How a team says an AI mistake out loud

The number is already in the board pack. It came from the forecasting assistant. A junior analyst notices it does not match the last extract they pulled. They draft a note, delete it, and let the meeting run.

Google's re

guide records what Project Aristotle found when the company asked what made a team effective. The researchers studied 180 teams, 115 project teams in engineering and 65 pods in sales, a mix of high-performing and low-performing teams. Teams ran from three people to fifty, with a median of nine. They ran hundreds of double-blind interviews with leaders and more than 35 statistical models. How the team worked together predicted effectiveness.

What the finding measures

Psychological safety came first. Teammates on a team with it feel safe to take a risk around one another. They feel confident that no one will embarrass or punish anyone else for admitting a mistake, asking a question or offering a new idea. The risk they weigh is being seen as ignorant, incompetent, negative or disruptive. Google's survey item is direct: "If I make a mistake on our team, it is not held against me."

The guide uses Amy Edmondson's definition: a shared belief that the team is safe for interpersonal risk taking. One statement she used asks whether a mistake on this team is held against you. Another asks whether members can bring up problems and tough issues. An AI miss is that kind of problem. The model helped, a person let the output travel, and the cost of saying so is social before it is technical.

The guide also records three moves from Edmondson. Treat the work as a learning problem. Acknowledge your own fallibility. Model curiosity and ask a lot of questions. I would use them on the miss itself: what the model produced, what a person checked and what the team now knows. A leader who puts a miss of their own on the table makes the next admission cheaper.

What the silence costs

People on Google teams with a stronger culture were less likely to leave and more likely to use the different ideas in the room. They brought in more revenue, and executives rated them effective twice as often. Those results belong to that study. If admitting the bad number gets someone punished, the next analyst deletes the note, and the team learns nothing about where the model fails.

The guide's worked scene is easy to recognise. A manager known for technical expertise grows intolerant of mistakes and of challenges to his thinking. He publicly tears down an idea the rest of the team thought was worth exploring, and he criticises that person to the wider team. Ideas dry up. His own proposal is later rejected because it lacks creativity. Put an AI draft where that idea was, and the person who would have said the number was wrong stays quiet. The pack goes upstairs clean and false.

Sitting together, consensus, extroversion, individual performance, workload, seniority, team size and tenure showed no significant link to effectiveness in the Google study. A small specialist group and a large one can both hide an AI mistake. The reply the speaker gets is the part that moves.

What to do the next time the number is wrong

1. In the meeting, ask whether anyone checked the model output and found a problem. Ask that before you ask who signed the page.
2. When someone names a miss, repeat what you heard, then move to what will change in the check. Keep the conversation on the work.
3. Tell the room about a miss of your own from the same month, so fallibility is already on the table when a junior person speaks.
4. Keep a short note of the AI misses the team was willing to say in the room. An empty note is a finding about the climate.

A psychological safety workshop can give an intact team a place to practise that reply on a real miss. The habit still has to show up in the meeting where the pack is already open.