Does AI make weak performers strong? The studies say it amplifies the judgment you bring. Same tools, opposite results for careful and careless users.
Does AI close the gap between your best people and your weakest, or widen it? This is the question every manager buying seats is quietly betting on, and the honest answer is that we don’t know yet in general. Randomized studies have found both. In one, the least experienced workers gained the most. In another, the weakest got worse.
Those results are less contradictory than they look. Read side by side, they point at one pattern: on anything beyond a narrow, well-specified task, AI multiplies judgment. It does not supply it. Before getting to the evidence, it helps to watch the multiplier work on one person.
One analyst, one question, and four checks
A revenue analyst at a subscription software company gets a Tuesday request from her VP: why did churn rise last quarter? She has a research agent, a spreadsheet assistant, and until Thursday. (She is illustrative, a composite built from the shape of the work, not a real person.)
In an illustrative afternoon, the tools produced the volume and every step that kept the answer true was hers.
The two tools saved her most of a day. They also produced two errors that would have survived a skim: a citation that didn’t support its sentence, and a formula that was correct for a definition of churn her company doesn’t use. She caught both because she knew how the billing system treats a paused account and because she opened sources instead of trusting them.
Hand the same tools and the same Tuesday to someone without that knowledge and they finish faster still. Their answer is better formatted, better cited, and wrong.
The model: AI amplifies the judgment you bring
Here is the model in one sentence: AI raises the volume of what you produce, and it raises the reach of your mistakes by the same factor, so what you get out depends on the judgment you put in.
That makes two independent axes. One is how much you use AI. The other is how good your judgment is on the task in front of you.
AI moves people sideways on this grid. Only practice moves them up.
The analyst sits top right. Her counterfactual twin sits bottom right, and from the outside the two look identical: same tools, same speed, same polish. The difference shows up later, when someone acts on the number.
Four findings from the research test that model. Two of them look like they break it.
Finding 1: inside the frontier AI levels people; outside it, it misleads them
The most cited test of AI on knowledge work gave 758 Boston Consulting Group consultants access to GPT-4 and split the work into tasks the model could do well and one it could not.
On tasks AI handles well, weaker consultants gained the most. On the task it handled badly, using it made people less likely to be right.
Fabrizio Dell’Acqua and colleagues called the boundary the “jagged technological frontier,” because it is invisible and uneven: two tasks that look equally hard can sit on opposite sides of it. Inside it, the top half gained least, which reads like AI as an equalizer.
Outside it, the people who did best were the ones who noticed the model was wrong and overrode it. That noticing is judgment, and the tool did not provide it. It actively competed with it, by producing a confident answer that was easy to accept.
Finding 2: AI lifts novices when the task already contains someone else’s judgment
This is the strongest counter-evidence to the model, and it deserves a straight hearing.
The support assistant lifted novices because it was carrying the best agents' judgment to them.
Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied a generative AI assistant rolled out to more than 5,000 customer support agents at one software firm, first as a 2023 NBER working paper and then in the Quarterly Journal of Economics in 2025. Productivity rose on average, and the gains went mostly to novices and lower-skilled agents. The most skilled barely moved, and their conversation quality dipped slightly.
Look at where the assistant’s suggestions came from, though. The authors note that top performers contributed many of the examples the system was trained on. What reached the novices was their best colleagues’ judgment, compressed and delivered mid-chat, on a task with a known shape: a customer has a problem, and there is a playbook.
That is the reconciliation. AI levels people on work where someone else’s judgment is already built into the task. Where the user has to supply the judgment, it amplifies whatever they bring.
Finding 3: when the user has to choose, the gap widens
The cleanest test of the second case gave small business owners an AI mentor and let them decide what to do with its advice.
Strong and weak entrepreneurs got similar advice. What they chose to act on split the outcomes in opposite directions.
Nicholas Otis, Rembrand Koning and colleagues ran this as a field experiment in Kenya. On average the AI mentor did nothing measurable. That average hid a split: high performers improved and low performers got worse. The authors trace the difference not to the questions people asked or the advice the model gave, but to which advice the entrepreneurs chose to implement.
A mentor that answers every question fluently is only as good as the person deciding which answers apply. That is the bottom-right square of the grid, measured.
Finding 4: nobody can feel their own multiplier
If AI amplifies judgment, you would hope people could at least tell which way it is amplifying them. The best evidence says they can’t.
Experienced developers predicted a speedup, measured a slowdown, and still believed they had been faster.
METR’s July 2025 randomized trial let 16 experienced open-source developers use AI tools (mostly Cursor with Claude Sonnet models) on some of their real issues and banned them on others. With AI, issues took 19 percent longer. Before starting, the developers forecast a 24 percent speedup; afterwards, they still estimated 20 percent.
These were strong engineers, and the lesson is about perception more than skill. The feeling of speed and the fact of speed came apart, and nobody inside the study could see it. If you can’t feel your multiplier, you have to measure it, and most people never do. The same gap shows up in review queues, covered in The Engineer After the Editor.
The same tool, two users
Every tool on the market today sits on both sides of this grid at once. Which side depends on the hands.
No product decides which square you land in. The scoping, rejecting and checking around it does.
Coding agents make this sharpest. Claude Code, Cursor’s agent, GitHub’s Copilot coding agent and OpenAI’s Codex can all take a task, edit across many files, run commands, and hand back a diff or a pull request. A strong engineer uses that reach on a task small enough to verify, with the check written before the code. A weak one hands over a vague feature request, gets back a large diff that compiles, and merges what they can’t evaluate.
Deep research tools from OpenAI, Google and Perplexity produce long reports with citations, and a citation feels like verification. It isn’t until someone opens it. The careful user also asks a narrow question, because a broad one invites a confident survey of the internet’s average opinion.
Writing and spreadsheet assistants fail the same way. The draft and the formula both look finished, and finished-looking work is the hardest kind to reject.
What this gets wrong
The grid is a simplification, and four of its edges are soft.
The model is strongest on unscripted work today and weakest as a permanent law.
Judgment is domain-specific. The analyst who catches the paused-account bug might accept a bad legal clause from the same chat window ten minutes later. “Good judgment” is really a long list of places where you know what wrong looks like, and the grid flattens that list into one score.
The frontier also moves. In early 2026 METR reported that its follow-up study, run with newer tools, suggested developers are probably faster with AI now than in 2025, but it called its own data an unreliable signal, partly because many developers would no longer agree to work without AI. Every capability jump pulls some tasks inside the frontier, where the leveling effect from Finding 1 applies. The model says nothing about how fast that happens.
The axes are not fully independent either. Brynjolfsson’s team found that support agents kept some of their gains during periods when the tool was unavailable, which suggests they learned from it. A well-used assistant can be a practice partner. That is the most hopeful edge of the model, and the least measured.
What survives these caveats is the direction. On work without a script, the tool does not know what you don’t know, and it will not stop you from shipping it.
Your next move: practice the checks
Adopting a tool takes an afternoon. Moving up the grid takes reps, and the reps are specific. Before you hand a task to an AI, write one sentence describing what a wrong answer would look like. After it answers, check one claim against something it didn’t produce. Throw away at least one draft on purpose, and notice why.
Those three habits are what the analyst did on her Tuesday. None of them needed a better model. They needed her.
20+ years of experience in technology leadership, innovation, and digital transformation. Building and scaling technology ventures.
Only if you find it useful
No pitch here. If these pieces are worth your time, you can get new ones in your inbox. If not, skip it with a clear conscience, nothing is being sold. Rare emails, no spam, leave whenever you like.
We use cookies to enhance your experience. Also, a friendly heads-up: some content here might be slightly exaggerated or embellished because, well, we're humans and sometimes we get excited about cool tech stuff! 😄
"In a world of AI precision, we proudly maintain our human tendency to occasionally overstate things."
