What psychological safety requires when someone challenges a model

A junior analyst flags that the forecast the model just produced does not match the last three quarters of actuals. The room goes quiet. Someone jokes that the analyst needs to trust the tool more. The forecast goes into the pack unchanged.

Jayshree Seth and Amy Edmondson describe a pattern leaders are starting to see as AI moves into daily work. The productivity gains don't show up as expected. People start to second-guess themselves, and trust erodes in ways that are hard to pin down. Their fix is to make doubt about AI output something a team can talk about, and to be clear about when human judgment overrides the model.

Why challenging the output feels risky

Speaking up about a wrong number on a spreadsheet was already hard. Questioning a model can feel worse, because the tool arrives looking objective. The person who pushes back risks looking like they're behind on the rollout. Teams that already punish bad news will punish this news faster.

The cost shows up later. Errors travel further when nobody wants to be the first to say the model got it wrong. People also stop learning how the tool fails, so each check gets weaker than the last.

What leaders have to make ordinary

Edmondson and Seth put the work on the team and the person leading it. I read their argument as three habits a manager can build without buying a new platform.

1. In the meeting where you'll use the output, ask out loud what would make you distrust this answer.
2. Name the point where a person must override the model, and write that rule down where the work happens.
3. Treat the first challenge as useful information about the team before you treat it as a training gap for the person who spoke.

These habits teach a team how to argue with a dashboard. Without them, people tend to trust it blindly.

What to watch this month

Listen for jokes that punish doubt. Watch whether the most junior person in the room ever questions an AI draft in front of the most senior. If they never do, the team's climate is the problem, and the model will make it worse.

Collective Campus has run innovation programmes with enterprises since 2015, and we've done AI capability work with teams since the generative AI wave began in 2022. Our psychological safety workshop is where intact teams practise how a leader replies when someone raises a problem, including a problem with a model.