Accountability and AI

Max Ringelmann measured this over a century ago. As he added people to a rope, the total force rose, but each person’s contribution fell. Later work showed the cause was not poor coordination. In 1974, participants pulled less even when their “teammates” only pretended to pull. The belief alone was enough. Psychologists named it social loafing. Its root is invisibility: when effort cannot be seen, people adjust without noticing. A 1979 follow-up had people shout while believing others shouted with them: with one other person they gave 82 percent of their solo effort, and with five others just 74 percent.

Now put a new partner on the rope. It does not tire, does not complain and appears to pull with great strength. That partner is AI.

The research suggests we treat it exactly as we treat a capable teammate. We ease off.

In a 2023 study, people checking circuit boards for defects found fewer faults when told a robot had already inspected them. They looked, but they looked less carefully. Among the 42 participants, those teamed with the robot found an average of 3.3 defects, against 4.2 for people working alone.

In a study of recruiters, those given a highly accurate AI tool put in less effort and made worse judgements than those given a weaker one. The better the partner, the deeper the sleep. The field experiment, by Fabrizio Dell’Acqua at Harvard Business School, asked 181 professional recruiters to review 44 resumes; those given the weaker AI worked harder and spent longer on each one.

In a 2025 survey of knowledge workers, the more people trusted generative AI, the less critical thinking they reported applying to its output. The Microsoft Research study of 319 knowledge workers found that higher confidence in generative AI went with less critical thinking, while higher self-confidence went with more.

This is the Ringelmann effect with one dangerous difference. A human teammate who slacks off is eventually noticed. An AI never loafs, so we stop asking whether anyone is still pulling. The output arrives fluent and complete. The slack hides inside it, in the unchecked figure, the unread clause, the plausible error.

The answer is not to take AI off the rope. It is to stop treating it as a teammate who shares the load, and to treat it as a tool whose output someone owns.

Name a person accountable for every AI-assisted output, by name, not by team.

Make human review visible: what was checked, by whom, and what was changed.

Keep the judgement-heavy parts with people, where the cost of a quiet error is highest.

Reward catching AI’s mistakes, not just shipping its work quickly.

Responsibility sits with whoever designs the work. Before AI joins the next project, decide who answers for what it produces. A rope with a tireless partner on it still needs a hand that knows when to pull.