Leaders need honest answers about how staff use AI

This opinion piece, written by Christoph Trattner, professor at the University of Bergen and head of the research centre SFI MediaFutures, argues that business leaders need honest accounts of how employees use AI. A generic note that AI was used says little. What matters is which parts of the work the tool did and how the result was checked. Trattner treats this openness as part of a company's quality control, since a tool's use is hard to follow up if people prefer not to talk about it. He builds on an August call in Digi by Accenture's Trygve Halvorsen for clear ownership and governance of corporate AI use.

Openness can carry a social cost. A 2025 PNAS study by Jessica Reif, Richard Larrick and Jack Soll ran four preregistered experiments with 4,439 participants in total. People who used AI in their work both expected and received more negative judgements of their competence and motivation. Trattner stresses these were experiments on social judgement, not a survey of Norwegian workplaces, but says leaders should check whether staff fear that disclosure will damage their reputation. That does not mean all use is acceptable: sharing confidential data in an unapproved service can be a serious breach, and an employee who delivers an unchecked analysis should be held responsible.

He suggests a useful disclosure describes the work, for example: "AI suggested test cases. I assessed which requirements they covered, added missing cases and ran the tests." A colleague can verify that, which a vague claim of responsible use does not allow. Such a description does not prove everything is correct, but it gives a starting point for control.

Companies could begin with three tiers. Spell-checking and formatting within approved tools and agreed data limits fall under general rules without reporting. Substantial AI contributions to code, analysis or decision material should come with a short description of the contribution and the checks. Use involving confidential information or affecting decisions about people should need prior approval or be banned.

This requires management to provide approved alternatives, explain the limits with examples, and give staff somewhere to ask. Managers must also schedule time for checking, and deadlines should not rest on how fast AI produces a draft. Leaders who describe their own AI use and the errors they caught set the example, and quick reporting of mistakes helps limit damage. Trattner's closing question for leaders is what happens in their organisation when someone honestly describes how they used the tool.