Read that as a design decision and it looks like restraint. Read it as a balance sheet and it looks like something else. Anthropic did not simply decline an aesthetic. It declined a revenue line, publicly, in writing, in a year when every competitor was building one. OpenAI’s CMO said so within a day.

That is what taste actually is, and it is why the word is suddenly everywhere. Not because businesses discovered beauty. Because the price of looking good collapsed, and something else had to start doing the work that polish used to do.

The word everyone is using and nobody is defining

The discourse arrived fast. Fast Company ran a piece in June on winning in the era of taste. Forbes councils have declared human taste the most valuable marketing asset of the year. Adobe surveyed 850 American small business owners and solopreneurs in June and found that 56 percent believe that when everyone has the same tools, the edge goes to whoever applies their own point of view.

Every one of these pieces is directionally right and almost none of them will tell you what taste is. It gets defined as an eye, a filter, a vibe, a sensibility. All of which make it sound like something you either have or you go buy from someone who has it.

I want to argue for a less flattering and more useful definition. Taste is not a sensibility. Taste is a cost you can afford to pay because of something true about your business, and which your competitors cannot afford to pay because nothing equivalent is true about theirs.

What actually happened to the price of looking good

There is a piece of economics from 1973 that explains this decade better than most things written about it this year. Michael Spence’s work on job market signalling made a simple observation: a signal only carries information if it is expensive to send, and specifically if it is cheaper to send for the people it is supposed to identify. A credential means something because it is costly. If everyone could get one on a Tuesday afternoon, it would mean nothing, and employers would go looking for whatever was still expensive.

That is precisely what happened to brand production between 2023 and now. Good typography, a clean site, a competent brand film, decent copy, a coherent grid. All of it used to be expensive, and because it was expensive it functioned as proof of seriousness. Now it costs an afternoon. The signal did not get worse. It stopped being a signal, because it stopped being scarce.

So the market went looking for whatever was still costly, and found judgment. Not the ability to make the good-looking thing, but the willingness to not make it. Patagonia ran a full-page ad in the New York Times on Black Friday 2011 telling people not to buy its jacket, and in 2022 Yvon Chouinard gave the company away rather than take it public, structuring ownership so profits fund climate work.

None of those are aesthetic choices. They are expensive refusals, and the aesthetic is only what the refusal looks like after it has been practised for twenty years. This is the part the taste discourse keeps missing. You cannot install the look without the cost, because the look was never the thing carrying information. The cost was.

The creator economy is a market for lemons

The second reason taste went from a nice-to-have to a survival trait has nothing to do with AI and everything to do with volume.

Every solo consultant, every boutique studio, every founder-led brand now presents at roughly the same production quality. The decks look the same. The sites look the same. The LinkedIn posts have the same three-line cadence and the same false vulnerability. A buyer scrolling through eleven of you cannot distinguish the operator with nine years of pattern recognition from the one who read a thread last month, because both artefacts are equally clean.

When quality becomes unobservable, buyers do not reward quality. They discount everything, and then they buy on price or on whoever was loudest. Which is why so many extremely good people are stuck at rates that make no sense relative to what they can actually do.

Taste is the escape hatch, but only in the technical sense. It is a warranty. It is the thing you say and do that a less capable competitor genuinely could not say or do without getting caught, because it requires knowing something specific, refusing something profitable, or having lived through something you cannot summarize.

What I build before anyone asks for it

The sequence, roughly, is this. First, I get the founder to articulate what they actually believe about their category, and then I stress-test it until it is specific enough to be wrong. Vague beliefs are safe and therefore worthless as differentiation. Second, we build the refusal list, which is the genuinely uncomfortable part, because it means naming the revenue you will not take, the clients you will not serve, and the tactics you will not run even when they work. That list is the taste. Everything visual is downstream of it. Third, only then do we design the expression, and it has to be unique to the business model, because taste borrowed from a brand with different economics is just costume.

I will not pretend the instrument is complicated. It is a set of questions in a specific order, and the order is most of the value. What I will say is that founders consistently underestimate how much of their growth problem is a coherence problem, and how much of what they call a brand is actually just a list of things they have never decided.

Taste is not the pretty part. It is the part that has already been paid for.

There is an assumption I run into often enough that I have started to find it interesting rather than annoying. It goes like this: an agency has ten people, I am one person, therefore the agency should cost more than me. It sounds like arithmetic. It feels like common sense. And it is almost exactly backwards.

I want to explain why, and I want to do it properly, with the research rather than with assertion, because the reasoning matters more than the conclusion. If you disagree with me at the end, I would rather you disagree with the argument than with my invoice.

First, what I actually do, because the misunderstanding usually starts here.

I work with founder-led businesses on three things. Brand strategy, which means positioning, narrative, content strategy, and visibility. Marketing, both organic and paid. And growth systems, which means a single source of truth, a CRM that actually works, and every touchpoint connected so a person can move from stranger to client without falling through a gap.

Most agencies do one of those three. Almost always the second. Some do the second and third. Very few do the first, and that is not a criticism of them, it is a description of how their business model works. I will come back to that, because it is the whole argument.

The thing the data says, that the market ignores

In 2013, Les Binet and Peter Field published an analysis for the Institute of Practitioners in Advertising based on the IPA Databank, drawing on 996 case studies covering 700 brands across 83 sectors over more than thirty years.

Their finding was simple and inconvenient. The optimal split of marketing investment is roughly sixty percent long-term brand building, forty percent short-term sales activation. Not the reverse. As Binet has put it, the sales effects of activation are easy to see and easy to measure, but they only represent forty percent of the payback, while sixty percent comes from long-term brand.

The two do different jobs on different clocks. Brand building creates memory structures and familiarity in the minds of people who are not buying yet. Sales activation converts the people already predisposed to buy. Its effects are fast and transient, and it has little effect on long-term growth if it is not supported by branding. Performance harvests what the brand has sown. Without sowing, the harvest gets leaner every year.

Brands that over-invest in activation at the expense of brand building tend to see diminishing returns over time as price sensitivity increases and brand preference erodes. Price sensitivity increasing is the polite way of saying it. In plain language: the more you rely on performance alone, the more your customers start to see you as interchangeable, and interchangeable businesses compete on price until there is nothing left to compete with.

That is what you are buying when you buy sales without positioning. A short spike, and a slow slide toward being a commodity.

Why this got worse in the last two years

If the sixty forty problem were the whole story, you could argue it has been true since 2013 and the world has coped. But something changed, and it changed fast. Generative AI made competent content free. Not good content. And competent turns out to be the exact thing that stops working when everyone has it.

They work by predicting the most probable next token, which means they excel at summarizing existing wisdom and struggle to generate the spiky, contrarian, or novel insight that arrests attention. The result has been described as content slop: a landscape of digital sameness where brand voices lose their distinctiveness and marketing copy sounds as though it were written by a committee of algorithms.

And it is not only the words. The same training-data dynamics apply to AI-generated visuals. Models trained on the same image datasets produce the same compositional choices, the same layouts, the same palettes that feel modern and clean, the same iconography. A brand isn’t just losing its voice, it is losing its face at the same time.

So the situation is this. Volume is free, sameness is the default, and audiences are actively fleeing it. In that environment, the scarce thing is not more content. It is a business that sounds like itself.

Why positioning is the actual work

Here is the sentence I keep coming back to.

Taste is not something you build. It is something you carry.

You cannot install taste into a business the way you install a tracking pixel. It is not a deliverable. It comes out of a founder’s particular history and judgement and refusals, and the job of a strategist is not to invent it but to find it, name it, and make it legible to a stranger in ten seconds. That is a very different task from running ads, and it fails in a completely different way when done badly.

Let me make it concrete.

Two firms sell apples. The first sells apples. It markets hard, prices low, runs paid and organic, and does everything correctly. Its apples are good. Nothing about the transaction is memorable, because an apple is an apple, and the only lever left is price.

The second firm sells apples from a specific place. Grown in a particular valley, by growers who have done it for generations, harvested in a particular window, and available in a limited quantity each season. Same fruit, chemically. Entirely different business.

This is not a marketing trick. It is one of the better documented effects in agricultural economics. There is even a paper in the American Economic Review titled, straightforwardly, “What Is the Value of Terroir?” There is a study on geographical indication recognition and willingness to pay for the Grójec apple, which is to say someone has literally run this experiment on apples.

But my favourite finding in that literature is stranger and more useful. In an Italian study of protected-origin ham, consumers who lived in the same area where the certified product was made were willing to pay a lower premium than consumers living farther away.

Sit with that. Distance increased the premium. The people close enough to see the ham as ordinary paid less. The people for whom it arrived carrying a story paid more. The story is not decoration on the product. In economic terms, the story is part of the product, and it is the part that survives comparison shopping.

An agency will rarely build this for you. Again, not because agencies are bad. Because positioning cannot be produced the way a posting calendar or a paid campaign can be produced. It cannot be assigned to whoever is free on Tuesday. It cannot be templated across a client roster, because the entire value of the output is that it is not like anyone else’s. A business built on repeatable process has a structural reason to avoid the one deliverable that refuses to repeat.

That is also, honestly, why agency pricing sits where it sits. A retainer built on predictable, delegable, headcount-based work is priced accordingly. It is a fair price for what it is. It is simply not the same product.

The part that costs more, and this is the part I do for my clients

Read more

The confession went like this. The previous season had worked, so he reached for the process that produced it: go somewhere beautiful, absorb it, return to the atelier and translate. He chose Barcelona, and Gaudí, which on paper is unimpeachable; the sacred texts of organic form, the façades that seem to have grown rather than been built. And nothing came. In his own account he entered a possibility-denying, misery-making cycle, the formula strangling the work it was supposed to produce. The retreat that had once been a source became, when repeated, a cage. He had found the thing every ambitious person eventually finds: that the surest way to lose the magic is to try to reproduce the conditions under which it last appeared.

What he did next is the reason the collection matters. He mentioned it within these words: The Call of the Void - l’appel du vide, that intrusive flicker at a great height that whispers step off, and he stepped. Off the method, off the reference-and-translate machinery of contemporary couture, into materials the discipline is trained to refuse. Latex. Silicone poured and set. Paint baked into sheets and sculpted like fabric. The anatomical bustiers, sculpted rather than molded from a body, were fabricated with a Paris atelier that ordinarily makes hyperreal silicone infants for film. He took the wheel, he said afterward, and drove off the cliff. The collection is the footage of the fall.

It would be easy, and wrong, to read this as mere shock. Roseberry is working inside a lineage and he knows it precisely. Elsa Schiaparelli built the house on exactly this seam, the surrealist proposition that the beautiful and the unsettling are not opposites but collaborators, that a lobster belongs on a dinner dress and a drawer belongs in a torso. The 2021 anatomical corsets, the gilded body parts, the perennial return to the flesh: this house has always insisted that couture is not decoration laid over the body but an argument about the body. The silicone is not a stunt. It is the maison’s oldest idea, spoken in a material Elsa never had.

But the collection is doing something beyond honoring its own DNA, and Roseberry said what it was, almost in passing, in the way people say the truest things. He had recently reviewed portfolios at a university and found AI throughout the students’ senior work. His unease was not the familiar complaint about craft or shortcuts. It was subtler and more difficult to dismiss. When the work is generated, he said, you don’t feel you are getting a reflection of the person you would be considering.

Sit with that, because it relocates the entire argument.

For two years the debate about machine-made work has been an argument about quality, and that argument is finished. The models make good things. They make a good dress, a good essay, a good portfolio, and they make them faster than any human atelier can. Anyone still insisting the outputs are bad is losing an argument with the evidence. But Roseberry has quietly declined to have that argument at all. He is not asking whether the work is good. He is asking whether it is a reflection of a person, and proposing that in the domains that matter most, that is the only question that was ever load-bearing.

Consider what the void actually is, once you strip away the French and the drama. It is the precise part of creation that cannot be optimized. It is the moment the reliable method dies in your hands and you are left with no choice but to surrender to something you have never done and cannot be sure will work, and you do this as a particular person, with particular fears, standing at a particular edge that is yours and no one else’s. This is the one thing a model structurally cannot do. It has no edge. It has a training distribution and a temperature parameter. It can produce the appearance of surrender flawlessly, because surrender has a describable style and style is exactly what it learns. What it cannot do is be at risk cannot actually fear that this collection, this one, will be the failure, because it has nothing staked on the outcome and no cliff beneath it to give the height meaning.

So Roseberry’s collection resolves into a single proposition: the fear was the content. The block, the misery, the surrender, the turn to silicone and latex, these are not the backstory to the work. They are the work. You can feel the person in the clothes precisely because he very nearly could not find him, and the trace of that search is the thing on the runway. The garments are legible as his in a way a flawless AI-generated Schiaparelli-in-the-style-of never could be, not because the machine’s version would be worse, but because it would be no one’s.

This is where couture stops being a niche and starts being a diagnosis of the culture around it. What Roseberry has identified on his runway is the scarce resource of the next decade, and it is not competence. Competence has been commoditized; it pours out of the machines by the terabyte and it will only get cheaper. The scarce thing is legible humanity; the perceptible sense that a specific person stood at a specific edge and chose to step off, and that what you are holding is what they came back with. In a market drowning in the competent and the plausible, the only durable premium is the signature of a self. This is as true of a fragrance, a building, a company, a sentence, as it is of a silicone gown.

He showed it, fittingly, on a mirrored runway at the Petit Palais. I doubt the mirror was decoration. A surrealist’s mirror never is. He wanted the room to see itself inside the work to understand that the reflection was the entire proposition, the whole of what was for sale and the whole of what could not be faked. A machine, asked to produce the same collection, would have handed back something cleaner. Better lit, more symmetrical, free of the misery that made it.

It simply would not have been a reflection of anyone.

I want to be careful here, because I am not writing this to be cruel to people who were not treated gently themselves. I owe the generation above me a great deal, and some of what I owe them is the exact thing I am about to question. So let me say the honest version first. There is no substitute for experience. There is a kind of knowing that only arrives after you have lived long enough to be wrong many times, and I do not have it yet, and they do. That is real, and I am not going to pretend it away to make my argument cleaner.

But here is the thing I have started to notice, and once I noticed it I could not stop. Older than you is not an argument. It is a fact about time. And somewhere along the way, in the culture I come from, a fact about time got promoted into a theory of truth. The claim stopped being I have seen more and became I am more. Those are not the same sentence. One of them has diagnostic value. The other is just a hierarchy wearing the costume of wisdom.

This is a letter about the difference. It is also, in the end, a letter to myself, because the frightening part of writing it is realising how easily I could become the thing I am describing.

The scoreboard we were raised on.

Every South Asian childhood has a scoreboard, and you are always losing on it.

Your cousin got into medical school. The neighbour’s son is in Canada now, dollars. That girl from the family you see at weddings is engaged to a doctor. You learn the metrics before you can question them: the profession, the visa, the salary, the skin tone, the marriage, the square footage. Your worth is not a thing you possess. It is a position you hold relative to a room full of people you did not choose to be compared to.

There is a name for the machinery underneath this, and it is one of the oldest findings in social psychology. In 1954 Leon Festinger proposed social comparison theory: human beings have a fundamental drive to evaluate themselves, and when there is no objective ruler available, they measure against other people. Festinger’s less quoted insight is the one that explains the aunties. We do not compare ourselves to just anyone. The comparison bites hardest when the other person is similar to us, because similarity is what makes the comparison feel fair, and therefore relevant. This is why your parents never compared you to a stranger. They compared you to your cousin. The closer the blood, the sharper the knife.

And when that comparison comes out unfavourable, it produces a specific feeling researchers call relative deprivation: not the pain of lacking something, but the pain of lacking it when someone like you has it. You can be objectively fine and still feel robbed, as long as the person across the dinner table has slightly more. An entire culture organised around upward comparison is an entire culture organised around a feeling that can never be satisfied, because there is always a cousin doing better, and if there is not, there is always a cousin’s cousin.

Here is what I have come to believe. The older generation did not invent this because they were cruel. They inherited an instrument and handed it down without a manual. What they never got to ask was whether the scoreboard measures anything worth winning.

The professions we were told to worship.

Become a doctor. Become a lawyer. Become an engineer. And underneath it, unspoken, the real instruction: become someone I can say a single word about at a wedding and watch the other person’s face change.

I want to be fair to this, because there is a lazy version of my argument that just mocks immigrant parents for wanting stable careers, and that version is beneath the people I am writing about. Medicine is not glamorous, and that is precisely the point they got wrong, but not in the way the mockery suggests. Medicine, law, the serious professions, they are service. They are long nights and other people’s emergencies and a life arranged around someone else’s crisis. There is prestige in them, earned prestige, but the prestige is the smallest and least honest part of the story. Nobody sat their child down and said: if you do this, you will belong to your patients, your clients, your cases. Your time will not be yours. That is a noble thing to choose. It is a cruel thing to be assigned.

What was really being worshipped was never the work. It was the word. The label that travels, the profession that photographs well in conversation. And this is where the culture reveals what it actually values, because it will take a miserable child in a prestigious profession over a fulfilled child in an illegible one, every single time.

The psychology here is not subtle once you see it. Edward Deci and Richard Ryan spent forty years building self-determination theory, and its central finding is close to the opposite of how we were raised. Human beings have three basic psychological needs: autonomy, the sense that your life is your own; competence, the sense that you are good at something; and relatedness, the sense that you belong. When those needs are met, motivation becomes intrinsic and durable and people flourish. When behaviour is driven instead by external reward and punishment, by the need to satisfy a parent or win an approval, the motivation is fragile and the wellbeing underneath it thins out.

Now, I can hear the objection, because I have made it to myself. Autonomy sounds like a Western value. Is this just an imported individualism dressed up as psychology? It is a fair challenge, and the researchers actually answered it. Chirkov and Ryan ran the study across cultures and found something important: autonomy is not the same as individualism, and it is not the same as independence. You can autonomously embrace a collective value. A person can choose their family, choose their duty, choose to care for their parents, and that choice is autonomous as long as it is genuinely theirs and not merely installed in them. The need for autonomy is not a Western preference. It is human. What varies by culture is only how much room a child is given to author their own reasons.

The tragedy of the professions we were told to worship is not that they are collective, or dutiful, or family-serving. Those can be beautiful reasons to choose a life. The tragedy is that we were rarely allowed to choose them. They were handed to us with the choosing already done, and a handed-down life, however prestigious, sits on the same shelf as a golden spoon. It looks like a gift and functions like a cage.

The gossip that was never really about us.

If you grew up in this culture you know the sound of it. The phone calls after the wedding. The dissection of who wore what, who married whom, whose son failed which exam. The cousin who was a warning one year and a triumph the next. The particular energy, which I can only describe as enemy-to-lover, of relatives who compete with you relentlessly and then weep at your departures.

For a long time I experienced this as simple unkindness. I have stopped believing that. I think the gossip was never really about the people it named. I think it was a generation metabolising its own anxiety out loud.

Consider what the older generation actually lived through. The scholarship on South Asian intergenerational trauma is direct about the sources: colonisation, partition, displacement, forced migration, and the long economic precarity that followed. This is a generation, and a generation before it, that learned in its bones that stability is temporary and the world is not safe. Researchers describe how those histories amplify certain patterns in families: caution around instability, emotional restraint, and an overemphasis on protection that gets rerouted through achievement. When you have watched the ground disappear once, you do not raise children to be happy. You raise them to be unassailable.

Gossip, in that light, is a threat-detection system that never got switched off. It is a way of constantly locating the family’s position on the scoreboard, checking for danger, confirming rank. The aunties are not being petty. They are running the same safety software their own mothers ran, on a world that no longer requires it, and passing the software forward with the updates disabled.

Understanding this changed my anger into something closer to grief. It is hard to stay furious at a frightened person once you have seen the fear. But, and this is the part I will not soften, understanding where a pattern comes from is not the same as agreeing to carry it. You can hold enormous compassion for the wound and still refuse to be the next person it happens through. Both things are allowed. In fact I think that is the whole task.

The projection we are only now learning to name.

Here is the thing that happens when you grow up and start to think for yourself. You question something. Gently, even. You ask why. And the response you get is out of all proportion to the question, because the question was never the real problem. The real problem is that your questioning is experienced as a verdict.

This is the part I had no language for until recently, and I suspect many people reading this do not either, so let me try to name it precisely. When a person has spent their whole life seeking validation, external and internal, arranging their worth around whether they measured up, then a child who declines to measure up is not just disagreeing. That child is threatening the entire framework the parent built their life inside. If worth comes from the scoreboard, and you announce that you do not believe in the scoreboard, you have not rejected a rule. You have rejected them.

So it comes back as projection. The insecurity gets turned outward and lands on you. You are ungrateful. You have changed. You think you are better than us. You have been corrupted by other ideas. And underneath every one of those accusations is a sentence the person cannot say, possibly cannot even hear themselves think: if you are right that the scoreboard is not real, then what was all of my suffering for?

That is an unbearable question, and I have compassion for anyone standing at the edge of it. But I want to say something clearly to my own generation, because we are the ones being handed the bill. Their insecurity is not your inheritance to absorb. You did not cause the wound and you cannot heal it by shrinking. The most loving thing you can often do is to stay steady, stay kind, and decline to accept a debt that was never yours to pay. Respect and self-erasure are not the same currency, no matter how often we were taught to spend them as if they were.

And to return to where I began: the fact of being older confers experience, and experience is genuinely worth deferring to. But deference to experience is a specific thing. It means tell me what you have seen, because you have seen more. It does not mean your opinion is automatically more valid because of the year you were born. The first is wisdom offering itself. The second is authority demanding itself. One of them I will bend toward for the rest of my life. The other I am done pretending is an argument.

What I will not inherit, and the little I hope to pass on.

I do not know if I speak for my generation. I am suspicious of anyone who claims to. So let me narrow it to the only person I can honestly write a rule for, which is the woman I might become, and the child she might one day confuse for a scoreboard.

To that future version of me, here is what I am asking you to refuse to pass on. Do not build a scoreboard. Do not measure a child against their cousins, their neighbours, the engaged girl at the wedding. Do not hand them a profession with the choosing already done and call it love. Do not run the gossip software just because it was running when you arrived. And when they grow up and question you, and they will, do not mistake their questions for a verdict on your life. Their independence is not a referendum on your worth. Let them ask.

And here is the small, unglamorous list of what I actually hope they want from me, when they are grown and I am the older one and the temptation to lecture is strongest.

Confidence. Enough belief in them that they can choose their own path and trust the ground will hold, because I held it while they learned to walk on it.

Empathy. The habit of weighing a stranger’s problem with the same seriousness they give their own, which is the entire difference between a person who has succeeded and a person who is worth knowing.

Humility. This one I want to say plainly, because it is the whole point. Choosing a certain profession does not make you better than anyone. Living in a certain part of the world does not. Driving a certain car does not. Running a large company does not. None of the metrics on the old scoreboard, not one of them, makes a human being worth more than the human beside them. Shine in your moment. God, please, shine. But do not gloat, because the gloating is just the scoreboard talking, and I am trying to raise someone who never learned to hear it.

I know I might fail at this. That is the honest ending. The pull of the culture that made me is strong, and I can already feel, in small ugly moments, how easily I could become the aunty on the phone after the wedding. Writing this is my attempt to catch it early. Consider it a note left for myself, taped to the inside of a door I have not walked through yet.

They were older than me, and they had seen more, and some of what they saw broke them in ways they were never allowed to name. I can hold that with love. I just will not hand it forward. The line stops here, gently, with me.

I want to make a different argument, and I want to make it carefully, because I run a consulting practice on the premise that most claims cannot survive contact with proof. So here is my claim, stated plainly enough to be checked: there is a version of AI slop that genuinely degrades you, and a version that quietly funds you, and the line between them is not effort. It is verification. By the end of this essay you will know exactly where that line sits in your own work.

Not all slop is created equal

First, definitions, because this word is doing too much work. Slop, properly, is generic AI output published as a finished product. You typed a prompt, it typed a paragraph, you hit post. Nobody verified anything. The reader can feel the absence of a human the way you can feel an empty house.

That is different, categorically different, from using AI to activate your research and fast-track your thinking onto a strategic structure. When I use a model to compress forty tabs of research into a brief I then interrogate, or to turn my scattered voice notes into a framework I then pressure-test against a real client’s numbers, the output is not slop. It has been observed. Someone with skin in the game checked it against reality. This is the same distinction I keep finding everywhere I write, in AI governance, in sustainability reporting, in fashion supply chains: the difference between a claim and a proof is whether anyone actually looked.

In 1998, the philosophers Andy Clark and David Chalmers published a paper called “The Extended Mind,” arguing that cognition does not stop at the skull. The notebook, the calculator, the phone: when a tool is reliably available and its outputs are trusted and used, it is functionally part of your mind. Nobody accuses you of not really remembering your schedule because your calendar holds it. The extended mind thesis is twenty-eight years old. Gen Z is simply the first generation to extend the part of the mind that drafts.

When generic AI output is actually good for you

Here is where I break with the purists. If you have paying clients or any consistent source of income, there is a category of work where generic AI output is not a moral failure. It is the correct tool. Specifically:

1. Everyday redundant tasks that need doing, not authoring. Confirmation emails, meeting recaps, calendar wrangling, the invoice reminder you have sent forty times. Nobody’s soul lives in these. Speed is the only virtue they have.

2. Getting things off your plate so your actual creative work gets your actual brain. Researchers Evan Risko and Sam Gilbert named this in 2016: cognitive offloading, the use of external tools to reduce the mental demand of a task. Their finding was not that offloading makes you dumb. It is that offloading frees working memory for the tasks that need it. The essay, the pitch, the sidequest that becomes a business.

3. Your business’s admin and logistics that do not need double-checking. SOPs, internal documentation, the operational plumbing where the audience is future-you, not the market.

4. First drafts of anything internal. The zero draft is where most people stall. A mediocre generated draft you ruthlessly rewrite beats a perfect draft that never exists.

5. Structure before judgment. Handing the model your mess and asking for the skeleton, so the hours you spend go into the thinking only you can do.

Notice what is not on this list: anything a client pays for, anything with your name on it, anything that represents your taste. Slop in the basement, craft in the shopfront.

The generation living the consequences

Gen Z has a controversial reputation on work ethic, I know. We are, allegedly, allergic to paying dues. But here is what I think is actually happening. Every generation before us discussed AI’s consequences as impending. A storm on the forecast. We are the first generation for whom the consequences are not impending. They are the weather. We graduated into the entry-level rungs thinning, into a job market fragmenting into contracts and gigs and portfolio careers, and so we did the rational thing: we started building businesses through the creator economy, stacking income streams the way our parents stacked years at one employer.

And when the ladder is missing its bottom rungs, offloading your menial work is not laziness. It is the only sane allocation of the one asset we actually have, which is time. The redundant tasks go to the machine. The judgment, the taste, the relationships, the proof: those stay with us, because those are what get paid.

Which brings me to the money. This is not theoretical for me. Earlier this year I signed a consulting client, a founder running a boutique service business, and AI touched almost every step between “this person exists” and “the contract is signed.” Not by writing slop at her. By structuring me.

Because here is the thing nobody tells you about working for yourself: the hard part is rarely the work. The hard part is the sequence. What to say first, what to build second, when to name the price. AI tools are unreasonably good at turning your foggy intentions into a decisive mechanism, a defined series of steps you execute instead of a vibe you circle. What follows is that mechanism, from finding her to closing her, plus the four-part proposal structure that did the actual closing.

Read more

A boutique luxury retail business, founder-led, beautiful product, a woman who had built the thing from nothing and had the taste to know what it should feel like. She came to me for brand strategy. The positioning felt muddy. The campaigns were not landing. Revenue was fine and had been fine for a while, which is its own kind of alarm.

I spent the first two weeks doing what I was hired to do. Audience work, messaging, competitive positioning. And I kept running into the same wall, which was that I could not answer basic questions about her own business, and neither could she.

Which campaigns brought in customers who bought twice? She did not know. What did a customer cost her from Instagram versus from email? Nobody had ever calculated it. How many people were in the pipeline right now? Depends who you ask, and there were four possible answers living in four different places.

She was spending real money every month on marketing that could not be evaluated. Not badly. Just invisibly.

By the end of week two I told her I thought we were solving the wrong problem, and that if she wanted, I would rather spend the engagement somewhere else entirely.

This piece is about what I found, why it is almost always what I find, and how you can tell in about twenty minutes whether the thing you are calling a brand problem is actually a brand problem.

The reason she was wrong about her own problem is a well-documented one.

In 1984 Eliyahu Goldratt published The Goal, and the idea at its centre has never stopped being true. Every system has a constraint. One point, usually one, that determines the throughput of the whole. And the corollary, which is the part people skip: improving anything that is not the constraint produces no improvement in the system. It produces the feeling of improvement, and a lot of activity, and metrics that move somewhere other than the bottom line.

Her constraint was not brand. Her brand was, in the pure sense, quite good. Her constraint was that she had no single source of truth, which meant every decision downstream of that was being made on vibes, including all the brand decisions I had been hired to make.

Sharpening the positioning would have felt productive. It would have produced a nice deck. It would have changed nothing, because the mechanism that turns positioning into revenue was not connected to anything.

Her brand looked fragmented because her operations were fragmented. Customer data in one place, campaign data in another, the actual sales in a third, and no one system that could see across all of it. So her marketing spoke in one voice, her follow-up in another, her post-purchase experience in a third, because they were being run by different people out of different systems with no shared picture of who the customer even was. That is not a positioning failure. A customer experiencing three different versions of a company is not confused about the messaging. They are correctly perceiving three different companies.

Kevin Lane Keller’s work on brand equity makes the same point from the other direction: brand strength is built by consistency across every touchpoint, accumulated over time. Consistency is not a creative achievement. It is an operational one. You cannot decide to be consistent. You can only build the thing that makes consistency the default and inconsistency the exception.

Which is the whole argument of this piece, so let me say it plainly. Brand strategy is a promise. Systems are the only reason the promise gets kept more than once.

The diagnostic. This is the part I would keep: If you are trying to work out whether your problem is brand or backend, do not look at the brand. Look at whether you can answer questions about your own business without asking anyone.

Here are the five I now use on every first call.

One: Where did your last ten customers come from, specifically, and can you check rather than remember? If you have to remember, you do not have attribution. You have anecdote.

Two: Ask two people on your team the same number, separately. Pipeline value, leads this month, anything. Different answers means no single source of truth, and every strategy conversation you have is a negotiation between competing fictions.

Three: Is there any number you look at weekly without someone preparing it for you? If a person has to build it, it will not survive a busy month, and the months you most need to see clearly are the busy ones.

Four: What happens to a lead between the moment they show interest and the moment they buy? If the honest answer involves the word “usually,” that is not a process, it is a habit, and it is doing different things depending on who is on shift.

Five: What did you decide to stop doing last quarter, and what evidence did you use? If nothing was stopped, nothing was measured, and you are accumulating marketing activity rather than choosing it.

Fail three or more and you do not have a brand problem. You have a visibility problem wearing a brand problem’s clothes, and no amount of repositioning will touch it.

I want to be careful about the tone here, because this is the point in an essay like this where a consultant usually implies the founder was careless. She was not careless. She was doing what almost every founder does, which is build the parts of the business that customers can see, because those are the parts that produce immediate feedback. The backend produces nothing you can feel until the day it produces everything.

What we actually did, briefly, before the part that is paid.

We stopped the brand work. We consolidated everything into one system, connected the money to the marketing so that every lead carried its source all the way through to the sale, and built a weekly view she could read in ninety seconds without asking a soul.

Then, and only then, we did the brand strategy. It took less than half the time it would have taken in week one, because by then the answers to the strategy questions were sitting in the data rather than in a room full of opinions.

The positioning we landed on was not the one either of us would have guessed in week two.

The rest is for paid subscribers. Below: the sequence I use to rebuild this, in order, including the step most people do first and should do fourth. And the harder half, which is what happened when I finally ran this same diagnostic on my own business and did not like the answer.

Read more

For most of my twenties I believed that if I could just find the right words, the right order, the right tone, the man in front of me would finally get it. My husband, a client, a man on a panel, my father. I had faith in explanation the way some people have faith in God.

I was very good at it. That was the problem. Being good at explaining is how you end up doing it for a decade.

Somewhere in the last year I stopped. Not out of contempt, and not because I lost the argument. I stopped because I noticed what the explaining was costing me, and because I finally understood that I had been asking for the wrong thing the entire time.

There are five things here. The first four are things I have given up trying to make men understand, with the research that explains why the attempt was always going to fail. The fifth is what I ask for instead, and it is smaller and stranger and far more useful than understanding.

If you are a woman reading this, I think the fourth one will be the one you have never had words for. If you are a man reading this, skip to the fifth. You do not need to understand any of the first four to act on it.

One. We are not overthinking. We are running a risk model nobody asked us to build.

Here is a thing that happens. I bring up a scenario. He says, that’s not going to happen. I say, I know, but if it does. He says, why do you do this to yourself.

What he is seeing is anxiety. What is actually running is a model.

There is a well-documented cognitive strategy called defensive pessimism, studied most closely by the psychologist Julie Norem. People who use it set deliberately low expectations before a high-stakes event and then mentally rehearse everything that could go wrong, in detail, in advance. It looks exactly like spiraling. It is not. Norem’s most interesting finding is what happens when you take it away: when researchers put defensive pessimists in a good mood or blocked them from imagining failure, their performance got worse, not better. The rehearsal was doing something. It was converting anxiety into preparation.

So when a woman maps out four ways a launch could fail before she agrees to the launch date, she is not catastrophising. She is running the only instrument she was ever given.

And she was given it early. Sociologists studying women’s daily lives have a term for the invisible, continuous labour of anticipating harm: safety work. Text me when you get home. Share your location. Keys between your fingers. Do not take that route after nine. Say you have a boyfriend, it works better than saying no. None of us were taught to strategise. We were taught to be careful, relentlessly, from about the age of eleven, and careful at that volume for that long simply becomes a strategy. It transfers. It shows up in a board meeting.

I stopped explaining this because the explanation itself became evidence against me. Justifying your risk model in real time sounds like worrying about worrying.

Two. Ambition is not a wish to be masculine. That translation is happening in your head, not ours.

I have been told I am intense, intimidating, a lot, too direct, surprisingly aggressive for a ‘woman’. I have also, more than once, been told as a compliment that I think like a man.

For a while I tried to argue about the specific words. Then I found the theory that explained why the words were never the point.

In 2002, Alice Eagly and Steven Karau published what is called role congruity theory. The argument is that prejudice toward women leaders emerges from a perceived mismatch between two roles: what people believe a woman is supposed to be, and what people believe a leader is supposed to be. When one person occupies both, the observer experiences an incongruity, and the incongruity gets resolved by deciding something is wrong with her.

Notice where the problem is located. Not in her behaviour. In the perception. This is why the same email is decisive from him and abrasive from her. Nothing in the email changed. A congruity error fired in the reader. I want to say the thing underneath this clearly, because it is the part that gets lost. Wanting to run a company is not wanting to be a man. Wanting to be the most competent person in the room is not a rejection of femininity. We have had four waves of feminism and we are still fielding the suggestion that ambition is a costume, that if you scratch a driven woman you will find someone performing masculinity because she was not offered anything better.

I like being a woman. I like it enormously. I also intend to take up the entire room.

The men who understand this are not the ones who tell you that gender does not matter. They are the ones who stop translating. Who hear a woman say she wants something large and do not silently convert it into a diagnosis.

Three. We are not losing to AI because we adapt worse. We were sorted into the jobs it eats.

This is the part of my work I get paid for, so let me be precise. In 2025 the International Labour Organization published a refined global index of which occupations are exposed to generative AI, built from nearly thirty thousand occupational tasks. The gendered finding is stark. Female-dominated occupations are almost twice as exposed to generative AI as male-dominated ones, twenty-nine percent against sixteen. In high-income countries, the jobs sitting in the highest-risk category account for 9.6 percent of women’s employment and 3.5 percent of men’s.

The ILO’s own explanation for this is not that women are slower to adopt new tools. It is occupational segregation. For a century, clerical, administrative and business support work was one of the main routes women had into the formal economy. That work is routine, well-documented, and highly automatable, which is exactly why it is now the most exposed category in the index.

And there is a second finding that matters more than the first. Women remain underrepresented in AI-related roles, which limits their access to the jobs being created and shapes how these systems get designed and deployed in the first place. So the same group most exposed to the disruption has the least influence over the thing doing the disrupting.

I have stopped explaining this one at dinner parties, because the conversation always resolves into a debate about whether women are as good at technical work, which is not the question and never was. Nobody is arguing about aptitude. The distribution was set decades before the technology arrived. You cannot out-adapt a hundred years of sorting in eighteen months, and asking women to try, individually, one upskilling course at a time, is a way of turning a structural problem into a personal failing.

We are not less analytical. We were pointed somewhere.

Four. Our composure costs something. You are reading the absence of a signal as the absence of a problem.

This is the one I have never been able to say properly, so I am going to use two pieces of research to say it for me.

The first is emotional labour, named by the sociologist Arlie Hochschild in 1983. She was studying flight attendants, and she made a distinction that has stayed with me: surface acting, where you produce the correct expression while feeling something else, and deep acting, where you work on yourself until you actually feel the required thing. Both are labour. Both are performed on shift, all day, mostly by women, and neither appears anywhere in a job description.

The second is allostatic load, a concept from the neuroscientist Bruce McEwen. Your body has systems for adapting to stress, and they work. The cost is that adaptation is not free. Run those systems continuously, for years, without recovery, and the wear accumulates in the body. Not as a feeling. As physiology.

Put those together and you get the thing I could never explain.

When a woman is calm through something difficult, that calm is not evidence that the thing was manageable. It is the output of work. The composure is the labour. And because it is being performed well, it produces no visible signal, so everyone around her concludes she is fine and adjusts nothing.

Then, years later, something in the body gives. And it is treated as an unrelated medical event.

I know a striking number of ambitious women in their late twenties and thirties with autoimmune conditions, gut problems, cycles that stopped, inflammation nobody can source. I am not going to make a clinical claim about causation, because I am not a doctor and this deserves better than a confident line in an essay. But I notice that the women I know who have spent a decade being impressively fine are not, on the whole, well.

I stopped explaining this because the explaining requires you to stop performing the composure in order to describe it, and the moment you do that, the composure is what people miss.

Five. We are not asking to be understood. We are asking to be believed.

Here is what I got wrong for a decade.

I thought the goal was comprehension. If he could feel what it feels like, he would act differently. So I explained, and refined, and found better analogies, and got angry, and explained again, and none of it worked, because comprehension of that kind requires shared experience and he does not have it and will never have it. There is no sentence that transmits twenty years of being told to be careful. There is no analogy that installs a risk model.

The philosopher Miranda Fricker has a name for the thing that was actually going wrong. In her 2007 work she describes testimonial injustice: when a listener gives a speaker less credibility than they deserve because of a prejudice about who is speaking. It is not disagreement. Disagreement happens after the content has been heard and weighed. This is a discount applied before, so that the same information carries less weight coming from her.

That is the actual injury. Not that he does not understand. That he does not credit.

And once I saw the difference, the whole thing became simple, and much lighter.

Understanding is a large ask. It requires him to have lived a life he did not live. Belief is a small one. It requires only that he accept I am a reliable narrator of my own experience. He does not have to know what it is like to run the risk model. He has to accept that when I say a thing is a risk, I am reporting, not fretting. He does not have to feel what the composure costs. He has to accept that it costs something, on my word, without a receipt.

That is it. That is the whole ask. And it turns out most decent men can do it immediately, once you stop demanding the thing they cannot give and start asking for the thing they can.

I want to end somewhere other than where this genre usually ends, because I am tired of the ending where women are right and men are hopeless. It is satisfying for about four minutes and it changes nothing.

We are not different species. I want to be careful here, because that framing is comfortable and it is also the opposite of everything I just argued. If it were biology, none of this would be anyone’s fault and none of it could change. It is not biology. It is training. Different curricula, taught from different ages, with different consequences for failing, and neither of us chose ours.

You cannot transmit a training somebody never underwent. That is not a moral failure on his part. It is a fact about what experience is.

Which means the peace is available, and it is available right now, and it does not require anybody to become fluent in a language they will never speak. It requires one thing on each side. From him, credit without comprehension. From us, and this is the harder one, the decision to stop spending our health on being understood by people who structurally cannot.

I gave that up this year. I have more of myself than I did.

The explaining was never going to work. But I did not stop because I lost. I stopped because I finally figured out what to ask for.

We had decided to take a jet ski out. My husband and I, both in Type III life vests, both reasonably confident in the way you are confident about things you have not done. I want to be clear about something before I go further: I am not a swimmer. Not a weak swimmer, not an out-of-practice swimmer. I am a woman who has spent her whole life staying where her feet touch.

The jet ski flipped.

I remember the water taking us and then giving us back. We surfaced holding hands, and I was floating, and it had taken perhaps three seconds, and those three seconds had the length of an afternoon. I turned my head one way and then the other, trying to locate the shore, trying to locate anything.

And there it was. Isla Verde, a strip of sand with small heads on it, people in the shallows, someone’s speaker carrying Bad Bunny across the water with total clarity. A beach having an ordinary Saturday.

Nobody was looking at us. Or rather, some of them certainly were, and what they saw was two adventurous idiots who had dumped a jet ski on purpose, the way people do, laughing about it in open water.

Inside my body was the loudest event of my life. Outside it, nothing was happening at all.

I have thought about that gap for a long time now, and I have since found the research that explains it. What I want to give you here is the mechanism, because it turns out this is not a story about the sea. It is a story about how catastrophe is signalled, and how reliably it is not, and it is happening around you right now in rooms with no water in them. There is a version of this fact that should frighten you and a version that should relieve you. Both are true. I am going to give you both.

Let me stay in the water a moment longer, because the details matter.

What I remember most is not fear exactly. It is the arithmetic. I remember calculating the distance to the beach and understanding, with a clarity I have never had about anything, that I could not swim it. Not on my best day. Not with an hour and no waves.

I remember being aware that I was making no noise. This is the part I could not explain to myself afterward. I was not being brave. I was not managing my panic for my husband’s sake. I simply did not scream, and it did not occur to me to scream, because some part of me had concluded that screaming was unnecessary. That anyone looking at me could obviously see what was happening.

They could not see it. There was, in the most literal sense, nothing to see. I was upright in the water, held there by a well-designed piece of equipment, my head above the surface, my face doing whatever a face does when its owner has left it.

Which brings me to the object at the center of this essay.

A Type III life vest is what is called a flotation aid. It is built for calm water where help is close, and it does exactly one thing extremely well, which is hold a conscious person upright and afloat. It saved me. It is the reason I am writing this instead of being a paragraph in someone’s travel warning.

It is also the precise reason nobody came.

The vest that kept me alive is the vest that made me look fine. I have not been able to get past that. Everything that was functioning correctly in that situation was also, at the same time, erasing the evidence that anything was wrong.

Here is the research.

In 1998, Thomas Gilovich, Kenneth Savitsky and Victoria Medvec published a paper in the Journal of Personality and Social Psychology naming something they called the illusion of transparency: the tendency to overestimate the degree to which other people can discern your internal states. Their explanation is that when you are inside a heightened emotional experience, that experience is so overwhelming that you cannot adjust away from it when you try to imagine another person’s view. You anchor on your own phenomenology. You assume the spotlight can see through you.

In one of their studies, people who had been instructed to lie estimated that nearly half the observers would catch them. About a quarter actually did.

And the paper did not present this as a curiosity about liars. From the beginning, the authors framed it around a specific implication: what this bias does to people’s willingness to intervene in emergencies.

I read that and felt something reorganize. I had spent a long time carrying a small private shame about that afternoon, which was that I had not called for help. That I had frozen. That some failure of nerve on my part was why nobody came.

That is not what happened. What happened is that I was a normal human being experiencing a documented bias, in a life vest doing its job, in front of a beach that had no way of knowing.

Now let me widen this, because the water is not the point.

The illusion of transparency does not need an ocean. It needs only a person having an internal experience they assume is visible. Which is to say it needs only a person.

Nobody can tell. The company looks like it is working, in the same way a person in a life vest looks like they are swimming.

Think about the friend whose messages got shorter and warmer at the same time. The mother in the pickup line who is entirely composed. The colleague whose work has not slipped at all, which is exactly the problem, because holding your output steady while everything underneath collapses is the most expensive thing a person can do and it produces no external signal whatsoever.

We have all been on both sides of this, and the reason it keeps happening is that both sides are broken simultaneously.

Because there is a second mechanism, and it belongs to the beach. It is called pluralistic ignorance, and it was named in this literature by Dale Miller and Cathy McFarland in 1987. In an ambiguous situation, each person looks around to see how others are reacting, and calibrates their own reaction accordingly. If everyone is calm, everyone concludes there is nothing to be alarmed about. The calm is self-reinforcing. It has nothing to do with the actual danger and everything to do with the fact that the only available evidence is other people’s faces.

So consider the full architecture of what happened to me. From inside the water, I could not signal, because I was certain I already was. From the beach, nobody could read it, because a hundred relaxed people were reading each other. Two separate failures, moving in opposite directions, meeting in the middle of the Atlantic where I was floating.

This is not an unlucky configuration. This is the default. This is what happens by design, every day, in every office and every marriage and every friendship, and the fact that most of the time nobody actually drowns is not because the system works. It is because most distress is survivable.

So here is what I would ask you to take from this.

Stop waiting to be read. Your interior is not visible. Not to your husband, not to your closest friend, not to the person who has known you for twenty years and can finish your sentences. They are not being negligent. They are looking at a face, and a face is a very poor instrument.

And stop reading calm as information. The calmest person in your life right now is not necessarily fine. They are, at minimum, someone whose internal state you have no access to, and the confidence you feel about how they are doing is manufactured entirely inside your own head.

If there is one thing to do tonight, it is this. Pick the person you are not worried about. Ask them a question specific enough that a general answer will not fit.

Read more

There is a particular silence in the boutique after you say yes.

The SA is folding tissue paper around something you have wanted for four years, and you are standing very still, because you understand that you have done something larger than a purchase. You have passed a law.

The law reads roughly: I am now a woman who owns this. Which means I am now a woman who wears it. Who has places to wear it. Whose weeks have the shape this object implies, whose calendar contains the dinners and the flights and the rooms where a quilted flap makes sense. The bag is not the point. The bag is the enforcement mechanism for a version of yourself you have not been yet.

Nobody spends eight thousand dollars on storage.

I want to talk about what happened three days ago in Brussels, because it is the same gesture, at the scale of a continent, and it failed for the same reason.

August 2, 2026 was supposed to be the day. For two years it sat circled in every compliance calendar in the Western world: the date the EU AI Act’s obligations for high-risk systems became enforceable, the moment the world’s first comprehensive AI law grew teeth. Recruitment algorithms, credit scoring, diagnostic support, anything touching Annex III. Conformity assessments, technical documentation, human oversight, post-market monitoring. Real fines attached, calculated as a share of global turnover.

It did not happen. On June 29, the Council of the EU gave final approval to the Digital Omnibus on AI, the first substantive amendment to the Act since it passed. The high-risk obligations that organized two years of enterprise planning moved to December 2, 2027. The versions embedded in regulated products moved to August 2028.

The reason matters more than the delay. Europe did not blink because industry lobbied hard enough, though industry lobbied. It blinked because the machinery underneath the law was not built. Member states had not finished designating the national authorities meant to supervise compliance. The harmonised standards that tell a company what compliance actually looks like, in technical terms, were unfinished. The conformity assessment infrastructure was thin. The law described a world in exacting detail and then arrived at the door of that world to find no instruments for observing it.

Europe wrote the rule before it built the ability to see whether the rule was being followed.

This is the Chanel bag.

Fashion has exactly one honest metric, and it is cost-per-wear. Price divided by the number of times the thing has actually been on your body. It is not a clever metric. It is arithmetic. It requires nothing but a number you already know and a number you have never once tracked.

Try it right now, from memory. Not the bag you regret. You know that number, roughly, and the knowledge sits in you like a small stone. Try it on the thing you reach for constantly, the one you would replace immediately if it were stolen. You do not know. Nobody knows. We buy in the language of investment and account in the language of feeling.

So here is the sequence, and I want to be precise about it, because the sequence is the whole argument.

You did not buy the bag because you observed your life and concluded that your life required it. You bought the bag in order to summon the life. The purchase was legislative. It declared a standard, and it created an enforcement mechanism, which is guilt, which activates every time the bag stays in its dust sleeve for a fourth consecutive week. What the purchase did not create, at any point, was a measurement layer. You have a law and a punishment and no instruments.

The result is not that you stop caring. It is worse. The result is that you begin optimizing for the part of the system you can see.

Goodhart’s Law, in its cleaner formulation, holds that when a measure becomes a target, it ceases to be a good measure. In fashion this arrives dressed as the investment piece. The moment “investment piece” becomes the standard by which a wardrobe is judged, we begin optimizing not for the outcome the phrase was built to capture, which is wearing one excellent object into the ground across twenty years, but for the legibility of the investment. The flat lay. The visible logo hardware in a photograph taken on a chair, indoors, in a home the bag has not left. The measure was a proxy for a practice, and the proxy is much easier to satisfy than the practice, so the proxy wins. Quietly, and without any moment where you decide to cheat.

James C. Scott, writing about why ambitious state projects fail, gave us the word for what is happening here. He called it legibility: the way institutions can only govern what they have first rendered visible, countable, standardized. His argument was that the process of making a thing legible destroys something in it, the accumulated practical knowledge he called mētis, which lives in the illegible parts and does most of the actual work.

Your closet has a legible third and an illegible two-thirds. The front third is what you believe your wardrobe to be, and it is the part you legislate. The back two-thirds is where your actual life is stored, and it is governing you continuously, without your consent, because you never look. Every purchase you make is made against a mental model of your existence that the real data would immediately contradict. Not cruelly. Just factually.

Now the part where the metaphor stops being a metaphor.

Same continent. Same year. Different room. While Brussels was conceding that it could not yet enforce its AI law, it was preparing to require the fashion industry to do the precise thing it had just admitted was not yet possible in AI.

Under the Ecodesign for Sustainable Products Regulation, Article 9, every textile product sold into the EU will eventually need to carry a Digital Product Passport: a structured, queryable record, reachable by QR code or RFID, carrying verifiable information on fibre composition, origin, environmental impact, and end-of-life handling. Not a marketing claim. A technical record that a customs officer can pull up. Textiles sit on the first-priority list. The delegated act defining what the passport must contain is expected in 2027, with a transition period pushing real implementation toward 2028.

And the industry cannot currently produce this data. Not because it is unwilling. Because supply chain information is fragmented across tiers, material traceability is inconsistent between suppliers, internal systems do not speak to each other, and supplier documentation is incomplete, across chains that routinely cross five or more countries and dozens of intermediaries. A brand can tell you its values with total fluency and cannot tell you, at the level of a single garment, where the fibre was grown.

So Brussels is running the same experiment twice, in two industries, and in both the binding constraint is identical. It is never the ambition. The ambition is abundant and well-drafted and morally serious. It is the plumbing. It is whether anyone built the thing that lets you see.

I have been calling this the observability deficit, and I would like to stop calling it an AI problem, because it is not one. It is a structural feature of how ambition behaves when it meets reality, and it shows up wherever someone declares a standard faster than they build the capacity to observe it.

Which is to say: it shows up in your business.

Declaring a niche is passing a law.

You do it in a single afternoon, usually after reading something. From now on, only mission-driven founders. Only Series A. Only clients who value the work. The declaration feels like clarity, and clarity feels like progress, so the declaration gets treated as if the strategic work is now complete.

But ask the questions the law implies and the instruments are missing. Which of your last twelve clients actually converted from cold outreach versus referral? Which engagement type produced revenue per hour that you would accept again? Which work did you not dread on a Tuesday? Where did every single inbound lead in the past six months originate, specifically, not approximately? Almost nobody can answer these, including people who have been in business for years, including people who are good at what they do.

So the niche gets enforced by feeling. You turn down work that violates the positioning even when the data, had you collected any, would have told you that this exact category of work is where your referrals come from. You keep offers that make sense in the story and lose money in the ledger. You are enforcing a law against a business you have never measured.

And in your own life it is the same instrument, turned inward. The five a.m. wake-up is legislation. The content cadence is legislation. So is the reading list, the gym block, the rule about not checking your phone before nine. Each one is a standard declared in advance of any observation, enforced by shame, and optimized, inevitably, toward the visible proxy. The streak survives; the outcome it stood for does not. You know what your morning is supposed to look like. You have no idea, in the way you would need to know it to actually change anything, where your last two weeks went.

You have courts. You never ran a census.

The correction is unglamorous, which is why it does not spread. Observation precedes enforcement. Always, and by a wide margin. You cannot govern what you have not first agreed to look at, and the looking has to happen before the rule, not after the failure.

In practice, for the next fourteen days:

• Track where the hours actually go, in fifteen-minute blocks, without editing and without improving your behavior while you measure. The point is not discipline. The point is a baseline, and a baseline is only useful if it is embarrassing.

• Build a ledger of every client or project from the last year with three columns: revenue, hours, and whether you would take it again. Your niche is in that third column. It has been sitting there the whole time.

• Run cost-per-wear on ten things you own. Not to feel bad. To find out what your actual life requires, as opposed to what your purchases have been arguing your life should be.

None of this is a rule. That is the point. Rules come later, and they come out of the data, and when they finally arrive they are enforceable because they describe something you have actually seen.

Europe now has until December 2027 to build the measurement layer it should have built first. It got something rare, which is a second chance with the deadline moved rather than the ambition abandoned. Most of us do not get that. We get the quieter version, where the law stays on the books, unenforceable, and we conclude that the failure was in our discipline.

It was not your discipline.

The bag was never the problem either. The bag was an answer to a question you had not asked yet. Go ask it.

If you work in AI governance, sustainability reporting, or you are a founder currently enforcing a positioning you have never measured, I would like to know whether this matches what you are seeing

.

In November 2025, a team of researchers at the National University of Sciences and Technology in Islamabad published a paper that should have rewritten the conversation about sustainable AI governance. It didn’t. Almost nobody outside of a narrow ML sustainability corridor read it. But the paper, which introduced G-TRACE, a cross-modal, region-aware carbon accounting framework contained a finding that reveals something fundamental about the gap between AI governance as it is theorized and AI governance as it is practiced.

The finding was this: the Ghibli-style image generation trend that swept TikTok, Instagram, and Pinterest between 2024 and 2025 consumed an estimated 4,309 megawatt-hours of energy and produced approximately 2,068 metric tonnes of CO₂ emissions. That is roughly equivalent to the annual carbon output of 450 passenger vehicles. A single aesthetic trend, not a model training run, not a new architecture, not a deliberate deployment; produced tonne-scale climate consequences through inference alone.

What makes this significant for governance is not the number itself. It is that no organization involved in the lifecycle of that trend; not the model providers, not the platform operators, not the regulatory bodies tracking AI’s environmental footprint had the operational infrastructure to observe it happening in real time. The emissions were reconstructed after the fact, through proxy indicators: hashtag counts, platform analytics interfaces, estimated FLOPs per image generation call, and regional grid carbon intensity data.

This is what I call the observability deficit. And it is, I want to argue, the central structural problem in sustainable AI governance today — more consequential than any gap in policy language, regulatory ambition, or technical standard.

The inference pivot and why it changes everything

The field’s understanding of AI’s environmental footprint has undergone a quiet but consequential shift. For years, the dominant framing was about training: the enormous compute required to produce a frontier model, measured in GPU-hours, petaFLOP-days, and the now-canonical comparisons (training GPT-3 emitted as much CO₂ as five transatlantic flights, and so on). Training was legible. It happened in known facilities, on known hardware, over known timeframes. It could be estimated, attributed, and debated.

What changed, sometime around mid-2025, is that inference overtook training as the dominant source of frontier-model lifecycle energy consumption. The Stanford AI Index 2026 confirmed what ML.Energy benchmarks had been suggesting: inference now accounts for roughly 63% of total lifecycle energy across frontier models. The crossover happened as deployment volume scaled faster than new model training cycles.

This matters enormously for governance because inference is operationally diffuse in a way that training is not. Training happens in identifiable clusters; a known model, a known data center, a known period. Inference happens everywhere, continuously, in response to tens of millions of individual user actions, across geographically distributed compute infrastructure with heterogeneous carbon intensity profiles. A single text query to GPT-4 consumes roughly 0.95 watt-hours. A single Ghibli-style image generation call consumes significantly more. Multiply by the 25.8 million Ghibli images the G-TRACE team estimated were generated across platforms, and individual actions aggregate into system-level climate impact — with no organizational entity tracking the accumulation.

The governance implication is stark: the majority of AI’s environmental footprint now lives in a domain that existing governance architectures were not designed to observe.

The framework-infrastructure gap

There is no shortage of governance frameworks addressing AI sustainability. The OECD and World Economic Forum published AI in Strategic Foresight: Reshaping Anticipatory Governance in late 2025, drawing on a survey of 167 practitioners. The EU AI Act, which began phasing in substantive obligations in February 2025, includes provisions for environmental impact reporting. ISO published TR 20226:2025, which explicitly incorporates sustainability indicators across the full AI system lifecycle; computational load, resource use, carbon impacts, disposal, and operational characteristics. The Green Software Foundation has developed carbon-aware software development standards. UNESCO’s AI ethics framework calls for environmental assessments as a core component of responsible AI deployment.

These frameworks share a common structural assumption: that the organizations implementing them possess the operational infrastructure to generate the data the frameworks require. Specifically, they assume the existence of what enterprise ML engineering calls an MLOps governance layer, automated pipelines capable of model versioning, lineage tracking, inference monitoring, compute attribution, and environmental telemetry integrated into deployment gates and CI/CD workflows.

A March 2026 paper on ethical AI frameworks formalized this expectation in what the authors call a triple-gate structure: metric gates (quantitative performance and safety thresholds), governance gates (legal and procedural compliance), and eco gates (carbon and water budgets as first-class deployment criteria). The paper explicitly argues that “ethics must move from an after-action audit into a gatekeeping mechanism embedded in build, deploy, and serving paths.” This is correct. But it presupposes an MLOps maturity level that the vast majority of organizations deploying AI, and certainly the vast majority of organizations governing AI — do not possess.

Consider the organizational profile of a typical AI safety or climate-focused nonprofit: 8 to 30 staff, grant-funded, running on Google Workspace and Airtable, deploying commercially available LLM APIs for internal research and communications, with no dedicated ML engineering function, no model registry, no inference logging, no compute attribution system, and no environmental telemetry of any kind. These organizations are often the very entities producing the governance frameworks, convening the policy discussions, and advising governments on responsible AI deployment.

The gap between what the frameworks demand and what the organizations can operationally deliver is not a minor implementation detail. It is a structural disconnect that undermines the credibility and enforceability of the entire sustainable AI governance project.

The Ghibli case and the problem of cultural accelerants

The G-TRACE research team’s analysis of the Ghibli trend illustrates a category of AI climate risk that existing governance models are particularly ill-equipped to address: demand spikes driven by cultural accelerants.

The paper defines cultural accelerants as “fandoms, remix culture, and meme dynamics” that “create unanticipated energy demand spikes for GenAI systems.” The Ghibli case is instructive because the emissions were not generated by a corporate deployment decision, a model training run, or an infrastructure expansion. They were generated by distributed individual behavior, where millions of users independently choose to generate images in a particular aesthetic style, aggregating into a compute demand pattern that no organizational actor planned, authorized, or monitored.

The research team collected data across six platforms: Instagram, Twitter/X, TikTok, Pinterest, Facebook, and Tumblr, using publicly available analytics interfaces and hashtag volume as proxy signals. The methodological constraint is revealing: because no platform or model provider publishes real-time per-query energy data, the researchers were forced to reconstruct energy consumption from indirect indicators. They used baseline energy estimates per image generation call, scaled by estimated generation volumes, weighted by regional grid carbon intensity using deployment geography data.

This is sophisticated work. But it is also, fundamentally, archaeological. It reconstructs what already happened. It cannot prevent, redirect, or even observe in real time. And this is the core problem: sustainable AI governance as currently constituted is a retrospective practice operating in a domain that requires prospective observability.

The AI Sustainability Pyramid that the G-TRACE team proposes: a seven-level model linking carbon accounting metrics (L1 through L7) with operational readiness, optimization, and stewardship, correctly identifies that governance effectiveness depends on measurement maturity. But at L1 (basic energy accounting) and L2 (carbon intensity attribution), most organizations; including the most prominent AI governance bodies have not yet begun.

What an operational telemetry layer actually requires

If the observability deficit is structural, the question becomes: what would it take to close it?

The honest answer is that closing it fully at the ecosystem level requires coordination mechanisms that do not yet exist; shared inference reporting standards, carbon-aware API pricing, and mandatory compute attribution in model serving infrastructure. These are policy problems that exceed the capacity of any single organization.

But at the organizational level, the level at which AI governance is actually practiced — the requirements are more tractable than the field’s current discourse suggests. They fall into four categories:

Model and tool inventory. Before an organization can govern its AI-related environmental footprint, it must know what AI tools it is using, where, and how frequently. This is not a sophisticated technical requirement. It is an operational hygiene requirement, the equivalent of an asset register that the majority of mission-driven organizations deploying AI do not maintain. An internal audit of AI tool adoption (which commercial APIs are in use, which teams use them, for what task categories, at what volume) is the precondition for any subsequent measurement.

Inference-side usage telemetry. For organizations consuming AI through commercial APIs (which is most of them), inference usage is partially observable through billing data, API call counts, token volumes, model identifiers. This data is already generated; it is simply not aggregated, analyzed, or connected to environmental metrics. Connecting billing telemetry to per-model energy estimates (such as those compiled by ML.Energy) produces a rough but meaningful carbon footprint estimate without requiring any custom infrastructure.

Data lineage and model provenance tracking. For organizations that fine-tune models, maintain internal datasets, or deploy AI in programmatic workflows, data lineage; the record of which data was used to produce which output, through which model, at what time is the substrate that enables both auditability and environmental attribution. Without lineage tracking, an organization cannot answer basic governance questions: which model generated this output? What data informed it? How many compute cycles did it require?

Governance integration into operational workflows. This is the layer where frameworks become enforceable. Policy-as-code: the practice of encoding governance rules as automated checks within operational pipelines, transforms governance from a documentation exercise into an operational constraint. An organization that embeds a carbon budget threshold into its deployment pipeline (flagging or blocking actions that exceed a per-query energy ceiling) is practicing sustainable AI governance. An organization that publishes a sustainability framework but has no mechanism to monitor its own AI usage is not.

The organizational capacity problem

Stating what the telemetry layer requires is simpler than explaining why it does not exist. The gap is not primarily technical. It is organizational.

Mission-driven organizations working on AI governance and climate policy allocate their cognitive and financial resources correctly to their core mission: research, policy advocacy, convenings, publications, and public engagement. Operational infrastructure is systematically deprioritized because it does not produce visible outputs that funders, boards, or peer organizations recognize as impact.

This creates a paradox: the organizations with the greatest intellectual authority on sustainable AI governance are often the least operationally equipped to practice it. Their internal systems, the CRMs, project management tools, data stores, and workflow automations that constitute their organizational infrastructure are typically fragmented, underdocumented, and maintained by generalist staff rather than systems specialists.

The 2024–25 Australian Institute of Company Directors study on not-for-profit governance found that while 74% of NFP directors discuss ESG issues annually, more than 55% of boards had taken no steps to address climate-related risks in their own operations. This pattern is not unique to Australia. It describes a structural feature of mission-driven organizations globally: governance attention directed outward (at the systems they seek to influence) rather than inward (at the systems they operate).

For AI governance specifically, this means that the organizations publishing frameworks on responsible AI deployment, compute transparency, and environmental impact reporting often lack the internal infrastructure to apply those frameworks to their own operations. They cannot audit their own AI tool usage because they have no tool inventory. They cannot report their inference-related emissions because they have no usage telemetry. They cannot enforce governance gates because their workflows have no gates.

Toward operational readiness as a governance criterion

The field needs a conceptual reframe. Sustainable AI governance is currently understood as a policy problem: what rules should govern AI’s environmental impact? This is necessary but insufficient. It must also be understood as an operational capacity problem: do the organizations responsible for governance possess the systems infrastructure to implement, monitor, and enforce those rules?

This reframe has practical implications:

Funders: foundations, government agencies, international organizations — should evaluate grantees’ operational infrastructure alongside their programmatic output. An AI governance organization that cannot demonstrate basic internal AI usage tracking is less credible as a governance authority than one that can.

Standards bodies: ISO, NIST, the OECD should incorporate organizational readiness assessments into their governance frameworks. A standard that specifies what must be measured but not whether the implementing organization can measure it is incomplete by design.

Organizations themselves, particularly AI safety, AI governance, and climate-AI nonprofits — should treat operational systems architecture as a core competency, not an overhead cost. The same rigor applied to model evaluation, policy analysis, and stakeholder mapping should be applied to internal data models, tool governance, and workflow automation.

The Foresight Institute’s Existential Hope program frames the challenge of steering toward desirable futures as requiring not just vision but institutional readiness; the organizational capacity to translate foresight into action at the pace required by technological change. This framing applies directly to sustainable AI governance. The vision exists. The frameworks exist. The institutional readiness to operationalize them does not, and until it does, the governance project remains aspirational rather than functional.

Coda

The 2,068 tonnes of CO₂ produced by the Ghibli image generation trend did not result from any governance failure in the traditional sense. No regulation was violated. No ethical principle was breached. No organization made a decision to emit. The emissions emerged from the aggregate behavior of millions of individuals interacting with commercially available tools, behavior that no governance architecture was designed to observe, much less constrain.

This is the condition under which sustainable AI governance must now operate: a domain where impact is diffuse, inference-driven, culturally accelerated, and operationally invisible to the organizations tasked with governing it. Meeting that condition requires not better principles, but better plumbing.

The missing layer is not ethical. It is infrastructural.

Maryam Khan is the founder of Equis Advisory (forthcoming), where she works as a strategy & system-ops partner for organizations in the ethical AI and sustainability space. She writes and speaks about the operational infrastructure of responsible AI governance, and how to make its usage more sustainable.

Most people doing “impact work” aren’t failing because they don’t care.

They’re failing because the systems behind the work can’t carry the weight of the mission.

I’ve spent the last few years working in operations acr…

Read more