The Average is Automated

Let’s imagine, for a moment, that we’re observing a group of students in a class. Each arrives with a different combination of curiosity, interests and ways of seeing the world. In theory, after three years of learning, we might expect those differences to compound. The more they learn, the further their paths should diverge.

Then we notice something. The teacher stops teaching and starts answering, and she becomes exceptionally good at it. Whatever the problem, she can produce a competent solution almost instantly.

Nobody struggles with the work anymore.

At first, this looks like a remarkable improvement. Her students work faster. Their grades improve. Even the weakest students can suddenly produce work they were previously incapable of.

But after a while, their work begins to look alike. Not because the students entered the class alike. They didn’t. Not because the teacher lacks knowledge. She doesn’t.

Their differences have become less necessary to the work.

We can stop imagining.

We are the students.

We now have unprecedented access to systems that can help us write, code, design, research, analyse and ship in seconds. We were supposed to diverge. Instead, we’re converging.

Welcome to the industrialisation of the mean.

The distance between idea and execution has collapsed. Yet much of what emerges on the other side looks oddly familiar: the same SaaS dashboards, the same landing pages, the same content formats, the same models repackaged into marginally different products. Wrappers on wrappers, all polished enough to work and familiar enough to understand.

For most of history, getting something you wanted required knowing quite a lot about how to get it. Most of the distance between intention and result still had to be crossed through some combination of knowledge, skill and time.

Civilisation has spent much of its existence shortening that distance.

Writing allowed knowledge to survive beyond the limits of individual memory. Machines took on physical work. Calculators took on maths. Search engines made retrieving information almost instantaneous.

We have been moving parts of the burden outside ourselves for a very long time.

There is nothing alarming about this. I do not believe anybody became less human because they stopped washing clothes by hand, nor do I feel particularly compelled to memorise every route I travel in case Apple Maps is rotting my intellect. Technology is useful precisely because it removes effort we no longer consider worth spending.

But friction was doing something besides wasting our time.

For most tasks, producing a convincing result required possessing at least some of the skill that result appeared to demonstrate.

AI separates those two things.

You can now produce polished prose without being much of a writer, functioning software without understanding much code, or a compelling strategy using concepts you would struggle to explain independently.

The result can look capable even when the understanding behind it is not.

That creates two fairly convenient conclusions.

The first is that access to a capability begins to feel like possession of it. The second is that possession no longer seems particularly necessary anyway.

If the desired output looks good enough, understanding how any of it was produced starts to feel less like competence and more like overhead.

To understand how the mean emerges, it’s worth being precise about what these models are actually doing.

At its foundation, a language model has learnt an enormous number of relationships from the data it was trained on. Given some context, it predicts what might plausibly come next and generates from those possibilities. The result is shaped both by what it learnt during training and by the context we provide it now.

That means there are a vast number of possible outputs.

Given all that possibility, why do we keep converging on the same things?

Enter Henry.

Suppose Henry asks Claude to “design me a landing page for my B2B SaaS company.” Although it sounds like one request, it isn’t. Henry has implicitly asked Claude to make hundreds of decisions.

What should dominate the page? How should the company describe itself? What should the type look like? How much space should there be? What comes after the hero? Should there be testimonials? A feature grid? Pricing? What should the buttons say? How restrained or expressive should the whole thing feel?

Henry has specified almost none of this, yet Claude is expected to give him a great landing page.

So it fills in the blanks using what it has learnt about the things Henry has specified. “B2B SaaS” and “landing page” sit inside a dense web of conventions and associations: hero sections, feature grids, social proof, calls to action, familiar layouts, and familiar ways of signalling that this is a serious piece of software intended for serious people who presumably have several browser tabs open.

None of those choices is the mathematical average.

They are defaults.

And defaults compound.

The typeface Henry never chose influences the visual tone. The hierarchy he never specified influences the layout. The layout makes some sections more natural than others. Those sections invite familiar copy. One plausible decision makes the next more likely.

By the time the page is finished, Claude may have made hundreds of perfectly reasonable choices on Henry’s behalf.

The problem is not that Henry used AI.

The problem is that Henry supplied the destination and delegated most of the direction.

The value of these systems is that they give us access to capabilities we do not personally possess. Henry doesn’t need to become a designer; that would defeat the point. But he can know what his customers care about. He can bring references. He can have preferences. He can reject the obvious solution and recognise when something technically competent is completely wrong for what he is trying to build.

The more of that Henry supplies, the fewer decisions the model has to make on his behalf.

So AI is not average in any literal sense. The possibility space is enormous, but possibility and direction are not the same thing.

When we leave the direction unspecified, we should not be surprised when the defaults begin to look familiar.

Henry, it turns out, has never run a company and doesn’t know much about building either, but fortunately for him, software increasingly knows a lot about software. He can prompt his way through a prototype, draft a sales strategy and prepare fairly convincing answers to questions he hadn’t thought to ask himself.

How hard can building Microsoft really be when you’ve got GPT Astra and three usage limit resets?

Henry is hardly alone.

For the first time, anyone can turn an idea into something functioning with little money, time or technical ability. At almost exactly the same moment, large amounts of capital are concentrating around a small collection of technologies, products and business models.

Production has become cheap just as the signals about what is worth producing have become impossible to miss.

The more signals a category accumulates, the more legible it becomes. More founders enter. More companies raise. Analysts construct comparables. Customers become more comfortable buying. Funds that missed the first winner start looking for the next one.

Eventually there are enough companies, funding rounds and revenue numbers that the category no longer looks speculative.

It looks like a market.

Humans look for evidence. Markets manufacture it.

And when building becomes cheap enough, that evidence does not merely attract capital.

It attracts production.

In the first quarter of 2026, roughly 80% of global venture funding went to AI companies. Four companies alone absorbed nearly 65% of all venture investment globally during the quarter. By the end of the first half of the year, OpenAI and Anthropic had taken 43% of all global startup funding.

So there is obvious demand, real revenue and an extraordinary amount of money being pumped into the same market at the same time.

There we go: huge numbers of people with access to similar capabilities, watching many of the same signals, and able to act on them almost instantly.

The market does not need to tell us to become more alike.

It only needs to make the familiar easier to explain, easier to fund and easier to copy.

You cannot escape a loop you are consistently being incentivised to run. Views, likes, revenue and rankings tell us what appears to be working. Other people’s choices are no longer simply visible; they are quantified, copied and made more familiar until familiarity starts helping them perform.

The internet did not invent imitation. It just gave imitation analytics.

Take outbid.lol, a wonderfully stupid website where companies literally pay to move higher up a leaderboard. People will, apparently, buy anything.

Within a couple of days came the variations: more pay-to-rank sites, pay to put your logo on someone’s Mac or iPad, even people offering space on their bodies. People saw a thing work, the numbers were public, the method was simple, and the cost of producing another version had become negligible.

Why wouldn’t you try?

Consider this an ode to the AI headshot generators, the AI meeting note-takers and the AI email assistants: categories that fill with near-identical products the moment the first one posts its revenue.

Entire companies now pitch themselves as “Cursor for lawyers” or “Perplexity for recruiters”, borrowing the credibility of something that already worked. The name tells investors what to expect. It also tells you exactly how different the product isn’t.

It is easy to watch somebody with no obvious advantage over you ship something in a weekend, go viral and make some money, and arrive at the obvious conclusion: well, why can’t I?

With a capable enough model, yesterday you were a chemist; today you’re shipping an ML benchmark and arguing with researchers on X. Sometimes that accessibility produces people doing genuinely remarkable things outside the boundaries they were previously kept inside. Sometimes it produces another wrapper.

The system is largely indifferent so long as people click.

The average has always existed.

Most things humans have made were not remarkable. Most paintings were forgotten, most buildings were ordinary, and most writers were not Shakespeare. History has a brutal filtering mechanism. The exceptional survives disproportionately well, leaving us with the strange impression that previous generations spent all day producing masterpieces.

The average has become cheap enough to reproduce without limit.

AI is not necessarily collapsing the entire distribution towards the middle. It may simply be increasing the volume of the middle at a rate we have never previously experienced.

What becomes scarce, then, is not output.

It is judgement.

And judgement was never handed to us. Much of it was built through the very friction we are now learning to remove. You learnt what good writing looked like by writing badly. You learnt what good code looked like by breaking things. Sometimes the struggle was how understanding got made.

So the tools that make judgement more valuable may also make it harder to acquire.

This is why “human in the loop” is less comforting than it seems. A human in the loop is only useful if they understand enough of the loop to intervene. Otherwise, we have kept a human in the system but not their judgement.

The answer is not to stop using AI. Use more of it. But choose where to keep the friction. Pick the things you want to understand, and do them the hard way, at least some of the time.

Ask it to show you the possibilities you would not have reached yourself. Ask it to attack your assumptions. Give it the weird references. Tell it what you hate. Make it defend its choices. Learn enough about the thing you are building that you can tell when the polished answer is nonsense.

You do not need to become the expert in every room you enter. That would defeat much of the point. But you do need enough understanding to recognise when the room is heading somewhere you do not want to go.

Our students never needed a worse teacher.

They needed to remain capable of disagreeing with her.

Perhaps the future distinction is not builder versus prompter, technical versus non-technical, or even human versus machine.

It is between those who use abundance to reproduce what already has signals behind it and those willing to use the same abundance to explore what does not.

The average has always existed. Now it has an API.