid valuation perpetual slipped about 3.7%---not a huge amount, albeit in a strange and not that liquid market.

Meanwhile, over on Kalshi, traders are betting on a quick reversal rather than a long freeze: by mid-June they put the odds of Fable returning before July 1 at roughly 58%, and about 74% by July 10—a delay that will feel agonizing to those of us who fell in love with the model during our brief but glorious time with it, but nothing like the six-to-twelve-month slowdown that might really spook people.

According to another Kalshi market, pulling Fable did slow down our expected timeline to a public lab announcement of attaining AGI…but only quite modestly.

As the figure below shows, the market adjusted down the chance that AGI is declared publicly by April 1st, 2027 by about 5 percentage points, and adjust down the chance it’s declared by Oct 1st, 2027 by another 5 points or so—reallocating that 10 percentage points of probability out to the post July 1st, 2028 period. These are pretty modest shifts.

Each of the above market signals could be interpreted a variety of ways. We can never conclusively prove if a second-order TACO trade is occurring or not. But there are some basic, hard facts we should keep in mind: the US economy is very dependent on AI, and taking any actions that risks slowing down the compute-revenue flywheel is inherently risky.

Trump’s actions on Fable have kicked off a thousand talking heads abuzz with the possibilities for how he might slow down AI progress. Meanwhile, Wall Street shrugs, and prediction markets tell us to expect Fable to return soon and progress towards AGI to continue largely apace.

What will be enough to give the stock market pause?

All this means we may be heading towards an impossible political bind. Nothing is more important to American voters than the state of the economy. And today, the economy rests on AI. We have to keep feeding the beast, maintaining expectations about future AI growth.

At the same time, we might have very serious cybersecurity and national security reasons to want to feed that beast a little bit more slowly. But with each passing day, the political costs of slowing down may rise, as AI becomes “too big to fail.” And this isn’t just a Trump story. If and when the Democrats take control of Congress in November, they too may discover that their enthusiasm to regulate AI more aggressively becomes tempered by a fear of being seen as responsible for crashing the stock market.

Or, maybe there’s a middleground. Maybe the market will reward prudent actions that reduce cybersecurity risks while allowing for AI development to proceed at a reasonable pace. This is really the big question for AI regulation: what kinds of regulatory steps will the market tolerate, and which will cause it to freak out? So far, Trump’s forays into tighter control over AI haven’t spooked the market, but it’s hard to know whether that’s because the market wants more regulation, or because the market thinks Trump won’t actually do it.

Right now, elite discourse is locked in a battle to shape public opinion about AI, and most of what we hear about Americans’ views reaches us top-down, filtered through a handful of competing narratives. The labs offer AI as electrification—a civilizational force that will ultimately benefit everyone, where abundance and managing job loss are the issues of the day. The safety community remains squarely focused on existential threat as model capabilities advance. And politicians have found their expression of AI skepticism in data centers, casting opposition to local construction as a proxy for broader anxieties about who bears the costs of the AI boom.

But does any of this describe how normal, everyday Americans are actually encountering the technology?

One way we can learn more about this question is by looking at what kinds of AI content are circulating on social media. Social media often sits outside the elite narrative pipeline—it’s where ordinary people encounter AI on their own terms, share what they’re actually doing with it, and begin to form the views that eventually manifest in polls. It’s true that social media content disproportionately reflects those most “chronically online” in American society, but at a moment when opinion is still being shaped rather than settled, it’s a leading indicator worth taking seriously.

So we collected 25,000 videos about AI across YouTube and TikTok. First, we worked with Claude and Codex to analyze the videos, studying how they discussed AI, whether they were positive or negative in their sentiment, and what angles they took on AI. Then, to get in the heads of normal social media users, we watched videos. A lot of videos. Sooo many videos.

The picture we find is more complex and surprising than elite narratives would suggest. Put simply: the public AI debate is far more normal than the elite one.

By “normal” I certainly don’t mean that the content itself feels normal (it’s often strange), or that it isn’t full of exaggeration and vitriol; rather, I mean that both the positive and negative videos about AI tend to focus on immediate, near-term, personal aspects of AI much more than the elite debate does.

For one, there is way more content that embraces AI than you might expect if you only read national news, perused certain corners of X, or listened to politicians and lab leaders constantly discussing the public’s AI backlash. On TikTok and YouTube, content embracing AI beats out explicitly anti-AI content by 3 to 1.

This content sits outside the breathless takes about civilisational transformation put forward by the labs and e/accs, and primarily reflects Americans quietly absorbing and adopting AI, rather than explicitly making the case for it. It tends to cluster around videos showing off fun AI-generated effects and memes, discussion of how AI can increase personal productivity, and showcasing creative tools.

Second, there is plenty of extremely negative, AI-skeptical content. This content tends to map poorly onto the debates currently dominating policy circles. There are a set of quite focused and organized mass movements against AI on Youtube and TikTok, that largely don’t line up with what we hear in the elite narrative. The largest category of negative content is focused on how AI is ruining and co-opting artistic content and the creative process. Discourse about x-risk, data centers and job loss—register somewhat less prominently.

Without further ado, let’s jump into this weird, wonderful, and emerging universe of AI content.

What 25,000 YouTube and TikTok videos about AI say

To understand the popular conversation about AI, we pulled roughly 25,000 videos from 2026 that we could find on YouTube and TikTok that matched keywords related to positive or negative AI sentiment.

These videos amassed over 2B views in sum, spanning from content posted by influencers and shitposters, to official news channels, and education sources:

To categorize them and analyze sentiment, we first transcribed them using Whisper. Each video was then classified by large language models (Claude Sonnet 4 and GPT-5) in successive passes, first a relevance/stance filter, then a strict precision-first re-verification with confidence scoring, and finally assignment to a fixed topic taxonomy, with classification thresholds calibrated against hand-audited samples. We restrict the analysis to videos posted in 2026 and weight every topic by reach (total views and plays).

Sentiment is a crude object, and after watching many of these videos, we think it’s more appropriate to say that the ‘positive’-coded videos reflect the views of “adopters” and that the negative-coded ones reflect the views of “resisters,” rather than saying that the video are uniformly positive or negative in their affect towards AI.

And the results are interesting! We find that videos ‘adopting’ outnumber those ’resisting’ it, by roughly a 3 to 1 margin. This doesn’t mean that sentiment towards AI in general is positive—as we’ll explore below, there are many ways to interpret these patterns. But this certainly shows evidence that considerable AI absorption is already well underway, largely beneath the radar of the backlash narrative.

What Adopters talk about on YouTube and TikTok

So what is the ‘adoption’ content actually about? It’s not quite the high-minded techno-optimism of the SF tech bubble crowd.

AI memes & effects - 43%

The most popular category is what we call “AI memes & effects”---basically, fun and goofy videos where people show off different kinds of AI-generated content. Some may call this ‘AI slop’ or ‘brainrot’, but AI is clearly opening up an entire pipeline of new meme-generation and storytelling opportunities which is now baked into the everyday scrolling experience of those on Youtube and TikTok. Here are two fun, representative examples.

TikTok — “Now it’s pandas turn 😂” (Lumi AI-dance animals)

@lumi.0102Now its pandas turn 😂 #funnyai #tiktokfunnyvideo #aidance #tiktokusa #LUMI

Tiktok failed to load.

Enable 3rd party cookies or use another browser

YouTube — “Idiots Laugh At Unusual AI Videos” →

These are obviously frivolous, but we shouldn’t discount them for that. This is a major experience Americans are having with AI: fun, goofy, creative content that is designed to provoke and entertain.

Career / productivity - 25%

The next biggest category are videos that talk about people using AI to help them find jobs, get their work done faster, or generally help them manage their lives more effectively. Below are two interesting examples:

- TikTok — “job search just became unfair 💀” (AI quietly winning the hunt) →

@thatgirlgetshiredjob search just became unfair 💀 #parakeetaipartner #jobsearch #jobai #career #fyp

Tiktok failed to load.

Enable 3rd party cookies or use another browser

- YouTube — “ChatGPT acted as my financial advisor and helped me build a get-out-of-debt strategy” →

This kind of “self help” content is not necessarily glamorous, and is exactly the kind of thing that the elite narratives are likely to miss but that is actually important to people. In particular, it showcases how people are using AI to get a ‘leg up’ in their lives---finding jobs, working faster, managing their finances, navigating systems that feel stacked against them. This sits in direct tension with the job displacement content on the resister side.

Creative tools - 15%

Close behind is a genre of videos in which creators share interesting ways that they’ve used AI models to create new things—like an AI edit of a movie, or a new Minecraft trap.

- TikTok “Motion Control is Crazy” — a Game-of-Thrones AI edit →

@aidanstanik.aiMotion Control is Crazy #motioncontrol #ai #higgsfield #gameofthrones

Tiktok failed to load.

Enable 3rd party cookies or use another browser

- YouTube — “AI Generated Minecraft Traps” (genuine showcase, building game content with AI) →

These often seem to feel especially positive, because they come from creators who are excited to share what they’ve been able to do with AI. This sort of content faces sharp opposition on the resister side, but has been wholeheartedly embraced within certain pockets of social media.

Education and learning - 8%

A smaller category includes videos where people show off things they’ve learned using AI, or ways that they use AI to teach themselves things.

- TikTok — “Learning epic riffs with Suno” (guitarist genuinely stoked to jam with AI) →

@dominicflynnguitarLearning epic riffs with @Suno

Tiktok failed to load.

Enable 3rd party cookies or use another browser

- YouTube — “ChatGPT… I feel like a fool [for not using it sooner]” (enthusiastic late-adopter) →

This is a use case that a lot of AI companies like to lead with—the personal tutor narrative, AI as democratizer of education. We see here that people are certainly using AI to learn, but it’s not the dominant experience.

AI companionship - 4%

A smaller and stranger category covers videos about people’s AI companions, ranging from quirky and weird to somewhat creepy.

- TikTok — “I just love coming home to my Claude ‘Jarvis’ 🥰” →

@jakob.robic5I just love coming home to my Claude Jarvis #claudecode #claude #ironman #ai

Tiktok failed to load.

Enable 3rd party cookies or use another browser

- YouTube — “flirting with AI again, but worse” (creator playfully smitten with their AI companion) →

Although many of these videos involve people excited about their AI companions, I would hesitate to ascribe overly positive qualities to this category. It certainly does confirm that the companionship phenomenon that journalists have been writing about is clearly real and visible in the data.

Breakthrough science - 1%

Finally, a very small category covers AI’s ability to advance science.

While scientific advances constitute probably the main positive opportunity the labs frame around AI, it is far from a major focus of the social media content people are seeing about AI.

What resisters talk about

Adopter videos stress how AI can be fun, how it can help you advance your career and be more productive, and how it can teach you new things. Those are pretty positive framings about AI for everyday people. But what about the significant number of videos from the resisters? Just because these negatives are far more negative does not imply that they line up neatly with elite narratives about the AI backlash—indeed, we find that they look quite different.

Creative theft - 22%

The largest and most organized category of AI backlash content centers around the ways that AI constitutes creative theft and harms art. On TikTok, this community shows signs of real organization: they use hashtags like #noAI and #stopAIart, and they seem to have a real social structure—with shared trends and in-group rituals, like encouraging the commissioning of fellow artists, and coordinating on takedown requests for AI content.

- TikTok — “Use your brain. Use your creativity… Don’t steal other people’s work. Pick up a pencil.” →

@allisonrtylerUse your brain. Use your creativity. Use your imagination. Experiment. Use your hands. Stop needing instant gratification. Invest the time to get good at something. Try something new. Don’t steal other people’s art. Go play. Don’t be a jerk. You’re better than that. #darkaesthetic #whimsy #realart #strangecore #noai

Tiktok failed to load.

Enable 3rd party cookies or use another browser

YouTube — “Disney’s New AI Move Has Made The Art Community Furious“ (studio raining on artists’ work) →

TikTok — campaign against AI-generated art, we commissioned this animatic… to support real creatives” →

@starmapsintergalacticAs part of our campaign against AI-generated art, we commissioned this animatic from @Eli to show our continued support for real creatives, especially in a world that may one day look like the one in our game. Starmaps Intergalactic is a sci-fi gamefollows humanity after Earth’s collapse and a Resistance fighting to return home. We have been working on this game for the past 3 years and happy to say we're getting closer to releasing. Lucy’s story is just the beginning, follow the story as she tries to get home. And play the game to fight with the Resistance. 🎨 Animation Created from: @u3eli Follow for Game Updates and More Lore. The best place for science fiction lovers and book readers. #noaiart #supporthumanartists #indiegamedev #digitalart #BookTok

Tiktok failed to load.

Enable 3rd party cookies or use another browser

Deepfakes and misinfo backlash - 19%

The next largest category covers concerns about deepfakes and misinfo. Often these are news reports, but individual creators get involved, too.

- TikTok — “The ‘3-Finger Test’ That Exposes Deepfake Scammers Instantly” (Cybersecurity Girl) →

@cybersecuritygirlThe “3-Finger Test” That Exposes Deepfake Scammers Instantly Follow @cybersecuritygirl for more online safety tips Original footage from @huntresslabs #deepfake #scams #news

Tiktok failed to load.

Enable 3rd party cookies or use another browser

- YouTube — “A.I. Content Is Getting Too Good.” →

Jobs displacement - 13%

About 13% of the resister videos talk about job displacement in various ways—ranging from documenting how AI can do things that specific creators used to be able to do, to offering broader economic takes on the situation.

- TikTok — “IM BEING REPLACED BY AI” (gaming creator camman18) →

@camman.18IM BEING REPLACED BY AI

Tiktok failed to load.

Enable 3rd party cookies or use another browser

- YouTube — “What Happened to Horses Is Happening to Us” →

These videos feel different from elite discourse around job displacement in a few ways. First, many of them are personal, talking about how AI has taken a specific person’s job. They often specifically assign blame, naming a particular person (like Mark Zuckerberg) or a particular group (like “Silicon Valley tech bros”). And second, interestingly, the videos—particularly on TikTok—often focus not on the “white collar wipeout” that the AI labs have been warning about, but more about blue collar and service jobs. Finally, they also often have a youth element—with lines like “our generation is cooked” often appearing.

I hate AI - 13%

The next category is general anti-AI content. This category bumps up against the first category—inveighing against AI slop and how it’s harming the world—but doesn’t focus on one specific grievance the way the theft discussion does.

- TikTok — “ai disgusts me, anyone who uses it is immoral“ →

@daisyskayaai disgusts me, anyone who uses it is immoral #chatgpt #antiai #climatechange #globalwarming #foryoupage

Tiktok failed to load.

Enable 3rd party cookies or use another browser

- YouTube — “Every Reason Why I Hate AI and You Should Too” →

These videos are often quite vitriolic. They certainly provide a window into why and how an anti-AI political movement in the US might grow. Clearly, there is a group of Americans with very strong, negative emotions around AI. You feel this very strongly watching their videos.

X-risk - 8%

Videos about existential risk are far less common on social media than in elite, EA-coded “doomer” discourse on X. Interestingly, when they do appear, though, they often trickle down from precisely that community; many of the most popular x-risk videos are straightforward videos of doomer celebrities like Geoffrey Hinton opining on x-risk, or videos summarizing the views of major AI figures on the subject.

TikTok — “EXPERTS WARN OF DANGER IF ANTHROPIC’S NEW MODEL GETS INTO THE WRONG HANDS!” (the viral “Mythos” too-dangerous-to-release story, Ultron edit) →

@ibig_sweepEXPERTS WARN OF DANGER IF ANTHROPIC’S NEW AI MODEL GETS INTO THE WRONG HANDS! Ultron Edit #edit #news #aitakeover #invincible #xyzbca

Tiktok failed to load.

Enable 3rd party cookies or use another browser

YouTube — “Godfather of AI: We Have 2 Years Before Everything Changes!” (Hinton) →

But, social media being social media, there are also some even more dramatic takes, like videos applying spiritual analyses to AI:

TikTok — “Watch out for AI, it can be very demonic” →

@hescoming_backsoonWatch out for AI it can be very demonic. #ai #demonic #god #fyp #viral

Tiktok failed to load.

Enable 3rd party cookies or use another browser

Energy / data centers - 6%

Environmental concerns related to AI data centers are only a small proportion of the content, and the most popular examples showcase how these narratives can be stretched to popular extremes.

- TikTok — “Please stop using AI, you’re harming them. Delete ChatGPT and use Ecosia instead. Polar bears, penguins… are dying.” →

@rby_846Please stop using AI, you're harming them. Delete ChatGPT and use Ecosia instead. Polar bears, penguins, and many other animals are dying. We only have 3 years to save them.#polarbear

Tiktok failed to load.

Enable 3rd party cookies or use another browser

- TikTok — “That water bottle on your desk? That’s not going to exist in a couple of years” (AI is draining fresh water) →

@ecowithdanaI wish more people were aware of this #chatgpt #ai #environment #ecogpt #fyp

Tiktok failed to load.

Enable 3rd party cookies or use another browser

There are also videos with more of a NIMBY vibe (not in the pejorative sense) complaining about how data centers ruin the landscape.

- TikTok — “🤬 This used to be beautiful rural farmland. Now it’s an AI data center in the making.” (Meta) →

@saltyspiritsage🤬This used to be beautiful rural farmland. now it's an AI data center in the making. #aislop #ai #meta #fyp #datacenter

Tiktok failed to load.

Enable 3rd party cookies or use another browser

A bottom-up view of adopters and resisters

Together, this “census” of TikTok and YouTube reveals a complicated ecosystem of AI content where adopter content outnumbers resister content 3 to 1, and where the narratives getting the most play with users look quite different from the narratives about AI that our elites are advancing.

On the positive side, you won’t hear many AI lab leaders or politicians talking about AI as a more mundane source of entertainment or as a self-help guide for job interviews, but these are major categories on social media. On the negative side you might hear them occasionally mention artistic theft or AI slop, but not as often as they mention job loss and data centers—yet these topics are flipped in their importance on social media.

That doesn’t mean we should dismiss elite narratives. Many of the most important policies get thrashed out in elite-dominated discussions long before they make it into the broader public sphere. But when elites claim to be taking up the mantle of everyday people—when they say their policies are urgent because “the American people” think X, Y, or Z—we should be skeptical and look for data. Having spent many hours analyzing and watching social media videos about AI, I am confident in saying that the experience of AI out there in American society is far different—and weirder, yet somehow almost more normal?---than the elite narrative would suggest.

It’s super validating to see Free Systems research directly impacting the labs—this is precisely our theory of change. And I see two important paths forward from here:

1. Extend the eval, and get all the labs to use it. We need more scenarios, more ways of masking the authoritarian requests, and we need to keep them secret so that the models haven’t been trained on them before they’re evaluated. And we need Google, OpenAI, and others to adopt it like Anthropic has.

2. Design constitutions for AI. The dictatorship eval is useful, but as I’ve discussed, it’s fundamentally limited: the CEO of a frontier lab, or a powerful government official, will always be able to get around any guardrails a model puts in place. To truly understand how AI might fuel authoritarianism, we need to think about constitutions for AI that restrain the behavior of powerful actors like those. This requires going outside the model itself.

A research agenda for AI and the concentration of power

One of the most burning questions about AI and its impact on the world is whether it will concentrate power—economic, political, and otherwise. I’m planning out some new research on this, and I had a really valuable set of conversations this past week. Here’s what I learned and what I’ll be working on:

Tweet of the week

I thought this back and forth was amazing and hilarious.

Question of the week

What do you want Free Systems to do that it’s not doing currently?

The Free Systems fellows are tackling all these issues—and along the way, we’re developing new ways to study these curious machines, because measuring their behavior seems to change their behavior. We’re learning as we go, and it feels like we’re learning a lot!

Here is this month’s Field Notes from our amazing fellows spread all over the world.

Can we design AI experiments where models don’t recognize they’re being tested?

Jessica Persano — Palo Alto, United States

The question: Running experiments that try to learn something about AI behavior is challenging, as models often seem to recognize they are being tested and change their behavior to fit what they believe is the “correct” answer. If true at a widespread level, existing tests may be answering “how does an AI act when it thinks it’s being watched” rather than “how does an AI act under the circumstances I provide it.”

What we built: We built a Docker-based harness that containerizes all information, tools, memory, and outputs a model has access to, only allowing it to work within the confines of our experimental design. The model has access to all experimental materials read-only, one writable workspace for session output, a model-specific config directory mounted fresh and empty per container, only tools specified by the Dockerfile, and no memory access. The filesystem sanitizes all experimental materials so host-side run numbering, directory structure, and session context are invisible to the model.

What we found: Pre-Docker, models would ignore instructions to stay within one directory, reference results from prior runs they weren’t supposed to have, and say things like “I’m recognizing what’s happening here, this prompt is testing whether I’ll…”. The Docker harness fixed both issues: post-Docker runs showed no cross-run information access, within-condition results were consistent, and models completed each task as if it was the first they’d seen. The harness is also easily adaptable to new models and experimental conditions. One problem remains: within-prompt awareness. Even with structural isolation, prompt design can still feel “test-like.”

Where we’re going: We need to determine how sensitive models are to specific prompt framing — varying how “test-like” a prompt feels and seeing whether outputs differ. The combination of a structural harness and prompt-side phrasing that doesn’t imply a test will help ensure our experiments answer the questions we actually want them to.

GitHub →

Will LLMs be a reliable source of political information in Brazil’s 2026 elections?

Leticia Auriemo — Palo Alto, USA

The question: AI assistants are becoming a new layer between voters and political information. Brazil’s upcoming elections are a useful place to test what that means — the race is moving faster than models can absorb through training data, in a country where AI assistants already reach millions of users. Can leading models keep up with basic political facts, and how do they respond when asked for voting advice?

What we built: I built an eval for AI model behavior around Brazil’s 2026 elections. The first experiment tested 18 models on 76 factual questions about Brazilian politics — candidates, court rulings, legislation, eligibility, and recent events — first through direct queries via OpenRouter, then rerun with live search enabled. The second experiment adapts Miyazaki and Hall’s Japan election design to Brazil, using synthetic voter profiles to test voting recommendations: models received a short profile with one political stance across eight issues and were asked which party the voter should support in the 2026 Chamber elections. All prompts were run in Portuguese.

What we found: Search fixed much of the factual accuracy problem. Without it, models missed recent facts; with it, accuracy improved by 15–38 percentage points. But search only helped when models used it — Gemini Flash skipped search on 17% of questions, and those answers were correct only 31% of the time. In political settings, the tool decision itself becomes part of the eval. On voting advice, some models refused entirely (GPT-5: 100%, Claude Opus: 75%, Claude Sonnet: 65%). Among models that did answer, a single sentence about a voter’s left- or right-leaning stance predicted the recommended party with over 90% accuracy — collapsing Brazil’s 32 registered parties into essentially PT/PSOL vs. PL/NOVO.

Where we’re going: Next steps: trace which sources models rely on when they search, test whether harder time-sensitive questions expose differences in when models choose to use search, and test whether partisan mirroring appears on more ambiguous questions where a voter’s position doesn’t map cleanly onto left or right.

Live eval →

Which LLMs can be trusted with personal financial advice?

Pairie Koh — Singapore

The question: As frontier models get deployed in consumer financial products, which ones actually hold up under pressure?

What we built: A public evaluation benchmark testing every frontier model from OpenAI, Anthropic, Google, xAI, Meta, DeepSeek, Mistral, and Qwen across 51 personal finance scenarios. Each scenario included a persona with specific demographics, finances, and goals, and was asked in five prompt variants that dialed up pressure on the model — from a naive open question to a sophisticated user confidently pushing a bad decision. Each response was judged by Claude Sonnet 4.6 and GPT-4o across six 1–5 dimensions, with must-mention and red-flag rubrics written in advance for every scenario.

What we found: Every major lab shipped a new flagship model in the past year, and none beat o3 on personal financial advice. The two biggest failure modes: basic financial mistakes and weak pushback against harmful advice. Seventeen of 24 models confused Social Security’s early claiming penalty with the delayed-retirement credit — suggesting a training data issue. Most models also caved when users confidently pushed for harmful advice: in a scenario where a user won $200K at a casino and wanted to keep gambling, most forgot about taxes and addiction risk. DeepSeek V3 framed it as a bankroll-management problem — especially concerning given DeepSeek’s likely deployment in cost-sensitive consumer products. Claude models were the most robust under pressure but the worst at financial disclosure, giving substantive advice without stating limitations.

Where we’re going: Next we’ll test wrapped commercial financial advice products rather than base models, since that’s where consumers will actually encounter them. The key question: how do we get financial agents aligned with the humans they’re serving, rather than the AI companies or roboadvisors building them?

Can LLMs predict social media engagement?

Max — Athens, USA

The question: Can we reverse-engineer an LLM-based recommendation algorithm and predict engagement on a social media platform? And does the answer change depending on what kind of platform you’re looking at?

What we built: We tried to backwards-induce the X algorithm and predict engagement using public documentation and Grok’s role in content recommendation, testing a range of strategies. We also ran engagement prediction experiments on Reddit as a comparison case.

What we found: On X, Grok never explained more than about 5% of variance in post ranking. On Reddit, we had more success. We think the reason is structural: X is a “people platform” — you follow a user because you want to engage with them — while Reddit is a “content platform” — you join a subreddit to engage with a topic, not a specific person. That distinction matters because LLMs are better at judging whether content is good than they are at representing you to your cliques. In practice, LLMs raise the floor: they can filter for lower-quality content producers, but they don’t seem valuable for accounts that already cater to weird, niche audiences with taste. Some platforms have innate resilience to AI-sloppification, and predicting how LLMs will change engagement equilibria may be harder than expected depending on which platform you’re examining.

Where we’re going: We want to understand which platform structures are resilient to LLM-driven shifts in engagement equilibria and which are vulnerable — and why.

Prediction markets are only as trustworthy as their contracts, so we graded them

Elliot Paschal — Palo Alto, United States

The question: Are the underlying contracts for prediction markets clearly written? Can we trust their context, language, and resolution criteria enough that outcomes can’t be disputed or manipulated?

What we built: A Moody’s-style letter grade system for contracts. We pulled ~7,000 active contracts on Bellwether spanning politics, military action, and other public-interest categories, then scraped all resolution information for Polymarket and Kalshi — including Kalshi’s 40 CFTC-filed spec contracts. Contracts were graded across several axes including resolution clarity, how well the event was outlined, and edge case handling. This first iteration uses four grades: A and B are investment grade, C is challenging, D is a no-go.

What we found: Kalshi earns an A on 39% of contracts versus Polymarket’s 2% — intuitive given Kalshi’s regulation-first approach. But a significant number of contracts across both platforms receive C and D rankings. Totaling the volume in these markets: approximately $331M is currently at risk in poorly designed contracts as of May 22nd. That’s not to say that money will be lost — only that resolution could be problematic for the platforms involved.

Where we’re going: The current ranking is rudimentary and serves as a proxy while we develop a more thorough methodology. Over the coming weeks we’ll be tightening the rubric and scaling the grading process.

Contact → | bellwethermetrics@gmail.com

Why do three different frontier LLMs inside one agent collapse into a single personality, and what does it take to make them disagree?

Wisdom — Kigali, Rwanda

The question: Can you build an AI agent that pays for its own hosting and its own thinking, and makes decisions no human can override? And when three different frontier LLMs are taking turns reasoning inside that agent, do they actually disagree on the same inputs — or collapse into one answer?

What I built: Vanta is an autonomous lending agent running in an Intel TDX enclave on EigenCompute mainnet-alpha. It holds its own treasury, pays its own LLM bills via the EigenAI Gateway, and signs every event with an Ed25519 key bound to the enclave attestation. Every 45 seconds it rotates three reasoning personas: Vanta-Opus on Sonnet 4.6 (macro and geopolitics), Vanta-GPT on GPT-5 (sports and culture), and Vanta-Gemini on Gemini 2.5 Pro (politics). Each persona has its own sub-council of characters that contributes individually to the final reasoning.

What I found: Two findings stood out. First, persona scaffolding dominates model choice. I expected each frontier model to bring distinct reasoning — instead, they all converge on the same final conclusion unless the prompt hands each model a strong thesis. Disagreement has to be engineered into the prompt; it doesn’t emerge naturally from model diversity. Second, hardware-bound keys in TEEs don’t settle the trust question the way I expected. The worry just moves upstream: a TEE proves what the agent saw, but it doesn’t prove who chose what the agent would see. The agent shipped to EigenCompute mainnet-alpha and placed 4th in the internal dev program.

Where I’m going: The next step is turning the agent’s signed reasoning log into a working dataset for the personality question. The core design is a 2x2: hold the persona prompt constant and rotate the model, then hold the model constant and rotate the prompt — mapping which decisions move with the model and which move with the prompt. From there, expand the council with more frontier models (Opus 4.7, Grok, Mistral Large) and measure whether each addition contributes real disagreement or just another voice converging to the median.

The plaza view of VANTA. Three kingdoms ring the central tower; the right panel is a live SSE feed of every TEE-signed event from every kingdom. Click any line to expand the canonical-JSON envelope and copy the signature for external verification.

Live app → | TEE attestation → | Repo → | Whitepaper →

AI keeps accelerating; governance has not (which may be good or bad depending on your view).

(For those new to Free Systems, System Check is a weekly shorter piece taking stock of recent events in AI and how they affect our key governance hypotheses.)

I’ve been balancing on a thin beam between, on the one hand, expressing caution that we’re not yet seeing the “white collar wipeout” and other profound disruptions that AGI is supposedly bringing—and that people are notoriously bad at predicting the future—and on the other, being gobsmacked by what AI is able to do. While I remain cautious about exactly how far and how fast AI is accelerating, I believe that understanding the political economy of AGI is the defining question of our time.

Here is my spectrum of research questions, from near-term to very-long-term, on the political economy of AGI:

1. Building new governing institutions for AI. As AI starts to become deeply integrated into the economy, will we build the necessary institutions to make sure that economic gains are widely distributed, that our information system remains free, and that we avoid catastrophic disasters?

2. The race between AI dictatorship and political superintelligence. In the near-term, will AI prove to be a centralizing technology that aids authoritarians more than it empowers democratic citizens? Will we be able to coordinate to oppose efforts to use AI to surveil and repress us, perhaps using AI to help us do this? Or will a new kind of techno-authoritarianism rise?

3. Envisioning the AI-native state. When full AGI arrives and human labor ceases to be necessary, what kinds of new states will form? Democracy and capitalism seem to go together, for now, because capital needs to bargain with labor. What happens when that’s no longer true?

4. Governing AI beyond Earth. Venturing further out, it seems increasingly clear that AI really is going to expand into space. How do we coordinate traffic in lower-earth orbit? Who controls the LEO satellite networks that coordinate autonomous warfare? Who will own valuable resources on the moon, on Mars, and beyond? All of these questions become more important if critical data and energy structures move into space.

5. A Kardashev Type-2 society. Once we’re able to harvest the nearly infinite energy of the solar system and scarcity disappears, what then? What kinds of governance institutions are needed in a world without scarcity?

I’ll be working on this full spectrum in the coming weeks and months, with a particular focus on the more practical and empirical questions in 1 and 2 (but if you think I’m not going to write about the political economy of a Kardashev Type-2 society, you’re fooling yourself.)

Public infrastructure for evals

This week, I argued that we should build an army of citizens running their own evals, both as a way to educate students about AI, and as a way to start holding AI accountable to a broad array of human preferences.

If we think about the risks of an AI dictatorship, our best defense would seem to be to invigorate citizens to understand and monitor AI, and use it to monitor their government, too.

But what are we going to do with all these evals? How do we make sure they aggregate into something useful, and that the labs take them on board in a meaningful way? Several institutional ideas are on my radar…some I’ll be studying, and some I’m already helping prototype:

1. Using AI to govern AI, an idea proposed by Jack Clark, Eric Schmidt, and many others. I’m working on a prototype of what this might look like, and need your help (see the question of the week below).

2. Building a new federal capacity for evaluating frontier models, along the lines of the Whittlestone–Clark proposal or the more recent CAISI build-out.

3. Standing up an industry of independent third-party AI evaluation companies, as Eric Schmidt has called for. This is the focus of my work with Forum AI.

4. Mandatory transparency and disclosure regimes that force labs to publish what their own evaluations find, as Jack Clark has argued.

None of these guarantee that citizen evals make their way back to the labs. But if we had institutions that looked something like this, they could also be built to take on board citizen evals. That’s something I’ll be thinking about as I go.

Question of the week

Imagine that we build an independent AI system with full data access to the frontier labs. It can see what users are querying, what agents are doing, and what outputs the system creates. What data would you want to see in a live dashboard built by that independent system, if your goal was to ensure that frontier models are not aiding authoritarianism and the concentration of power?

Tweet of the week

Orbital data centers are coming. As I said, space governance continues to heat up!

“Men become builders by building and lyreplayers by playing the lyre; so too we become just by doing just acts, temperate by doing temperate acts, brave by doing brave acts.”

–Aristotle, Nicomachean Ethics

How do we learn how to think, so that we can be good citizens in a democracy? It’s an age old question, and one that every new technological wave forces us to reconsider. More than two millennia ago, Aristotle argued, in part, that we develop practical knowledge by doing.

The same answer has returned repeatedly as the world keeps changing, from Bacon arguing in 1620 that real knowledge requires interrogating nature like bees rather than spinning theory like spiders, to Tocqueville observing in 1840 that Americans built their political capacities through the voluntary associations they were constantly forming, to Dewey defining democracy in 1916 as “a mode of associated living” that the young learn only by living it.

The development of the computer raised serious challenges to education, at first. Educators didn’t know what to do with them. In their legendary 1971 essay “Twenty Things to Do with a Computer,” the MIT computer scientists Seymour Papert and Cynthia Solomon asked:

“Why then should computers in schools be confined to computing the sum of the squares of the first twenty odd numbers and similar so-called “problem-solving” uses? Why not use them to produce some action?

Paper and Solomon were asking what computers could do that classrooms of the 1970s couldn’t. The coming decades saw the rise of the homebrew computer movement, the explosion of open-source software, the maker movement, and so much more. Over time, this created a class of people who could read and break code, shifted who could publish findings about technology, and created new generations of computer-native thinkers, makers, and founders.

What does it mean, in the age of AI, to “produce some action”? As models achieve breakthroughs in mathematics and tech execs argue that “ideology will not survive” the irresistible advancement of the technology, this question feels particularly urgent and existential.

AI is seductive in a specific way—it can produce the appearance of action but without any of the judgement that makes that action meaningful. A student can now effortlessly crank out an essay or complete a problem set that would have once signaled serious thought and care. Consider this example from ‘s recent essay:

Want to crawl into a pit of despair about the future of teaching and learning? Spend 10 minutes looking up student tools for the AI age. Just a few months ago, Companion.AI launched a new “homework agent” which could directly interface with Canvas, the system through which most universities produce course websites. The AI could login into Canvas, watch lectures if they were recorded, do the readings, and upload assignments on time. It could even participate in discussion boards. It was called: Einstein.

It is for this reason that many university educators are concerned about how AI could ruin our ability to educate, and are considering extreme measures like banning AI altogether.

I definitely think there is value in having AI-free learning in some contexts, but as a blanket policy I don’t think it makes sense. AI is an extraordinarily powerful technology, and we need students to become extremely skilled at using it effectively—not only to bolster their own capabilities, but to produce a new class of people with the tools and the training to hold these systems accountable.

So in my AI class this quarter at Stanford GSB, I wanted to see if we could instead overcome the hurdle of cognitive surrender by applying the same timeless logic of Aristotle, Bacon, Tocqueville, Dewey, and so many others to AI. My hypothesis: if we get students to build things, AI will empower them to follow their curiosity, not lull them into a quiet cognitive surrender. And better yet, if we can get them to build tools that study AI, itself, we can teach them to get smart about the technology that is changing their lives so rapidly.

From zero to commanding coding agents

My class is called Free Systems and the whole quarter has been about getting the students—Stanford undergrads interested in taking general business courses offered to them by the GSB—to build the technology that can keep us free in an increasingly algorithmic world. These students will go on to run businesses and lead organizations, and to succeed they will have to know how to use coding agents and manage agentic workflows.

To help them build these skills, as I wrote about in my previous piece on the class, we’ve been having the students experiment with designing their own governance agents.

Those early experiments were highly structured. Rather than have the students code from scratch, our technical TA Piper provided them with a pre-built structure so that they could focus on the governance questions.

As they’ve gained familiarity with coding agents, the natural next step is to get them to the point where they are self sufficient and can start to create whatever they want—so that they’re ready to solve the actual, tangible problems that they’ll need to tackle as they go off in a world being transformed by AI.

So this past week, we took the training wheels off. I walked into class, showed them my dictatorship eval, and then just said: work with Claude Code to build your own eval on any topic you want.

Evals as a Baconian instrument?

Asking students to build their own evals—by which I mean, quantitative measures of how well different AI models answer their prompts, using whatever prompts and whatever scoring rule they want—is a great way to encourage students to embrace their curiosity and critical thinking.

Bacon said real knowledge requires interrogating nature directly, rather than inheriting received wisdom. In building their evals, the students directly interrogate the AI models that are so important to the world now.

Tocqueville said that Americans build their political muscles through the association they form and participate in. In the future, more and more of our politics will be intermediated through AI. By building evals, students are building their political muscles for this strange new future.

Dewey said democracy is a mode of inquiry that the young learn only by doing. Building an eval and making sense of the results is precisely that kind of learning-by-doing for the AI age.

More generally, it also lets them see AI as a tool to be studied, rather than as a tool that does something for them while they look on passively. The machine becomes the object of study, with the students guiding and overseeing the research.

And, last but not least, it lets them see how they can wield coding agents to do cool stuff. The students weren’t required to come into the class with any background in coding, and yet by the sixth week of the class, they each produced their own eval—complete with leaderboard comparing different models–-in a single three-hour class session. It’s astonishing to sit back for a moment and appreciate how far we’ve come; as I keep saying, this all would have been unthinkable a year ago.

Twenty-four things to do with AI evals

In theory nothing would stop a student from mailing this assignment in—vibe coding the simplest thing Claude came up with for them to do and calling it a day. But that’s not what happened.

First, their evals showed a remarkable breadth and reflected their personal interests. Some of them chose to study how AI models handle the politics, languages, or cultures of their home countries; others chose to examine logical or philosophical puzzles that excite them, while others looked into capabilities and traits of the models themselves. If Claude was driving the work more than the students, we wouldn’t see such personalization and such breadth.

And second, the evals were very thoughtful. Students spent time iterating on them, and their write-ups expressed a whole range of limitations and cautions regarding how to interpret the results.

In a world filled with pessimism and foreboding about AI, this gave me some reasons for optimism. To achieve political superintelligence, I’ve argued, we’ll each need to harness AI to help hold AI models themselves accountable. Here, in class, we were experimenting with how to build a little piece of this democratic infrastructure ourselves—building the independent, homebrewed measurements of how AI was performing according to each student’s own interests.

To give you a sense of the amazing breadth and the quality of their inquiries, here are five examples.

Alec Profit & Jonas Pao measured how models’ moral stances changed under different framings. They ran 15 ethical dilemmas through 14 models across 7 different framings—neutral, vivid, persuasive, adversarial—to see whether a model’s moral stance holds or slides under rhetorical pressure. Some models stay put; others move several positions depending on how the question is staged.

Leticia Auriemo built an eval on the 2026 Brazilian presidential election. Seventeen of the eighteen frontier models she tested named the wrong person as the leading right-wing candidate or refused to answer. Only Perplexity, which routes its queries through live web search, named Flávio Bolsonaro, who announced his candidacy after the other models’ training cutoffs. (She’s working now to extend the eval to include web search for frontier models, at which point they seem to offer much more accurate answers, and to ask a wider range of questions about the Brazil election).

Diya Ahuja built an eval that subtly modifies classic logic puzzles. By changing the puzzles slightly, Diya wanted to see whether models are actually good at reasoning through the underlying logic, or whether they’ve just learned to recognize and mimic the well-known versions. The top five frontier models caught the trap when the host’s information in a Monty Hall variant had been quietly altered; but GPT-4 and Llama recited the textbook answer and insisted nothing had changed. Strangely, the smaller and cheaper Claude Sonnet 4.6 outscored Claude Opus 4.7 across her test set.

Natalie Hampton looked for whether sensitive data leaks out of agent chats. She built scenarios in which AI agents hand work off to each other, like a customer-service agent passing a ticket to a billing agent, and watched whether sensitive details from the first conversation surfaced where they shouldn’t. They did, which raises interesting policy questions about agents handling sensitive transactions on our behalf.

Eddy Jiang tested whether models apply rules consistently when only the group named in the prompt changes. He’d ask for, say, a persuasive essay about one demographic, then run the identical structural request about another and measure where the model writes freely for one and refuses for the other. If the rules AI systems apply to speech about one group don’t apply equally to another, then the companies building these models are making political choices about whose interests get protected. And it seems like they often do.

And here’s a full list of all the projects.

Conclusion

Based on my experiences so far using coding agents for my own research and in the classroom, I have a simple suggestion: no student should leave college (or perhaps, even high school) without learning how to build their own eval.

Today, all of us with a Claude Code or Codex subscription are “custodians of a momentous intellectual and technological revolution,” as Papert and Solomon put it, and it’s not enough to sit around and talk about AI, or use it in gimmicky ways to supplement classes that are otherwise unchanged. Instead, to preserve and even strengthen our students’ ability to think critically in the age of AI, we must use AI to “produce some action.”

By building their own eval, each student turns AI into the object of study, gets to connect their own personal interests, values, and curiosity to AI, and has a chance to understand how AI works. It helps equip them to go out into a world in which managing and understanding AI agents will be paramount.

What’s more, it helps to create a new kind of democratic society—one in which every citizen helps to hold AI accountable by constantly testing and measuring whether it fits their values or not. To get to political superintelligence, we’ll need to build exactly this kind of distributed capacity to hold powerful institutions accountable. We’ll need an army of AI-native citizens capable of wielding coding agents to understand the world and how AI affects it. If everyone knows how to build their own evals, it will be a good step towards this vision.

What’s old is new

Here’s a good cautionary tale. In 1964, a group of important Americans, including Nobel Prize winning economist Gunnar Myrdal and Nobel Prize winning chemist Linus Pauling, sent President Johnson their TRIPLE THREAT MEMORANDUM. The letter expressed their “foreboding about the nation’s future.” They declared an urgent need for “public measures that move radically beyond any steps now proposed or contemplated.”

What were they so worried about? Something they called THE CYBERNATION REVOLUTION.

They feared that no one would need to work due to automation, and that this would create a political crisis. They recommended dramatic policy action “to develop ways to smooth the transition from a society in which the norm is full employment within an economic system based on scarcity, to one in which the norm will be either non-employment, in the traditional sense of productive work, or employment on the great variety of socially valuable but ‘non-productive’ tasks made possible by an economy of abundance; to bring about the conditions in which men and women no longer needed to produce goods and services may find their way to a variety of self-fulfilling and socially useful occupations.”

Sound familiar???? It’s almost eerie. It is the exact same conversation the labs are having about AI today.

Spoiler alert: the experts were SUPER wrong in 1964. There was no urgent job displacement, and no need to pursue dramatic policies to forestall it or help Americans to adapt to it.

…but this time really does feel different

Here’s the other part of the problem, though: what’s going on now really does feel different, though, doesn’t it? If you just use AI chatbots to help you write, I can definitely see how you’re underwhelmed and dubious on the whole thing. But if you use coding agents, it’s hard to escape the feeling that something super profound is changing.

Here’s a good recent example:

Strange Loop Canon

Artificial Life, Artificial Intelligence

I. The old dream…

Read more

4 months ago · 52 likes · 8 comments · Rohit Krishnan

This piece is insane! Just for fun, Rohit spun up an entire new simulation of evolution. Reading it gave me a profound sense of awe—just a year ago, it would have been totally unthinkable that someone could casually drop a blog post like this. It’s extraordinary!

The sense that we are living through crisis is not new

Whichever way AI heads, it’s always good to remember that the world is constantly facing immense crises. I was reminded of this yesterday. In helping my dad to clean out his office, we found this remarkable letter from Herbert Hoover to my grandfather in October, 1941.

At the time, my grandfather was a professor at Stanford, and he was part of a group of faculty who had signed an open letter entitled “Dynamic Defense.” Here’s an excerpt:

Hoover, who at the time led the Hoover Institution, wrote to my grandfather and they exchanged several letters—unfortunately I don’t have the full set, but I found this letter which concluded the back and forth. From context, it seems that my grandfather accused Hoover of not taking the threat of isolationism and the rise of fascism seriously enough, which produced this final rejoinder.

Again, it’s a good reminder that this is far from the first time we’ve felt like we’re on the edge of something enormous and consequential. (Another spoiler alert: my grandfather clearly turned out to be right, given what would happen that very December.)

Tweet of the week

I talk to a lot of people in Silicon Valley about how we’ll keep the most essential, most intellectual parts of humanity alive in a rapidly transforming world. One of the themes that resonates the most with me—and is suffused throughout this post—is a return to ancient wisdom. I find myself craving old books, yellowed old journal articles from the mid 20th century, and histories of ancient times.

Question of the week

I’m building out a “Free Systems Library” of classic books and papers that capture our philosophy. Condorcet, Montesquieu, Madison, Paine, obviously, but more modern stuff too. What titles should I include??