The Originality Tax.
Exploring how every AI tool is, by construction, a median machine — and why producing genuinely original work in 2026 now costs cognitive effort that simply did not exist before.
In 2024, two researchers — Anil Doshi and Oliver Hauser — published the most quietly unsettling study of the AI-and-creativity literature so far. They asked one group of writers to produce short stories with no help. They asked another group to use a generative AI to brainstorm. Then they had thousands of independent readers rate the results.
The headline finding was the optimistic one. AI helped. Stories produced with AI assistance were rated as more enjoyable, better written, and more creative — especially when the writer wasn’t a particularly creative person to begin with. Generative AI, the paper concluded, “enhances individual creativity.”
The footnote finding was the dangerous one. The AI-assisted stories, when compared to one another, were significantly more similar than the unassisted stories. Better individual outputs, narrower collective idea space. A social dilemma, the authors called it: “writers are individually better off, but collectively a narrower scope of novel content is produced.”
That single graph — individual creativity up, collective diversity down — explains something most of us have felt in the last eighteen months but couldn’t quite name. The pitch decks all say the same things in the same voice. The strategy memos arrive at the same recommendations. The think-pieces converge on the same takes. The LinkedIn posts have started to share a faint, unmistakable family resemblance.
We have more content than ever, and it sounds more alike than ever.
This is a piece about the invisible cost of producing that: the cognitive tax that every professional now pays, whether they notice it or not, whenever they sit down to do work that is supposed to sound like them.
I’m going to call it the originality tax. It is the new, real, unmeasured cost of producing work that doesn’t sound like everything else.
What “median machine” means, structurally
Let’s start with a piece of mathematics that almost nobody discusses honestly.
A large language model is, in its mechanics, a probability distribution over what comes next given what came before. It was trained on enormous quantities of human-written text, and it learned to produce the most likely continuation. Not the most original. Not the most surprising. The most likely. That is the loss function. That is the thing the model was optimised to do.
The implications are easy to miss because they sound innocuous. The model is “helpful”. It “smooths your prose”. It “polishes your work”. But what it is actually doing — at the level of the cost function it was trained against — is moving your text toward a predicted average of how text like this is usually written. That’s the only thing it can do. It cannot give you a continuation more original than its training distribution permits, because its training distribution defines what original even means to it.
The first three suggestions any AI tool gives you are, almost by definition, the three most predictable continuations. Not the most useful. Not the most precise. The most predictable. The probability mass concentrates near the centre of the distribution, and that is what arrives in your editor as a polite suggestion.
Once you see this, you cannot unsee it. Every “make it more professional” pass is a step toward the median professional voice. Every “shorter and clearer” suggestion is a step toward the median form of brevity. Every “tighten the introduction” recommendation is a step toward the median introduction. Each step is small. Each step is an obvious improvement, when judged locally. None of them feels like a loss.
You only notice the loss collectively, at the level of the field — in the Doshi & Hauser graph, in the indistinguishable consulting reports, in the strange feeling of reading three professionals in your industry and being unable to tell them apart.
The mechanism is not malicious. It is built into the architecture. And there is no version of using AI for writing that escapes it.
The autocomplete tilt: where it shows up first
If the median pull were only an aesthetic problem, we could shrug and move on. Everyone uses the same Microsoft Word. Most professional writing was already mediocre.
The deeper problem, surfaced by a series of papers between 2023 and 2025, is that the median pull does not stop at prose. It reaches into thought.
The clearest evidence comes from a 2023 Cornell study by Maurice Jakesch and colleagues, titled — with characteristic academic dryness — “Co-writing with opinionated language models affects users’ views”. The setup was simple. Participants were asked to write short essays about whether social media is good or bad for society. Some used a writing assistant tuned to gently suggest pro-social-media completions. Others used one tuned to suggest anti-social-media completions. The participants were not told.
The result was unsettling. After a single short writing session, participants exposed to the pro-leaning model expressed measurably more positive opinions about social media — and the participants exposed to the anti-leaning model expressed measurably more negative opinions. They reported these views as their own. Asked whether the assistant had influenced them, they said no.
“AI suggestions don’t just speed up writing. They shape what people think. Because the user is the one who chooses to accept the suggestion, the brain internalises it as an original thought.” — Jakesch et al., extended interpretation in PsyPost, 2024
The mechanism here is the bit that should keep professional services partners awake at night. When AI offers you a sentence and you accept it, you are not just borrowing a phrase. You are absorbing a frame. You think you wrote it because you typed it, but the typing was the surface action — the deeper cognitive act, the framing of the thought, came from the model. The user’s sense of authorship survives. The actual authorship moves elsewhere.
Multiply that by every professional who uses AI to draft anything — emails, briefs, decks, op-eds, strategy memos, performance reviews — and you start to understand why the field is flattening. It isn’t that everyone is being lazy. It’s that everyone is being subtly shaped by the same shaping force, and they cannot feel the shaping because the shaping happens at the level of what they take to be their own thoughts.
A 2025 CHI paper from Agarwal and colleagues found something even more pointed: Indian users of AI writing assistants progressively adopted Western prose styles, even when writing in their own language about their own culture. The tool has a default voice, drawn from its training distribution, which is overwhelmingly English-language and Western-coded. The default voice wins. Not just what is written — how it is written.
This is the originality tax, levied early and quietly. By the time the polished draft lands, the cost has already been paid in invisible currency — in the texture of how the writer thinks.
Three taxes, named
The originality tax shows up in three distinct forms. They are paid in different currencies, by different people, at different stages of the work.
The autocomplete tax. The cost paid when a tool finishes your sentence and you accept the finish. You save time. You lose the thinking that would have happened in the moment of completing the sentence yourself. Over thousands of small acceptances, the texture of your prose drifts toward the model’s prose. Most professionals pay this tax constantly and unconsciously. They are not aware it is being levied.
The polish tax. The cost paid when a finished draft is run through a “make this better” pass. The output is more competent — and less distinctive. Specific word choices that signalled a personality get replaced with generic high-frequency synonyms. Idiosyncratic structures get smoothed into standard ones. The piece is more publishable, and less recognisable as the writer’s. This tax is paid deliberately, but most writers don’t notice they’re paying it because the result looks “more professional.”
The brainstorm tax. The cost paid when a writer asks AI to suggest angles, arguments, or framings. The angles offered are, by construction, the angles closest to what other people have already written about this topic. The writer feels generative — three options! — when in fact they are picking from three samples of the median. The writing that follows is competent and forgettable. This is the most expensive tax, because it is paid at the structural level of what the work even is.
You can ship a piece without paying the autocomplete or polish tax. It takes effort. You can write the awkward sentence the autocomplete keeps trying to smooth out. You can refuse the polish pass on the parts that are intentionally rough. Both are inefficient, both are slower, and both produce output that is more recognisable.
The brainstorm tax is harder to evade. To brainstorm without the model is to face the blank page — slow, frustrating, often unproductive in a single sitting. The model offers you something for the discomfort. Most writers, under deadline, accept.
What an over-taxed market looks like
If you wanted a single image of what an over-taxed market looks like, the recent series of consulting-firm AI debacles is the cleanest one available.
In October 2025, the Australian government revealed that a A$440,000 Deloitte report on welfare compliance contained AI-generated fabricated citations, hallucinated quotes attributed to a real federal-court judge, and several invented academic references. Deloitte refunded part of the fee. The story went global. A few weeks later, a Canadian provincial government discovered the same pattern in a near-million-dollar Deloitte healthcare report. Both clients had paid premium rates for premium expertise. They had received, in part, AI-tinted output that nobody had bothered to verify.
The temptation is to treat these as scandals about laziness. They are not. They are scandals about taxation. Inside Big-Four firms, the originality tax has been quietly accepted as a cost-of-doing-business. Internal accounts published in 2025 — including a leaked McKinsey post-mortem on its “Lilli” tool — describe consultants discarding up to half of AI-generated slides because the suggested frameworks were too generic; up to 25% of AI-drafted deliverables requiring substantial rewriting before they reached partner review; senior partners spending two to three additional days per engagement on quality control they didn’t used to do.
These are not “AI productivity gains.” They are the visible, measurable surface of the tax. The output is faster. The verification cost is enormous. The originality of the final product is, by most accounts, lower than what the same firms used to produce ten years ago.
And yet most professional services firms still report AI as a productivity win. Why? Because the tax is paid in a currency the income statement doesn’t track: distinctiveness. Voice. Edge. The things that, in the long run, make a firm worth hiring instead of any other firm. None of those show up in quarterly numbers. The savings do.
What gets lost when distinctiveness goes
There is a temptation here to romanticise pre-AI prose as if it were always good. It wasn’t.
Most writing has always been mediocre. Most consulting reports have always read like other consulting reports. The originality tax did not create the average; it deepened it.
What’s at risk is the stratum above the average. The professionals and brands whose work was, for whatever reason, recognisable. The firm whose memos felt different from the four other firms. The analyst whose voice you’d recognise blind. The designer whose work could be picked out of a deck without the byline. These were always rare. They were also, in a real sense, what made the market a market — the differentiating stratum that gave clients meaningful choices.
When the tax pulls everyone toward the median, the differentiating stratum thins. Not because those people stop existing — but because their work, run through the same tools as everyone else’s, comes out sounding more like everyone else’s. The signal weakens. Clients can no longer tell the firms apart.
The market response to this, eventually, is to start paying a premium for the people who somehow still sound like themselves. We are at the early edge of that. The boutiques that beat bigger firms in 2026 increasingly do so on voice. The newsletters that grow against the big institutional outlets win on voice. The individual contributors who get hired against teams win on voice. None of those are accidents. They are early arbitrage on the originality premium.
Practical implications
For early-career people: the most valuable thing you can do this year is pay the tax deliberately. Write your first drafts without the AI. Notice your actual voice — what words you keep choosing, what rhythm your sentences have, what kinds of metaphors come naturally. Then use the tool to sharpen specific bits, never to speak for you. The delta between your unassisted writing and your AI-assisted writing is the most precise diagnostic you have of how much of your voice is still yours. Watch it.
For hiring managers: stop treating speed of output as a signal. AI broke the relationship between speed and skill. The signal that matters now is the delta between someone’s AI-assisted output and their unassisted output. If they’re identical, you’re hiring the model. The juniors who will become your differentiated seniors in 2032 are the ones who can articulate, when asked, what they refused to let AI do for them.
For leaders: every workflow that prioritises throughput over distinctiveness is taxing your firm’s originality whether you measure it or not. Your competitors using the same tools are converging on the same shape. Clients are starting to notice — slowly, then suddenly. The strategic question is which budgets you’re willing to spend protecting the people who still produce work that sounds like nothing else. They are your differentiation. They are also, currently, more expensive to retain than to lose.
The uncomfortable truth
Originality used to be free. Everyone faced a blank page. The blank page was democratic — it pulled nothing out of you, suggested nothing, finished no sentence. Whatever appeared, however clumsy, was yours.
The AI-assisted page is not blank. It is suggesting, finishing, polishing, every time. Producing something that isn’t the median requires effort the blank page never demanded. The tax is real. It compounds. It is paid invisibly, in cognitive currency, by every professional who uses these tools without explicitly resisting them.
The most uncomfortable part is not that the tax exists. It’s that most people paying it don’t know they’re paying it. They think the smoothed-out, AI-tinted version was their own voice all along. The mechanism by which AI shapes thought — the autocomplete that nudges, the polish that flattens, the brainstorm that median-tilts — operates beneath the level of conscious noticing. By the time you can feel it, your reference point has already moved.
The professionals and brands willing — and resourced — to pay the tax will become rare and valuable. The ones who can’t, won’t. They will sound, increasingly, like everyone else. Their work will be perfectly competent. It will also be functionally interchangeable.
In ten years, when the field has flattened further, the question worth asking will not be “did you use AI.” Everyone will have. The question will be: did your work still sound like you, after?
Most people will not be able to answer.
The few who can will own a market the rest will be too tired to compete in.