What is falling asleep at the wheel?

Falling asleep at the wheel is what happens when people working beside a high-quality AI put in less effort and judge less well, because the AI's quality gives them little reason to stay engaged. The phrase is Fabrizio Dell'Acqua's. It names a result that sounds backwards: in his experiment, the recruiters given the better AI did the worse job.

A better tool, a worse result#

Dell'Acqua randomly assigned 181 professional recruiters to screen 44 résumés for a software engineering role. Some had no AI. The rest had AI recommendations whose quality was set at three levels: about 99 per cent accurate, about 85 per cent and about 75 per cent, as the Harvard Business School AI Institute describes the design. The paper's abstract reports the pattern that gave the idea its name. Recruiters with the higher-quality AI were less accurate than recruiters with the lower-quality AI.

The abstract also reports why, in the author's reading. Recruiters working with the weaker AI worked harder, spent longer on each résumé and were less likely to accept the AI's pick without question. They improved over time as they learned to work with it, and the gain was concentrated among the more experienced recruiters. The Harvard summary adds that the weaker tool's visible flaws kept recruiters checking its claims, while those with the stronger tool worked faster and leaned on it more.

What an hour of attention is worth#

The mechanism is economic before it is psychological. If an assistant is right nearly every time, an extra minute of checking almost never changes the answer, so a sensible person spends the minute elsewhere. The formal model in the paper makes that argument: as AI improves, the person's reason to pay attention shrinks, and the AI starts to substitute for them where it was meant to support them. That is why the pattern is hard to correct by exhortation. Telling a reviewer to stay vigilant asks them to spend effort at a rate their own experience says is wasted.

It sits close to what the human-factors literature calls automation complacency and automation bias, and to over-reliance, which turns on whether you could still catch an error. What the recruiter experiment adds is a randomised comparison in which the quality of the tool was the thing varied.

How far one working paper reaches#

The evidence is one preregistered field experiment, one task, and a working paper without a journal reference that I could find. The full text sits on a host that refused the fetcher, so this page rests on the abstract and on the author's institute's own summary, and it reports no effect sizes for that reason. The result shows that effort and accuracy moved with AI quality in this setting. It does not show that the same happens with every task, with people who carry personal accountability for the result, or over months rather than a single session.

The sensible reading is a design question. The Harvard summary's own advice is to measure combined human and AI performance as a target in its own right and to build interactions that make the person engage before a recommendation appears, rather than to use worse models. Two 2026 preprints on over-reliance test exactly that kind of design and find early, tentative support for it.

What a team can check this week#

Look for the places where an AI is trusted because it is usually right. In each, ask who last disagreed with it and when, and whether the reviewer spends measurably less time on its output than on a colleague's. A review step that never overrules a good tool carries the signature of the effect, and the count takes minutes.

Key sources

• Dell'Acqua, F. (2022). Falling Asleep at the Wheel: Human/AI Collaboration in a Field Experiment on HR Recruiters. Working paper, Laboratory for Innovation Science at Harvard.
• Mollick, E. Centaurs and Cyborgs on the Jagged Frontier. One Useful Thing.
• Harvard Business School AI Institute. Is AI making your team lazy?
• Maier, S. et al. (2026). Designing Against Deskilling. arXiv preprint 2609.20143.