007 / Essay ·
AI lifts the bottom and the top. Not the middle.
Some experiments find that AI helps the least skilled most; others that it helps the best. Both are right. An AI tool lifts those below its frontier and multiplies those above it, so the gain is V-shaped in ability, and which side shows up depends on how much the model knows.
Generative AI helps the weakest and the strongest workers most, and the competent middle least. When a large customer-support operation gave its agents an AI assistant, resolutions per hour rose by about 14% on average. The gain was concentrated at the bottom: novice and low-skilled agents improved by 34%, and the most experienced barely changed. Around the same time, a field experiment in Kenya gave small-business owners an AI mentor on WhatsApp. There the pattern ran the other way. The best-performing entrepreneurs gained and the weakest did worse than without the tool.
Results like these are usually filed under one of two theories. Either AI is skill-biased, like computers in the 1990s, and widens the gap between the able and everyone else. Or it diffuses knowledge, handing the novice what the expert knows, and narrows the gap. Each new experiment is counted as a point for one side. I think that, within a job, both sides can be describing the same curve from different ends.
- An AI tool does two separate things. It solves some problems for anyone who asks (its frontier), and it helps people get through more work (its leverage).
- The frontier lifts those below it most. Leverage pays those above it most. So the gain from AI is V-shaped in ability, and the worker who gains least is the one whose knowledge sits exactly at the frontier.
- Whether AI looks skill-biased or equalising depends on where the frontier sits in a given job. As models improve, the same job flips from one to the other.
- The two camps are not wrong about the technology. They are looking at different jobs, or the same job at different moments.
What a worker knows
Start with a simple description of skill, due to the economist Luis Garicano. A job is a stream of problems of varying difficulty: a customer who cannot log in, a refund that crosses two systems, a billing dispute nobody has seen before. A worker's ability is the hardest problem they can solve on their own. Everything easier gets solved; everything harder gets escalated, or dropped. More able workers produce more because they solve more of the stream.
This is a narrow view of skill, and deliberately so. It turns the question of who gains from AI into a concrete one: which problems does the tool let each worker solve that they could not solve before, and how much more of the stream can they get through?
Two things an AI tool does
Seen this way, an AI tool has two properties that are easy to confuse.
The first is its frontier: the hardest problem it can solve for anyone who asks. Below the frontier, knowledge has been codified. A novice who has never seen a problem can get the answer by asking. The frontier puts a floor under everyone's output, and that floor is worth most to those standing furthest below it.
The second is its leverage: the extra work it lets a person get through. It drafts the reply, finds the policy, formats the report. Leverage multiplies whatever a worker already produces, so it is worth most to those who already produce the most.
The two move with different kinds of progress. The frontier moves with model releases: a new model knows more. Leverage moves with interfaces, integration into workflows and the user's own fluency. They are usually bundled into a single number called "AI productivity". They should not be, because they push the distribution of gains in opposite directions.
The gain from AI, by worker ability
Gain in output from using the tool, Δ(a) = (1+λ)·max{a, k} − a, for a worker of ability a, with frontier k and leverage λ (Research Note 002, Proposition 1). Ability and problem difficulty are spread evenly between 0 and 1, and every solved problem is worth the same. Under these assumptions output without AI equals ability, so a gain of 0.2 is a fifth of the output of the most able worker. The V shape holds for any distribution of ability and difficulty; the straight lines do not.
The valley
Put the two together and the gain from AI has the shape in Figure 1. Below the frontier, it falls with ability: the less a worker knows, the more of the stream the tool unlocks for them. Above the frontier, it rises with ability: these workers no longer need answers from the tool, but leverage scales what they already do. The bottom of the valley is the worker whose own knowledge sits exactly at the frontier. They do not need the AI's answers, and they are not yet expert enough for extra throughput to be worth much.
The two theories are statements about the slope of this curve. Skill bias means gains rise with ability; knowledge diffusion means they fall. In the model the slope is negative on one side of the frontier and positive on the other. The question "is AI skill-biased?" has no answer until you say which workers, relative to which frontier.
The customer-support result now reads differently. Support is a job with a large stock of routine, documented problems, and a tool trained on the best agents' conversations knows most of them. Its frontier sits high relative to the workforce, so most agents are on the descending side of the valley, and the novices gain most. Studies of mid-level writing and of software developers using coding assistants found the same pattern for the same reason.
The Kenyan result needs one more ingredient, and it is worth being honest about it. In the pure valley no one loses. Losses appear when the tool also answers questions it cannot reliably answer, and its wrong answers look plausible. Then the ability that matters is spotting the bad advice, and that ability rises with experience. Research Note 002 shows that this kind of judgment becomes the skill-biased margin of AI, and that workers who trust the tool too much can end up worse off, with losses concentrated at the bottom. That is a story for another essay.
The sign flips
Where the frontier sits is not fixed. It moves up with every model release. So the average slope of gains on ability in a job, the number a study would estimate, changes as the technology improves.
Does AI favour the able? It depends on the frontier
Slope of a regression of the gain from AI on ability across all workers in a job, S(k, λ) = (1+λ)(1−k)²(1+2k) − 1 (Proposition 2), at the leverage set in Figure 1. Above zero, gains rise with ability on average; below zero, they fall. The marker follows the frontier set in Figure 1. For any distribution of ability, the slope falls as the frontier rises and crosses zero exactly once.
When the frontier is low, it sits below most of the workforce. Nearly everyone is on the rising side of the valley, and AI looks skill-biased. As the frontier passes through the distribution of ability, more and more workers fall below it, and the average slope turns negative. With 20% leverage and an even spread of ability, the flip comes once the tool can solve about the easiest quarter of the job's problems (k ≈ 0.26), when only one worker in four sits below the frontier. More leverage delays the flip, because it makes the rising side steeper.
Two things follow. First, the sign of an estimated AI×skill interaction is a property of a technology vintage and a job, not of AI in general. In this model, an experiment that finds knowledge diffusion in customer support says that the frontier sat high in customer support in 2020–21. It does not tell us that AI is equalising in law, medicine or research. Second, the same job should drift from skill-biased to equalising as models improve, and studies run two years apart in the same occupation can disagree without either being wrong.
Two dials, opposite effects
The model also says that the two kinds of progress have opposite distributional effects, even when their average effect is the same. Raising the frontier helps only workers below it, and helps the least able most. Raising leverage helps everyone in proportion to what they already produce, and so helps the most able most. A new model release and a better integration into the workflow can raise average productivity by the same amount and move the distribution in opposite directions.
The same split explains why studies of inequality can disagree on the same data. With leverage that is the same for every worker, AI never raises inequality in relative terms: the gap between the 90th and 10th percentile, measured as a ratio, does not widen. But leverage scales absolute gaps, so the difference between top and bottom in units of output can widen while the ratio narrows. A study that works in logs and one that works in levels can reach opposite conclusions, and both can be right.
The results are theorems in a deliberately simple model: one dimension of ability and a tool that never errs below its frontier. Two simplifications in the figures are not needed for the result. The V survives leverage that rises with ability, as long as it does not rise too steeply, and if the frontier is fuzzy rather than a sharp line, the kink becomes a smooth valley whose floor sits at an observable point of the ability distribution (Proposition 3). The experiments described here are consistent with the model; they were not designed to test it. Most report effects for two or three groups of workers, which cannot tell a V from a straight line. A direct test needs effects estimated decile by decile of baseline ability, in a job where one can measure which problems the tool solves on its own.
There is also a competing explanation. Enrique Ide and Eduard Talamàs, in a model of firms as knowledge hierarchies published in the Journal of Political Economy, reconcile the same experiments through the tool's autonomy: AI that only advises helps the least knowledgeable most, while AI that can do the work itself helps the most knowledgeable. The experiments so far involved advisory tools and cannot tell the two accounts apart. They come apart for tools whose main effect is speed rather than knowledge. Here such a tool favours the able; in their model an advisory tool favours the least knowledgeable.
The argument would be wrong if, with gains measured decile by decile of baseline ability in a job where the tool solves some but not all problems, the gain rose or fell steadily across the whole range; if a model upgrade that improves accuracy but not speed made the gains more skill-biased rather than less; or if a tool that mainly speeds people up helped the least able most.
What follows
For a firm deciding how to roll out an AI tool, the useful question is not whether its workers are skilled but where they sit relative to the tool's frontier in this job. Workers well below it gain most from the answers. Workers well above it gain most from throughput, and need tools built for leverage rather than for answers. The workers in the middle, competent but not expert, gain least, and are the easiest to overlook when the tool is judged by its average effect.
For researchers, the model points to a cleaner experiment than the ones run so far. When a vendor swaps the underlying model inside an unchanged product, the frontier moves while leverage stays roughly fixed. The theory predicts that the interaction between AI use and baseline ability falls after the upgrade, and that the bottom of the valley moves up the ability distribution. If it rises instead, the argument is wrong.
And for the debate itself: in this model, the question "does AI help the weak or the strong?" has a precise answer. It helps the weak with what it knows and the strong with what it does. Which effect dominates depends on how much it knows, and that keeps changing.
Methods and sources
The model, propositions and proofs are in Research Note 002, The Valley at the Frontier: Who Gains from Generative AI, Sections 3–5 and the appendix; all results are checked symbolically and numerically in the replication code. Figures use the paper's uniform benchmark with illustrative parameters; they are not estimates. The description of ability follows Garicano (2000). Evidence: Brynjolfsson, Li and Raymond (2025) on customer support; Otis, Clarke, Delecourt, Holtz and Koning (2024) on Kenyan entrepreneurs; Noy and Zhang (2023) on professional writing; Peng et al. (2023) and Cui et al. (2025) on software developers; Dell'Acqua et al. (2023) on management consultants. The competing account is Ide and Talamàs (2025), "Artificial Intelligence in the Knowledge Economy", Journal of Political Economy 133(12).