AI raises the floor more than the ceiling. It still needs real human expertise.

+15% on average

Productivity gain for support agents with an AI assistant. The least experienced gained the most; the most experienced saw small gains in speed and small declines in quality.

peer-reviewedBrynjolfsson, Li & Raymond · Quarterly Journal of Economics2025 · n = 5,172

+26%

More tasks completed with Copilot across three company trials, with larger gains for less experienced developers.

peer-reviewedCui et al. · Management Science2026 · n = 4,867

≈ experts

On the writing task, marketing specialists with AI all but matched the AI-aided in-house writers. The group furthest from the skill did not move: you need enough to refine the output.

reportHarvard Business Review2026-03 · n = 78

−19%

Less likely to be correct with AI on a task outside the model's frontier. Inside it, more tasks, faster, at higher quality.

peer-reviewedDell'Acqua et al. · Organization Science2026 · n = 758

−10% high performers: likely +15%

The same GPT-4 advisor: low performers lost, high performers likely gained. The gap came from which advice each chose to act on.

peer-reviewedOtis et al. · Management Science2026 · field experiment, Kenya

319 surveyed

Knowledge workers surveyed: the more they trusted the AI, the less critical thinking they reported. Confidence in their own skill ran the other way.

peer-reviewedLee et al. · Microsoft Research & CMU, CHI2025

−19% was −15%

Employment for ages 22–25 in AI-exposed jobs, relative to peers in less-exposed ones. Experienced workers show no comparable gap. Mainly reduced hiring, not separations.

working paperBrynjolfsson, Chandar & Chen · Stanford, ADP payroll2026-08

≈ −9%

Junior employment at firms adopting generative AI, relative to non-adopters, after six quarters. Senior employment at the same firms kept growing.

working paperHosseini Maasoum & Lichtinger · Harvard2026 · 65M workers, résumé data

−65%

Entry-level hiring at the twelve companies SignalFire calls the Tech Majors, since 2019. Their advice to new grads: demonstrate proof of work before anyone gives you permission.

reportSignalFire · State of Tech Talent2026-06

−17 pts 50% vs 67%

On a comprehension quiz, for developers who learned a new library with an AI assistant to hand. No significant speed-up either.

randomised trialAnthropic2026-01 · n = 52

−17% hints arm: ≈ 0

On the exam, after practising with a model that gave answers. The hints-only arm, built on teachers' solutions, largely removed the harm.

peer-reviewedBastani et al. · PNAS2025 · n ≈ 1,000

28.4% → 22.4%

Endoscopists' unassisted detection rate after three months of AI-assisted colonoscopy. Observational; the authors say exposure might reduce it.

peer-reviewedBudzyń et al. · Lancet Gastroenterology & Hepatology2025