Novices gain first, and what that means for everyone else

The pattern across every study

Support agents: the least experienced gained a third, veterans almost nothing. Writers: the weaker gained most, and the gap between best and worst narrowed. Consultants: the below-average gained forty-three per cent, the above-average seventeen. Expert developers on their own code: a loss. The same result, four times, in four occupations.

It is not a coincidence and it is not a property of the people. It is a property of what a model is.

The median, supplied

Recall from the lessons on training what a model learns: the patterns in the pile, weighted toward what was common and well-regarded. What it produces is, by construction, the likeliest good answer: the median of competent practice in its data. Not the best answer anyone ever gave; the answer most competent people would give.

Hand that to a novice and you have handed them what the veterans know. Their output jumps to the median. Hand it to someone already at the median and they get a faster route to where they were. Hand it to someone above the median and they get a suggestion worse than their own, which costs time to evaluate, and, if they trust it, pulls their work down toward the middle. The frontier study measured exactly that pull.

So the direction of the effect follows from one question: where does this person sit relative to the model’s median on this task? The answer differs by task, which is why the same expert developer is helped on an unfamiliar language and hindered on their own project.

Two things this does to a workplace

It compresses. The gap between the best and worst performer shrinks, because the floor rises and the ceiling does not. In the short run this is good for the people at the bottom, and it is the most hopeful finding in the field: a tool that most helps the least advantaged is rare. Whether wages follow performance, or whether a firm responds to a compressed workforce by needing fewer people in it, is not settled by any study yet and is the question that decides whether compression is good news.

It homogenises. A 2024 experiment gave several hundred people a creative-writing task, with and without model-generated ideas. Individual stories from the assisted writers were rated better, especially from the less creative writers. Taken together, they were more alike: the assisted group’s stories clustered, because they had drawn from the same median. Each writer was improved and the collection was flattened. That is the same mechanism seen from above, and it is the answer to why so much writing, design and code since 2023 has a recognisable sameness.

What changes for the expert

The expert’s advantage used to be partly in the median (knowing what good practice looks like) and partly above it: judgement, context, taste, knowing when the standard answer is wrong. The model gives the first part away to everyone. What remains scarce is the second, and the frontier study showed that the second is also the part the model cannot replace and can actively degrade if trusted.

So an expert’s job shifts. Less producing the competent draft, which the novice with the tool can now do. More checking it, knowing when it is wrong, and carrying the consequence. That is a real change, and for a person whose identity was in producing rather than judging it is an uncomfortable one. It is also, structurally, what a senior role was always supposed to be.

What changes for the novice

The novice’s gain is immediate and genuine, and it carries a cost that no study has yet measured over a career. The support agent who reached veteran productivity in three months did so by using the veterans’ patterns without having formed them. Whether that agent, five years on, has the judgement to check the tool (or has only ever known the median) is the question of the next lesson, and it is the one the whole module turns on.

Reading the next claim

When someone says AI makes workers more productive, ask which workers, on which tasks, relative to the model’s median. When someone says it will replace experts, ask which of the expert’s two advantages they mean. And when a firm says the tool has raised its average, ask what happened to its best, because the average can rise while the top comes down, and the studies suggest that it does.