Two sentences decide whether an organization actually learns to use AI: "I used AI to do this," and "the AI got this wrong, and I missed it." Where people can say both without fear, use becomes visible, errors surface early, and good practice spreads. Where either sentence is dangerous, the reliance goes underground and the mistakes stay buried until they detonate. The deciding factor is not the tool. It is psychological safety.
The concept has a precise origin. Amy Edmondson defined team psychological safety in 1999 as a shared belief that the team is safe for interpersonal risk-taking, and showed across 51 work teams that safety drives performance by enabling learning behavior: people ask questions, admit errors, and seek feedback only when doing so will not be held against them. The 136-sample meta-analysis by Frazier and colleagues, spanning more than 22,000 individuals, later confirmed the pattern, finding psychological safety strongly associated with learning behavior and information sharing. Those two behaviors, learning and sharing, are exactly what AI adoption requires, and exactly what fear suppresses.
The hidden reliance and the hidden mistake
AI creates two distinct things people are tempted to hide. The first is use itself. In a culture where speed is assumed to prove competence, admitting a machine helped can feel like confessing weakness, or like volunteering that your job could be done without you. So people quietly rely on the tool and take private credit, and the organization never learns who is using what or how. Deloitte's 2025 research found six in ten workers already treat AI as a coworker; a workplace that makes people hide that relationship cannot manage it.
The second thing people hide is the mistake. Because AI fails in confident, fluent prose, its errors slip through, and the person who shipped a wrong AI-assisted answer has strong reasons to stay quiet: it exposes both the tool's failure and their own missed check. But a buried AI error is the most expensive kind, because it looks authoritative and propagates. The only defense is a culture where flagging "this came from AI and I think it's wrong" earns thanks rather than blame.
Safety is the precondition for learning
Ricardo Semler has told how Semco built the safety structurally rather than hoping for it. Everyone rated their boss anonymously twice a year, and a manager who scored badly twice running lost the role, Ricardo included; teams could veto a manager outright. The point was to make telling the truth upward safe by design, so that power had to be earned from the people it affected. That is psychological safety engineered into the structure, not left to individual courage. An organization serious about honest AI use needs the same instinct: make it structurally safe to say what you actually did and what actually went wrong.
Without that, the numbers curdle. Gallup's 2025 survey found 65% of employees feel AI helps their productivity while only 12% strongly agree it has changed how work is done. Part of that gap is hidden practice, real use that never becomes shared, improvable, governable practice because admitting it feels unsafe. Fear converts a potential organizational capability into a collection of private, invisible habits.
Building the safety on purpose
The practical moves are concrete. Leaders go first, showing their own AI use in the open, including the prompt that failed, which signals that reliance is normal and fallible. Admitting an AI-assisted error is treated as good catching, not misconduct, because the alternative is people hiding the next one. Disclosure is made routine and low-stakes, so "drafted with AI, checked these three things" becomes an ordinary sentence rather than a confession. And, as at Semco, the safety is built into structure and incentive rather than left to hope, because courage does not scale and design does.
The organizations that learn to use AI well will not be the ones with the best tools. Those are available to everyone. They will be the ones where people are not afraid to say what they did with the tool and what it got wrong, because only in the open can use be improved, errors be caught, and the good prompt travel from the person who found it to everyone who needs it. Safety is not a soft complement to AI adoption. It is the precondition for it.
Sources
- Edmondson (1999). Psychological Safety and Learning Behavior in Work Teams (Administrative Science Quarterly). Defined team psychological safety; across 51 teams, safety drives performance by enabling learning behavior. https://journals.sagepub.com/doi/10.2307/2666999
- Frazier, Fainshmidt, Klinger, Pezeshkan & Vracheva (2017). Psychological Safety: A Meta-Analytic Review (Personnel Psychology). 136 samples, 22,000+ individuals; safety strongly linked to learning behavior and information sharing. https://doi.org/10.1111/peps.12183
- Deloitte (2025). 2025 Global Human Capital Trends. Six in ten workers already think of AI as a coworker. https://www.deloitte.com/us/en/insights/topics/talent/human-capital-trends/2025.html
- Gallup (2025). State of the Global Workplace. 65% of employees report a positive productivity impact from AI, but only 12% strongly agree it has changed how work gets done. https://www.gallup.com/workplace/659610/state-global-workplace.aspx