AI on the job: is one mistake enough to rule out a tool?

When AI makes a mistake, a team may question its use far beyond the case at hand. The managerial question is precise: how should we decide whether to continue or suspend its use without mistaking a memorable error for an assessment of reliability?

What the literature establishes

Berkeley Dietvorst, Joseph Simmons and Cade Massey (2015, Journal of Experimental Psychology: General) show, in forecasting experiments, that participants may avoid an algorithm after seeing it make mistakes, even when it outperforms the human forecaster. Observing errors undermines confidence in the algorithm more than in the human. The decision to use a tool therefore does not necessarily reflect its comparative performance.

A preference that can also favour the algorithm

Jennifer Logg, Julia Minson and Don Moore (2019, Organizational Behavior and Human Decision Processes) show, across several estimation experiments, that participants give advice more weight when they believe it comes from an algorithm rather than a person. This finding does not directly contradict the earlier one: receiving advice labelled as algorithmic is not the same situation as watching a tool fail. The source label and the experience of its errors are two distinct factors.

The premature conclusion

One might conclude that teams should be taught to tolerate AI errors to encourage adoption. That would confuse a psychological response with a management rule: some errors warrant suspension, particularly when they compromise safety or confidentiality. The useful criterion is not tolerance itself, but the severity of the error and the comparative reliability of the available alternatives for the same task. (our executive and employee training programmes)

What these findings cannot decide

These studies concern judgements and forecasts in experimental settings, not the sustained deployment of generative assistants within an organisation. They establish neither an acceptable error threshold nor a validation rule for medtech, academic or sporting applications. A stated preference for a tool demonstrates neither its safety nor its usefulness in the workplace.

A practical check in Lausanne

In an SHR “IA on the job” workshop in Lausanne, we propose an exercise for teams from medtech, higher education, international sport and the Lake Geneva region's scale-ups, using a fictional administrative case without sensitive data. Randomly assign participants to two versions of the same case file, showing identical results and an identical error but attributing them either to AI or to a person, then ask whether they would use that source again. The measure is the proportion of favourable responses in each group, reported alongside group sizes and reasons for refusal. Any difference provides a basis for discussing suspension criteria; it is neither evidence of individual bias nor validation of the tool. To go further: explore the AI on the Job training in Lausanne, or browse our executive and employee training programmes in Switzerland.

In pictures: AI on the Job in Lausanne

• ai on the job
• Lausanne
• research
• training lausanne