Psychological Safety in AI-Augmented Teams: What Leaders Need to Know

Google's Project Aristotle found that psychological safety is the single strongest predictor of team performance, more significant than individual skill or experience. This was a landmark finding. It meant that the quality of team dynamics, specifically, whether people felt safe speaking up, asking questions, and acknowledging uncertainty, mattered more than almost anything else you could optimise for.

But psychological safety is fragile. It's eroded quickly by threat, uncertainty, and lack of leadership vulnerability. And AI adoption, the way it's typically implemented, actively erodes psychological safety.

A February 2026 HBR article by Amy Edmondson found that AI introduction frequently violates the conditions necessary for psychological safety. People feel that decisions are being made by an algorithm they don't understand. Their expertise is being circumvented or devalued. Their autonomy is reduced. Their role is destabilised.

Research published in Nature in May 2025 found that AI adoption in workplace teams negatively impacts psychological safety and can lead to increased depression and anxiety among team members.

So we have a puzzle: the organisations that need high psychological safety most (those undergoing AI transformation) are precisely those most likely to see it eroded.

Why Psychological Safety Matters for AI Success

Experimentation requires safety. You can't learn how to work effectively with a new tool if you're afraid to experiment. You won't try novel applications if you're worried about being blamed for failure.

Escalation requires safety. If an AI system produces a recommendation that your team believes is wrong, they need to feel safe raising it. If they stay silent, you'll implement a bad decision.

Learning requires safety. Becoming genuinely capable with a new tool means making mistakes, asking questions, and surfacing uncertainty. If people are hiding their confusion or hiding failures, they're not learning.

Retention requires safety. People stay in roles where they feel understood, valued, and like they can grow. AI transformation that erodes safety will trigger talent loss precisely when you need to hold onto your best people.

Building and Protecting Psychological Safety in AI Adoption

First, acknowledge the legitimate threat and uncertainty. Don't pretend that AI adoption is only positive or that no roles will change. Instead, be honest: "This is a significant change. I understand it might feel destabilising. I want to be transparent about what's changing, what I don't yet know, and what we're committed to." This honesty creates more safety than false reassurance.

Second, model vulnerability and learning. As a leader, share your own AI learning journey openly. "I don't fully understand how this system works yet. Here's what surprised me. Here's where I made a mistake." This signals that not knowing is acceptable, that learning is valued, and that the leader is human.

Third, create explicit permission to experiment and fail. Many organisations claim to value experimentation but punish failure. Be different. Explicitly tell your team: "I want you to experiment with how we might use this AI tool more effectively. Some experiments will work, some won't. That's the point." Then, when someone tries something and it doesn't work, reinforce this by asking "What did you learn?" not "Why didn't you..."

Fourth, run an explicit safety audit. Use a simple anonymous survey: "On a scale of 1–10, how safe do you feel speaking up with concerns or questions about how we're using AI tools? Why did you give that rating? What would help you feel safer?"

Fifth, be visibly responsive to concerns. When someone raises a concern about an AI system, take it seriously. Investigate. Explain your decision. Being responsive to concerns is what actually builds psychological safety. Dismissing them destroys it.

Sixth, protect autonomy and human judgment. Be explicit about where humans make final decisions and where AI recommends. Protect people's sense that their judgment and expertise still matter.

Measuring Psychological Safety

Create a simple quarterly survey focused on three questions: (1) Do you feel safe expressing concerns about how AI tools are being used? (2) Does your manager listen to your concerns and take them seriously? (3) Do you feel that your expertise and judgment still matter in your role? Track these over time.

Psychological safety isn't a nice cultural addition. It's the foundation upon which successful AI adoption is built.

Try This

Run an anonymous psychological safety survey focused on AI adoption. Ask three questions: (1) How safe do you feel raising concerns about AI tools? (2) Do you feel heard when you express concerns? (3) Do you feel your expertise still matters? Track results quarterly.

Model vulnerability about your own AI learning. In team meetings, share something you're uncertain about regarding AI, or a mistake you made trying to use a new tool.

Create explicit permission to experiment and fail. Tell your team: ‘I want you to try new ways of using these AI tools. Some experiments will work, some won't, that's the point. There's no blame for experiments that don't work.’ Then, when an experiment fails, ask ‘What did you learn?’ and mean it.

The Paradox at the Heart of AI Adoption

The EY Agentic AI Workplace Survey (2025) captured the contradictory reality of how most employees are experiencing AI adoption: 84% are enthusiastic about AI agents, and 56% simultaneously worry about job security. These are not two different groups of people. They are the same person, in the same moment, holding both realities at once.

Leaders who receive this information and respond by emphasising the 84% are making a mistake. Leaders who respond by building a safe space for the 56% are building the foundation for successful adoption.

The reason is grounded in Edmondson's (2019) research on psychological safety: the belief that you will not be punished or humiliated for speaking up, making mistakes, asking questions, or voicing concerns. Edmondson's work shows that psychological safety is the single strongest predictor of team learning behaviour. Teams with high psychological safety experiment more, recover from failure faster, and share knowledge more readily.

The MIT Technology Review (2025) reported that 83% of executives confirm psychological safety measurably improves the success of AI initiatives. This is not a soft outcome. It is a direct driver of adoption success.

What Blocks Connection in Practice

The barriers to psychological safety during AI adoption are specific and predictable.

Performance pressure accelerates before capability develops. When organisations roll out AI tools, expectations of productivity improvement often precede the actual development of skill and confidence. People who are still learning to use a tool are simultaneously expected to produce at higher levels with it. This creates the conditions for hiding difficulties rather than surfacing them.

The knowledge gap becomes a status threat. In most professional environments, being seen as competent matters. When AI introduces a new dimension where some people are significantly more fluent than others, the knowledge gap becomes a status threat for those behind. Rather than asking questions and learning, people perform confidence they do not have.

Managers project certainty they do not feel. The DDI Global Leadership Forecast (2025) found that 71% of leaders are under increased stress, with 40% considering leaving their roles. Many of the managers I work with are navigating their own uncertainty about AI while simultaneously being expected to project confidence to their teams. This is not sustainable, and it is counterproductive. When a leader pretends to understand something they do not, they are modelling that pretending is safer than admitting.

What Connection Requires in Practice

Building Connection in an augmented team requires four specific leadership practices.

Hold honest team conversations. Not briefings. Not presentations about the AI strategy. Conversations where the leader names both the opportunities and the anxieties, and invites the team to do the same. The format matters less than the honesty. What makes these conversations effective is when the leader goes first and does not rush to reassure.

Approach resistance with curiosity. When a team member is not engaging with a new tool, the least productive response is pressure or assumption. The more productive response is a one-to-one conversation that opens with genuine curiosity: help me understand what you're finding difficult. In many cases, the resistance reveals something important about the tool, the workflow, or the team's capacity that would otherwise stay hidden.

Establish agreements about disagreement. One of the specific challenges of augmented working is that team members will sometimes disagree with AI recommendations. Without an explicit agreement about how to handle those disagreements, people either defer to the tool against their better judgement or override it without flagging the concern. Neither is a good outcome. A team-level agreement about how decisions get made when technology recommendations and human judgement differ creates the clarity that allows people to act confidently.

Make safety visible. The final practice is acting visibly on what the team says. If a team member raises a concern about an AI tool and nothing changes, the implicit message is that raising concerns here is not worth the risk. If a leader hears a concern, acknowledges it, investigates it, and reports back, the implicit message is that speaking up has value. This is how psychological safety gets built: not through declarations, but through repeated demonstrations.

The Leader's Own Connection

One dimension of Connection that receives less attention is the leader's own experience. The same research that identifies psychological safety as the protective factor in AI adoption also highlights that leaders themselves are under significant strain. Building Connection is not only about creating conditions for others. It requires leaders to attend to their own wellbeing and to be honest about their own learning edges.

A leader who is pretending to understand AI fully while privately feeling overwhelmed is a leader whose team will eventually sense the gap between what they say and how they behave. Authenticity in this context is not just a value. It is the condition for the trust that Connection requires.

The most effective leaders I encounter in this work are those who have found a way to hold the tension: being genuinely open about uncertainty while remaining a steady presence for their team. That combination, honesty about not knowing alongside confidence in the capacity to learn, is the foundation of Connection in an AI-augmented team.