A recent whitepaper from Leuphana University Lüneburg calls psychological safety a central precondition for AI adoption: only where it is high do mistakes and uncertainty in working with AI become something people can talk about. Where companies measure it, they mostly use a questionnaire in which employees rate statements on a scale. The scale shows whether something is missing; where it is missing and why is not in any of the ticked boxes.
What is psychological safety? Psychological safety is a team's shared belief that it is safe to take interpersonal risks there: to admit mistakes, raise problems, ask for help or disagree without being embarrassed or punished for it. The definition in common use today comes from Harvard professor Amy Edmondson (1999), who studied how psychological safety relates to learning behaviour in teams. In AI adoption, it helps decide whether employees discuss their uncertainty and mistakes with AI openly or keep them to themselves.
Edgar Schein and Warren Bennis were already using the term in 1965. The scale that many people associate with it today, by contrast, grew out of conversations: Edmondson developed it at an office furniture manufacturer from observations and interviews, and only then handed out questionnaires.
Edmondson's scale sits between two rounds of interviews
Her study in the journal Administrative Science Quarterly ran in three phases. First, over two four-day visits, Edmondson observed eight team meetings and conducted 17 interviews of 45 to 60 minutes each; from these conversations she wrote the statements for her questionnaire so that they would make sense in the language of this particular company. She then surveyed 51 teams in writing.
Finally, she went back to seven teams that had shown particularly high or particularly low learning behaviour in the questionnaire, and interviewed and observed them again. The wording came from the first round of interviews; in the second, Edmondson asked what lay behind the scores.
The scale itself consists of seven statements that team members rate on a seven-point scale. One of them reads: "If you make a mistake on this team, it is often held against you." Someone who agrees is saying something important, but not which mistake, held against them by whom, and since when.
AI adoption makes the gap visible
The whitepaper "Die KI-Transformation verantwortungsvoll gestalten" by Simon Berkler, Karoline Rütter and Sevda Helpap (June 2026) draws on 13 qualitative interviews with executives, organisational development practitioners, work scientists and technology experts. According to this analysis, organisations with high psychological safety deal more constructively with identity questions such as "What will my contribution be in future?" (translated from German). The fear of making mistakes, on the other hand, holds back AI adoption massively.
The authors therefore call for spaces in which fears of job loss and identity crises can be discussed openly. Anita Liebe of the Lufthansa Innovation Hub describes in the same paper what happens without such spaces: in her work she sees, every day, "a quiet sense of insecurity that is rarely voiced, because there is no room for it".
Edmondson's scale does not ask about the worry whether your own work will still be needed in two years. It does ask whether you can voice that worry in your team – just not what the worry is about or who has it.
What a pulse survey shows and what it does not
I have designed a pulse survey like this myself. In a concept for an AI rollout, I planned a weekly survey with three questions on psychological safety on a scale from 1 to 5, alongside a channel where concerns can be reported anonymously. I still think that is right, because a pulse takes moments to fill in and makes it visible when a team slips after a rollout.
What it does not show is the reason. If a team's score drops sharply, every explanation remains open: the new AI assistant, a manager who picks apart AI results in front of the whole group, a rumour about job cuts, or the worry that one's own usage is being logged. My concept lists one of these causes as a separate anti-pattern: AI shaming, meaning exposing people for their AI results.
The comment box under the pulse does not close this gap, because it does not ask follow-up questions. There is also a contradiction in the method: a questionnaire about openness assumes that people answer openly. The more delicate a topic, the more the feeling of being evaluated colours the answers; I have collected the studies on this in an article on how people answer an AI.
A conversation is about situations
A conversation deals with what people have experienced. Each statement on the scale can be turned into a question about a concrete situation, and the follow-up question brings out what lies behind it.
| What the scale asks | What gets asked in a conversation |
|---|---|
| If you make a mistake on this team, it is often held against you. | When did an AI result last go wrong in the team, and what happened afterwards? |
| Members of this team are able to bring up problems and tough issues. | Is there a concern about the AI rollout that nobody has said out loud so far? |
| It is difficult to ask other members of this team for help. | Who do people turn to when the AI assistant produces nonsense, and how does that feel? |
| It is safe to take a risk on this team. | What has someone tried with AI and not told anyone about, and why not? |
The answers to questions like these are longer, messier and harder to analyse than an average score. But they contain what is missing after a poor pulse result: who is affected, at which point in their working day, and what the people themselves see as the cause. That gives a manager something they can actually change.
Listening at scale used to be too expensive
Methodologically, Edmondson's approach would be the right one: observe, interview, measure and then interview again. Hardly any company can pay for it, because with 300 employees and one hour per conversation, you have 300 hours of interviews before the analysis even begins. For most teams, the alternative to the perfect interview is therefore no interview at all.
AI-led interviews change this calculation. Every person gets their own conversation, at the same time as everyone else, with follow-up questions that respond to their answers. At QUALLEE we run these conversations as text and without an avatar, participants know that an AI is asking, and no human listens in during the conversation; what gets analysed are patterns across all the conversations in a survey.
This conversation format does not create psychological safety. Anyone who has experienced that speaking openly had consequences will also answer an AI cautiously, and rightly so. Before a company starts listening, it therefore has to state clearly what happens to the answers and who sees the analysis; in my concept, over-monitoring sits next to AI shaming on the list of anti-patterns.
Whether AI intensifies work, as described in the AI burnout paradox, or whether a culture drifts away from its AI goals, as with culture dissonance, are separate questions. Psychological safety decides whether people talk about either of them.
After the next pulse
After her questionnaire, Edmondson went back to the teams whose scores stood out. The same can be done after a pulse, with more teams and more people per team than a research group on site can manage. Discussing a falling score only among the leadership team leaves out the second half of her method.
Frequently asked questions
Who coined the term psychological safety?
Edgar Schein and Warren Bennis used the term in 1965 in research on organisational change. The team-level definition in common use today comes from Amy Edmondson, who published a study of 51 teams at a manufacturing company in 1999. She describes psychological safety as a shared belief that a team is safe for interpersonal risk-taking.
How do you measure psychological safety?
Many measurements are based on Edmondson's scale of seven statements, such as whether members of the team are able to bring up problems and tough issues; team members rate them on a seven-point scale. The scale shows level and trend. Where things get stuck and why is shown by interviews with follow-up questions, like the ones Edmondson conducted before and after the questionnaire in her own study.
Why does psychological safety matter for AI adoption?
According to the whitepaper "Die KI-Transformation verantwortungsvoll gestalten" (Leuphana University Lüneburg, 2026), it is a central precondition for mistakes and uncertainty in working with AI to become discussable. The authors find that the fear of making mistakes holds back AI adoption massively, and that teams with high psychological safety deal more constructively with identity questions about their own role.
How can psychological safety be strengthened during AI adoption?
Helpful measures are a channel where concerns can be reported anonymously, the rule that criticism is aimed at the AI result and never at the person, and a few openly explained usage metrics instead of surveillance. Before any survey, it should be clear who sees the answers and what happens to them. Conversations with follow-up questions show where a team needs to adjust course.
Can AI-led interviews measure psychological safety?
Not as a single number, but they show where it is missing and why. An AI interview asks about concrete situations, follows up and can be offered to all employees instead of a small sample. It cannot create psychological safety; that takes managers who respond to what they hear, and clear commitments about what happens to the answers.
Sources
- Edmondson, A. (1999): Psychological Safety and Learning Behavior in Work Teams. Administrative Science Quarterly, 44(2), 350-383.
- Berkler, S., Rütter, K. & Helpap, S. (2026): Die KI-Transformation verantwortungsvoll gestalten. Wie Künstliche Intelligenz Organisationen verändert – und warum Organisationsentwicklung dabei eine Schlüsselrolle spielt (Shaping the AI transformation responsibly: how artificial intelligence changes organisations and why organisational development plays a key role). Whitepaper, version 1.0, June 2026, Leuphana University Lüneburg, TheDive, Inspiring Minds. Licence CC BY-SA 4.0.
Try it yourself
You can experience what a conversation feels like in which an AI asks follow-up questions in 20 minutes, right in your browser, without logging in.
