Different minds in the age of average

Generative AI is a regression-to-the-mean engine. It is trained on the distribution of what humans have produced, it predicts the most likely continuation, and at scale it pushes whatever it helps make — writing, design, code, plans, decisions — toward the middle of that distribution. The more AI mediates production, the more the median output becomes the default output. That is not a bug. It is the math.

I watch this from two directions at once. As someone who builds software, I watch AI flatten the distance between competent work and median work, and I wonder where anything distinct will come from. As a neurodiversity-affirming parent of an autistic child, I watch a mind that diverges from the modal — pattern-recognition that finds connections the median mind misses, attention that moves nonlinearly, a processing profile that produces genuinely different outputs from the same inputs — get told, over and over, that different is deficient. This piece is where those two views meet. Not a business argument. A description.

What “the age of average” actually means

The mechanism is simple enough to state plainly. A language model is a probability distribution over text. Given a context, it produces the most likely next word, and then the next. Sentence by sentence, this produces fluency. But multiply it across millions of documents, designs, codebases, and plans produced with AI assistance, and you get convergence: the outputs cluster around the center of the training distribution. The tail thins. The distinct, the contrarian, the structurally unusual gets rarer.

Hear the word prediction carefully, because this one is not speculative. The homogenizing effect has already been measured, in two distinct places. Doshi and Hauser (2024) had writers compose short stories with and without AI assistance and found the effect splitting in an uncomfortable direction: individually, the assisted stories scored more creative; collectively, they grew more similar to one another. Individual gain, collective convergence, in the same dataset. Shumailov et al. (2024) describe the slower and deeper version — when generative models are trained on output that generative models produced, the tails of the distribution erode, generation by generation, and the model’s range narrows. Collapse is their word for it.

Those are different mechanisms. One is about people using models; the other is about models eating their own output. Both point the same way. A system that predicts the most likely continuation, and is then fed its own continuations, has no mechanism for recovering what it has stopped producing.

This is not a claim that AI output is bad. It is competent, readable, and functional. The trouble is that it is the same competent, readable, functional output, with variations that are decorative rather than structural. Ask ten teams before and after for a plan and the “after” versions disagree on details but share a shape — the same framing, the same hedged conclusions, the same move-toward-the-middle recommendations. The differences between them are noise, not signal.

Adoption compounds it. The more organizations lean on the same class of tool, the more the modal output becomes the default not just inside one organization but across the market. And the thing that goes scarcest fastest is the genuinely off-center judgment — the contrarian read that turns out to be right, the pattern spotted from an angle the median doesn’t occupy, the decision made from a way of processing that the consensus can’t reproduce because it doesn’t share the architecture that produced it.

Cognitive variance, in plain terms

Here is the descriptive claim: when average output becomes cheap and abundant, what becomes rare and noticeable is variance — output that is genuinely off the modal distribution. And one of the largest reservoirs of cognitive variance available to any team, school, or community is neurodivergentneurodivergent. Having a mind that works differently from the typical — e.g. autistic, ADHD, dyslexic. A descriptive word, not a deficit label. cognition.

Let me be precise about what I mean. Autistic cognition, in many of its profiles, involves pattern-recognition that works at different granularities and with different salience weights than neurotypical cognition — seeing structure where others see noise, or skipping social structure that others foreground. ADHD cognition often involves attention that moves nonlinearly, crossing domains that linear attention never traverses. The autism literature has a name for part of the first pattern: enhanced perceptual functioning, a documented bias toward local detail and low-level features over the global, socially-weighted gist that non-autistic perception privileges (Mottron et al., 2006). The ADHD side of this is thinner and more contested — I am describing a population-level tendency, not handing you one clean study, and anyone who tells you the creativity literature on ADHD is settled is overselling it. Neither is a universal claim about every neurodivergent person; the variation within neurodivergent populations is enormous. The claim is directional and population-level: minds whose processing diverges from the modal tend to produce outputs that diverge from the modal. In a world that has made the modal abundant, that divergence is noticeable in a way it wasn’t before.

Notice what this does not say. It does not say neurodivergent output is better. Some variance is noise, and not every off-center judgment turns out right. The observation is narrower: neurodivergent minds produce different output, and when the center of the distribution is cheap, difference itself carries information the center can’t.

Maskingmasking. Hiding your natural way of thinking or behaving to fit in. Common among neurodivergent people; it's exhausting and linked to burnout. hides exactly this

Now the part where the description meets lived experience. The cognitive difference just described is precisely what masking suppresses.

As this site’s piece on autistic burnout and masking in knowledge work lays out, masking is the performance of neurotypical cognition — suppressing your actual processing to produce the outputs the modal expects, in the shape it expects them. Read that through the lens above: masking is, literally, the production of modal output from a non-modal mind. It is the thing AI now produces cheaply, performed by hand, at enormous personal cost. The Camouflage Profile on this site maps exactly what that performance costs you — effort, not identity.

So consider what a workplace gets when it “includes” neurodivergent people but quietly requires them to mask — to communicate only in neurotypical registers, to perform attention in neurotypical patterns, to sand down everything legibly different. It pays the inclusion cost, receives modal output, and loses the variance that made the person worth including in the first place. That’s the worst possible trade: the difference is real, the environment penalizes it, and what survives the filter looks like everything else. Meanwhile the person pays for that filter in burnout. Nothing about this requires malice — only defaults that were never examined.

The friction is a mismatch, not a defect

Where does the suppression happen? Mostly in protocols nobody designed.

Consider the friction that gets neurodivergent cognition pathologized in engineering teams: terse review comments that read as hostile, literal communication that skips social lubrication, nonlinear attention that doesn’t track the meeting agenda. The piece on double-empathy problems in your pull requests names these for what they are — not individual deficits, but protocol mismatches. An autistic engineer’s terse PR comment is not a communication failure; it is a different communication grammar, and a team hearing hostility in it is making an error in the other direction. Both sides are misreading each other. That is what double empathydouble empathy. Damian Milton's account of cross-neurotype misunderstanding: autistic and non-autistic people each misread the other's intent, so comprehension breaks down in both directions rather than one side lacking empathy. The friction sits between people, not inside the autistic person. means.

The term is Damian Milton’s (2012), and it was a correction aimed at an asymmetry: research had long treated autistic misunderstanding as an autistic deficit, when the same misunderstandings appear, symmetrically, from the other direction. Crompton et al. (2020) then tested the prediction that follows — if the problem is mutual, autistic people should understand each other at least as well as non-autistic people understand them — and found it. Information passed between autistic peers as reliably as between non-autistic peers, and better than it passed through mixed pairs. The deficit was never in one of the two minds. It was in the pairing.

The practical consequence follows directly. The fixes that piece argues for — structured PRs, async-first norms, explicit communication expectations — are usually filed under “accommodations.” But look at what they actually do mechanically: they remove the social-performance layer that difference has to be squeezed through. Structured feedback surfaces signal that tone-policing buries. Async norms route around the real-time social-processing tax. Explicit norms replace implicit consensus, which always, silently, defaults to the modal. Whether difference shows up as signal or gets spent as friction is, in large part, an environment decision. Most environments have simply never been asked the question.

The strongest objection

The argument above has a serious counter, and it deserves stating at full strength rather than as a strawman.

The counter is about floors. The median output of human work was, in most domains, mediocre — not because people are mediocre, but because most work is done under time pressure by people without the training or the hours that excellence needs. If AI lifts that median substantially, it is an enormous good, spread across billions of documents, decisions, plans and codebases that were never going to get a specialist’s attention. A world where the thousandth-best answer is competent is better, on almost any accounting, than one where it is bad. Set against that, a thinner tail looks like a rounding error — and a self-indulgent thing for someone with the luxury of caring about tails to mourn.

I think that is largely right, and it is not a refutation. Both can hold at once: the floor rises, and the tail thins. They are facts about different parts of the same distribution, and the second does not cancel the first. The claim here was never that homogenization is bad for the world’s average output — it is plainly good for it. The claim is about what becomes rare, and therefore what becomes valuable, and therefore what an environment is choosing between when it smooths its inputs and its people into one shape.

There is a more hopeful version of the counter too: that homogenization is transitional. Fine-tuning, local models, retrieval and deliberate sampling can all widen output again — the default is not destiny. That is probably true, and it is also the point. Widening the output is an act, not a drift. Defaults regress to the mean; recovering a tail takes someone deciding that variance is worth the effort. Which makes it an argument about who intends it.

The honest caveat

This argument goes wrong in one predictable way, so let me close that door myself. Cognitive variance is not universally valuable. Some variance is noise — dysregulation that destroys output, executive dysfunction that can’t be routed around, profiles that produce friction without signal. Not every neurodivergent mind is a reservoir of insight, and the slide from “different minds notice different things” to “neurodivergent people are superpowers waiting to be hired” is reductive, dehumanizing, and wrong. It flattens real people into a marketable caricature of useful difference while ignoring the daily costs and the many people the caricature doesn’t fit.

The claim here stays directional and modest: as AI makes modal output abundant, genuine cognitive difference becomes rarer and more visible; neurodivergent cognition is a significant source of that difference; and environments built on masking suppress it. That is an observation about distributions and defaults, not a hiring pitch and not an RCT. I believe it as someone who builds things with these tools and as a parent who watches one particular different mind meet a world that keeps calling different deficient.

What this asks of an environment

If the argument holds, it converts into a small number of concrete environmental decisions — none expensive, all deliberate.

Evaluate the judgment, not its register. A terse comment, a non-linear route to a correct answer, a question arriving out of order — these are the shapes non-modal thinking often comes in. An environment that scores fluency and confidence is scoring the mask, and the mask is precisely the layer that makes output look modal.

Keep the read that lost. Consensus processes are regression-to-the-mean engines of their own: the dissenting judgment gets absorbed, softened, or dropped, and the record keeps only the shape everyone agreed on. Writing down the position that did not win costs one paragraph, and preserves the one input the median could not have generated.

Make norms explicit, and default to asynchronous. Implicit norms always resolve toward the modal, because they are enforced by whoever needs least accommodation. Written expectations and async-first defaults do something mechanical: they remove the performance layer that difference has to be squeezed through before it can be assessed on its merits.

Do not confuse raising the floor with no longer needing the tail. Those are the two halves of the objection above, and the failure mode is treating the first as a licence to ignore the second. The floor is where most of the value goes. The tail is where the unrepeatable part goes. An environment can want both — and has to say so out loud, because only one of them arrives unasked.

Coda

The age of average is a description of where production is heading: the middle of the distribution, rendered cheap and endless. Whatever else that changes, it changes what difference means. A mind that doesn’t produce the modal output used to be, mostly, a problem to be managed. It is increasingly the interesting part — not because different is better, but because different is the one thing the engine can’t make more of.

Neurodivergent people did not need AI to make their cognition valuable. It always was; the culture’s accounting just never counted it. If the homogenization of everything else is what finally makes that visible, the visibility is welcome — but the work underneath it hasn’t changed. Different isn’t deficient. The environment decides how much of the difference it gets to keep. Build accordingly.