Dear Readers,
imagine an employee using an AI tool to prepare a report. The output sounds plausible, it is well written and it saves time, so he uses it without questioning it much further. Two weeks later, it turns out that a key figure was wrong. Who made the mistake?
This question is harder to answer than it may seem, and that is exactly the real issue.
Mistakes have always been part of working life. But as AI becomes part of everyday work processes, the logic of how mistakes arise is changing. In the past, a mistake could usually be traced back quite clearly to a person who had made a decision or misjudged a situation. Today, mistakes emerge through the interaction of people, models, and organizational context, within systems that are increasingly difficult to fully oversee.
Mistakes are therefore no longer just an individual matter. They have become a systemic phenomenon. For leaders, this marks an important shift.
Where errors arise in the use of AI
AI systems generate probabilities, not truths and a language model can produce plausible-sounding responses that are factually incorrect, so-called hallucinations. Without human review, such errors can flow directly into decision-making, often without anyone noticing.
Paradoxically, people with a moderate level of AI knowledge are often the most likely to overlook these errors, because they underestimate the complexity of the model and place more trust in its outputs than in their own judgment. Research describes this pattern as the Dunning-Kruger effect, and under time pressure it becomes even more pronounced, because the brain conserves effort where it can, while AI outputs usually sound convincing enough not to invite further scrutiny. You may recognize this from your own experience: an output seems coherent, time is short, and it simply remains unchanged.
In addition, an error in an AI system behaves differently from a human error, because it can be replicated automatically across processes and decisions before anyone detects it. And once a machine has been involved in the process, people often begin to feel as though it made the decision, which noticeably reduces vigilance and critical thinking without those involved necessarily being aware of it.
What leaders can do in practice
These effects cannot simply be dismissed, but they can be managed.
One way to address them is to ask employees to formulate their own assessment before looking at the AI output. People unconsciously anchor themselves to the first suggestion they see. Anyone who records their own view first is more likely to compare more carefully afterward and to give greater weight to their own judgment.
Small interruptions in the process can have a similar effect, for example a brief requirement to justify a result before adopting it. AI outputs often sound so complete and matter-of-fact that people hardly think to question them. A built-in pause can change that.
Training also helps when it focuses not only on what AI can do, but above all on where it fails. Anyone who understands how hallucinations occur and why models sometimes respond in biased ways is better able to judge when caution is needed.
When leaders make it clear in everyday practice that AI is there to support, while human beings retain final judgment and accountability, responsibility remains anchored where it belongs.
A look at everyday work
It is worth pausing for a moment to ask where AI is already shaping day-to-day work and whether the team knows how to deal with it. A few questions can help make visible what is already working well and where there is still room for improvement:
Teams that discuss these questions often realize quite quickly that this is less about technology than about mindset, about who really looks closely in the process and who simply relies on the assumption that it is probably correct.
Yours,
Dr. Sandra Dentler
Image source: Photo by TheYuri Arcurs Collection on Magnific