Picture this: One of your analysts uses an AI tool to prepare a client summary on a Thursday afternoon. The tool invents a number. She catches it, fixes it, and delivers the report on time. She tells no one. The next week, a colleague pastes internal pricing data into a public chatbot to speed up a forecast, then deletes the chat. Nobody files a report. Nobody asks. From where you sit, AI adoption looks smooth. Licenses are up, training is complete, and no incidents have reached your inbox.

That quiet is the problem. A September 2026 study from GoTo found that 83% of workers worry they could be blamed or even fired for an AI mistake, and 31% say that fear has made them reluctant to use the tools at all. Your people may not lack AI skills; they may lack permission to be honest about AI mistakes at work. If you manage managers or lead a team yourself, this is the part of AI adoption no software vendor can fix for you. It comes down to psychological safety—and it lives or dies with how you react on a bad day.

Here is what the research says—and what I would do if I were in your seat.

The Numbers Behind the Silence

Start with GoTo’s report, The Pulse of Work in 2026, which surveyed 2,500 knowledge workers and IT leaders across 10 countries. Beyond the 83% who fear blame, 55% admitted they do not always double-check AI output, even on high-stakes tasks. Forty-three percent said they had used AI output they suspected was low quality or contained fabricated information. Most striking, 14% said they reported an AI error to a manager and were told to keep quiet about it.

That’s not a one-off. A 2025 KPMG and University of Melbourne survey found that 57% of employees hid their AI use, 66% relied on output without checking it properly, and 56% said they had made mistakes at work because of AI. Glean’s Work AI Index 2026 found that 69% of AI users admit to shipping AI-generated work they haven’t verified or can’t fully stand behind.

Meanwhile, confidence is sliding. ManpowerGroup’s 2026 Global Talent Barometer, which surveyed more than 13,000 workers in 19 countries, found that AI use rose 13% in 2025 while worker confidence in using it fell 18%. In the UK, Ipsos Karian and Box found that only 38% of employees believe they would be supported, rather than punished, for missing an AI error. Among junior staff, that figure falls to about 29%.

Put those together and you get a picture most leaders don’t see on their dashboards. People are using AI more, trusting it less, checking it inconsistently, and hiding the results.

Why Smart People Keep Quiet

It is tempting to read those numbers as a training gap: Give people a better course, and the problem goes away. I do not buy it, and neither does the research.

Consider the situation from your team’s perspective. Employees are being pushed to use AI. GoTo found that 60% feel pressure to use AI to boost productivity, while 47% say AI use now appears in their performance reviews. In addition, 42% said they would rather risk making errors with AI than appear slow or replaceable. They are told to move fast, evaluated on their use of the tools, and afraid of what happens when those tools fail. Under those conditions, hiding a mistake is not a character flaw. It is a rational response.

Research on employee silence supports this conclusion. Elizabeth Morrison, Kelly See, and Caitlin Pan found that people who feel powerless are more likely to stay quiet, and that the perceived openness of the person they would report to can determine whether they speak up. In plain terms, your people are reading you. They are watching what happened to the last person who raised a problem—and doing the math.

There’s a social layer too. Slack’s Workforce Lab found workers worry colleagues will see AI-assisted work as less authentic. Deloitte found 23% of UK generative AI users perceive a stigma around using the tools. Nobody wants to be the person who “cheated” or the one who got burned by a chatbot.

What Shadow AI Tells You

Shadow AI is the name for employees using AI tools their employer doesn’t know about or hasn’t approved. Microsoft’s 2024 Work Trend Index found 78% of AI users brought their own tools to work, and 52% were reluctant to admit using AI for their most important tasks. Deloitte found 31% of UK generative AI users work with it without their employer’s knowledge, and that UK workers spend an estimated £958 million a year of their own money on AI tools for work.

Most leaders treat shadow AI as a security problem—and it is. But it is also a message. When someone pays out of pocket for a tool and hides it, they are telling you two things: The approved tools are not doing the job, and employees do not trust you enough to say so. If your first move is to block sites and send a stern memo, you will push the behavior further underground. You will also lose the information you need most: how the work is actually getting done.

Your Reaction Is the Policy

Every middle manager should hear this: Your company’s AI policy is defined by what happens the first time someone on your team admits a mistake.

Gallup’s research this year makes the manager’s role plain. In a survey of 102 CHROs, 99% said AI matters to their strategy, but half lacked confidence in their managers’ ability to guide employees through it. Gallup’s Q1 2026 study of 23,717 US employees found that where people strongly agree their manager champions AI, 33% say it has transformed how work gets done. For everyone else, that number is 4%. Where managers actively support their team’s AI use, 31% say culture improved over the past year, compared with 21% elsewhere.

Psychological safety matters at the outset. A February 2026 study of 2,257 employees at a global consulting firm found that psychological safety predicted whether people adopted AI tools in the first place, and the relationship held across roles, experience levels, and regions. It did not predict how often employees used the tools after they began. Safety gets people through the door. Without it, many never walk in.

When someone tells you they caught an AI error—or, worse, missed one—your response sets the rule for everyone watching. Thank them first. Address the immediate risk. Ask what happened without asking whose fault it was. Then explain to them and the team what will change as a result. If you cannot do those four things calmly, no policy document will save you.

Separate Honest Mistakes from Bad Calls

None of this means anything goes. Accountability still matters. The trick is being clear about what you’re holding people accountable for.

Behavioral scientist Gleb Tsipursky put it well in an August 2026 piece for UNLEASH: deliberate misuse, fraud, and repeated disregard for clear rules need consequences, while good-faith experimentation, confusion, and promptly reported mistakes need learning. Treat both the same way and you kill candor.

In practice, draw the line in advance and state it clearly. Tell your team: If you use an approved tool, follow the data rules, and tell me quickly when something goes wrong, you will not be punished for the mistake. If you paste client data into a public chatbot after we have discussed the rule three times, that is a different conversation. Most people can accept that distinction. What they cannot accept is having to guess.

Also watch for the opposite failure. GoTo found 17% of workers have already blamed AI for a mistake they made themselves. If “the AI did it” becomes a free pass, you’ve swapped silence for excuses. Ownership stays with the person who shipped the work, and that’s fair, as long as they had the time and support to check it.

Make Near Misses Normal

Safety-critical industries figured this out decades ago. Aviation and healthcare treat near misses as gold because they show you where the next accident is coming from. OSHA recommends simple ways to report hazards and close calls, including anonymous options and protection from retaliation. You can borrow that approach for AI.

Create a short, low-drama way for people to log AI near misses: an invented citation caught before publication, confidential data nearly pasted into the wrong tool, or a recommendation overridden after a second look. Ask what they were trying to do, which tool they used, what went wrong, and what prevented harm. Keep the focus on the workflow, not the person.

Then talk about these openly in team meetings, the way good engineering teams run blameless postmortems. When a near miss leads to a better checklist or a clearer data rule, say so. People report more when they see reports turn into fixes.

Go First

If you want your team to admit AI mistakes, admit yours. A Microsoft People Science survey found that where managers modeled AI use themselves, rather than just encouraging it, employees reported a 17-point lift in the value they got from AI and a 22-point rise in critical thinking about how they used it.

Use the tools in front of your team. Share a prompt that worked and one that failed. Tell them about the time AI gave you a wrong answer you almost forwarded to your boss. That two-minute story will do more for psychological safety than a dozen slides about a “culture of experimentation.” Senior leaders can do the same in town halls and skip-level meetings. When a vice president says, “I got this wrong last week, and here is what I learned,” every manager below receives permission to do the same.

Measure Candor, Not Just Usage

Most AI dashboards track licenses, logins, and training completion. Those numbers tell you who has access. They don’t tell you whether people trust you enough to tell you the truth.

Track a different set of indicators: the number of AI near misses reported each month, how quickly problems surface, and whether the same mistakes recur. Add two pulse questions: Do you know which tools and data are approved? Would you expect fair treatment if you reported a good-faith AI mistake? Culture Amp found that the share of organizations measuring employee belief in company success has fallen 19 points over five years—a sign that many companies have stopped asking uncomfortable questions. Do not be one of them.

One counterintuitive point. When you start doing this well, reports will go up. That’s good news. It means you’re finally seeing what was always there. If AI use is climbing on your team and reports sit at zero, be suspicious.

Conclusion

We spend considerable energy deciding which AI tools to buy and how to train people to use them. That work matters, but this year’s data points to a larger barrier: fear. Eighty-three percent of workers worry about being blamed for AI errors. Fewer than four in ten UK employees believe they would be supported after missing one. And one in seven workers who spoke up was told to stay quiet.

You cannot fix that with a memo. You fix it one conversation at a time, beginning with how you respond when someone walks into your office and says, “I think the AI got this wrong.” Thank them. Fix it. Learn from it. Tell the team. Repeat that response, and you will build something your competitors cannot buy: a team that tells you the truth while there is still time to act.

Leadership in the AI era isn’t about knowing every tool. It’s about building the kind of room where people feel safe saying what they don’t know. That’s always been the job. AI just raised the stakes.