Why 73% of Workers Hide AI Mistakes at Work in 2026

Workers are watching AI break things at work and choosing not to say a word. A new study from Radical Candor, covered by HR Dive on May 12 2026, found that 73% of employees have seen AI-related inconsistencies on the job while 60% report being afraid to speak up about workplace concerns generally. That gap, between what people see and what they say, is the real AI risk story of 2026. It is not that the models hallucinate. It is that the workers who notice have no safe channel to flag it, and leadership has no idea how exposed they actually are. The silence is the story.

What does the new HR Dive study reveal about AI silence?

The Radical Candor research, summarized by HR Dive, paints a picture of a workforce that has gone quiet at exactly the moment its feedback matters most. Seventy-three percent of respondents reported AI-related inconsistencies in their daily work, ranging from wrong outputs to broken automations to tools that contradict company policy. Half of those workers said negative feedback about AI mistakes was addressed only sometimes or rarely when they did try to raise it. That is a feedback loop that is structurally broken on both ends. Workers see the problem. Workers report the problem some of the time. The problem mostly does not get fixed.

The broader workplace silence problem is even more alarming. Sixty-seven percent of HR professionals said they witnessed workforce silence about concerns often or very often. And 46% of executives flagged the lack of honest feedback as one of their top organizational concerns, meaning the people running these companies know something is wrong, they just cannot get it out of the staff who would actually know what it is. There is a near-universal acknowledgment that the channels are clogged.

Radical Candor co-founder Kim Scott put the dynamic plainly in the HR Dive piece. "When an untrained manager hears criticism from an employee, too often they punish the messenger rather than reward the candor. That's where psychological safety gets destroyed." She is describing a feedback loop that gets sharper every time a worker tries to raise an AI concern and gets shut down, ignored, or quietly reassigned. After enough rounds, everyone learns the lesson: keep your head down, let the tool be wrong, hope it does not become your problem.

That dynamic does not exist in a vacuum. It builds on the American office silence era that has been quietly metastasizing since 2024, where workers have stopped flagging anything that might mark them as difficult, slow, or insufficiently bullish on the company's stated priorities.

Why are workers staying quiet about AI errors?

The answer is incentive design. When leadership is publicly committed to AI adoption, often in the same quarter as workforce reductions, no rational employee wants to be the person telling the boss the new system is broken. The signal sent in those moments is clear: AI is the path forward, and anyone who pushes back is either resistant to change or, worse, redundant. Workers can read corporate body language. They know what the safe answer is.

Aura Intelligence research cited in the same HR Dive coverage found trust in leadership has dropped sharply in industries running layoffs or sudden return-to-office mandates. Those are the same industries pushing AI hardest. Workers do the math. Flagging an AI error means publicly questioning a corporate priority while standing on uncertain ground, and the upside is essentially zero. Most pick silence. The math is not complicated and it is not irrational.

There is also a training gap. Seventy percent of managers told Radical Candor they were never taught to give or receive feedback before stepping into the role. That means even managers who would, in theory, welcome AI-error reporting often do not know how to receive it without sounding defensive, dismissive, or threatened. So the message workers learn is not that managers want the truth. It is that managers cannot handle the truth without making it the worker's problem.

That silence stacks on top of an AI adoption gap between high earners and frontline workers, where the people most likely to spot ground-level AI failures are the least likely to be in the room when AI strategy gets set. Meanwhile executives quietly fear losing their jobs over failed AI adoption, which only sharpens the pressure to declare AI rollouts a success regardless of what the data on the ground actually shows. The result is a top-down narrative that resists bottom-up correction.

What does that silence actually cost?

The cost lands in three places. First, operational risk. When a third of the workforce sees AI inconsistencies and only half of those reports get a real response, errors compound. Bad outputs feed downstream systems. Customer-facing decisions get made on faulty automation. The mistakes stop being a tool problem and start being a company-wide quality problem, and the version that finally reaches leadership has been laundered through three layers of polite framing until it sounds like a minor process tweak.

Second, retention. Workers who feel they cannot raise concerns about a tool they use every day eventually conclude they cannot raise concerns about anything. The mental ledger fills up. That accelerates what we have already documented around the corporate ick factor and worker retention, where a single moment of feeling dismissed becomes the reason a high performer starts replying to recruiter messages. Silence is a leading indicator of attrition, not a sign of stability. By the time the exit interviews catch the pattern, the best people have already left.

Third, productivity. The workslop crisis we covered earlier in 2026 documented exactly what happens when AI outputs are accepted without scrutiny: low-quality work piles up, downstream teams burn hours cleaning it, and the original productivity gain inverts into a productivity tax. Silence about AI errors is what lets workslop scale. Without a feedback channel, there is no friction between bad output and the next step in the workflow, so the bad output just keeps moving.

The pattern shows up in surveys too. Earlier this year workers said AI does more harm than good while employers cut AI training budgets. That mismatch is exactly the kind of signal a healthy feedback culture would surface early. Instead it shows up two years later as a productivity miss in an earnings call, and the executives reading the deck genuinely cannot explain why.

How does psychological safety change AI outcomes?

Psychological safety is the variable that decides whether a workforce flags AI errors early or absorbs them quietly until they become a crisis. The research on this is unusually consistent. Teams where workers feel safe raising concerns catch problems faster, iterate on AI rollouts more effectively, and retain employees longer. We covered the underlying mechanics in psychological safety and speak-up culture and again in the Google research connecting psychological safety to retention.

The AI dimension sharpens the stakes. AI systems learn from feedback or they degrade. A workforce that cannot tell its managers the new tool is producing nonsense outputs is a workforce that cannot help refine the model, the workflow, or the deployment plan. The company ends up with worse AI not better, while the leadership team congratulates itself on hitting an adoption KPI that measures whether the tool was deployed, not whether it works.

The contrast is visible in how workers respond when they feel safe to push back. Earlier coverage of workers actively rejecting AI as a coworker in 2026 showed that resistance is often a signal of unaddressed concerns, not stubbornness. When workers have a real channel to raise issues, that resistance often converts into engagement. When they do not, it converts into quiet quitting or a resume update. The same energy, two very different outcomes, decided almost entirely by whether anyone listens.

There is also a story about who speaks up at all. The Pinterest engineer who broke silence on layoff tracking is a clear example of what happens when workers stop trusting internal channels: the truth goes public, often on a platform the company cannot moderate. AI-error silence works the same way. Suppress it internally long enough and it surfaces externally, usually attached to a regulatory complaint, a viral thread, or a journalist's voicemail.

Worth noting that the dynamic is not universal across geographies. Workplace culture in China is shifting on respect and power in ways that put real pressure on management styles built on suppressing feedback, which is a useful contrast for US managers still treating candor as a threat to authority rather than as the input that protects them from blowing up an AI rollout in public.

What can workers and managers do this quarter?

For workers, the move is documentation. If you are seeing AI errors, write them down. Note the system, the input, the wrong output, and the downstream impact. Raise it in a setting where there is a paper trail, like a project channel or an email follow-up, rather than a hallway conversation. That protects you and forces the issue into a record someone has to eventually answer for. If your manager cannot handle the conversation, the documentation gives you leverage when you escalate or, eventually, when you decide to leave.

Watch for the warning signs too. The toxic workplace red flags we have catalogued apply directly here. If raising a basic AI concern gets you sidelined, that is the company telling you exactly what it values and exactly how it will treat you the next time the stakes are higher. Believe it the first time. People rarely regret leaving a workplace where the truth was punished.

For managers, the work is reversing the punish-the-messenger instinct. Scott's framing in the Radical Candor coverage is the right starting point: criticism from an employee is data, not insubordination. The fix is procedural. Build a recurring agenda item for AI-error reporting. Reward the first person who raises an issue, publicly and visibly. Treat a clean week with zero reported errors as a signal something is wrong, not right. If nobody is reporting errors, it is not because the AI is working. It is because the channel is dead.

It is also worth being honest about who is good at AI and who is not. Anthropic data showing the 10-percent performance divide between AI power users and newcomers is a reminder that the people most likely to spot subtle AI failures are the experienced users, and the people most likely to miss them are the recent adopters. A feedback system that only hears from the loud or the senior misses most of the signal. Build channels that pull from the people actually using the tool the most, not just the ones with the strongest opinions about it.

The deeper shift is cultural. Companies that treat AI rollouts as faith-based commitments will keep getting silence and bad outputs. Companies that treat them as iterative engineering problems, where errors are expected and reporting them is rewarded, will end the year with better tools and a workforce that still trusts them. The choice is being made right now, one suppressed concern at a time, in every team meeting where someone almost spoke up and then decided not to.

People also asked

Q: What percentage of workers see AI errors at work in 2026?

A: According to the Radical Candor study reported by HR Dive, 73% of workers said they have seen AI-related inconsistencies at work, including wrong outputs, broken automations, and tools contradicting policy. Half said those errors were addressed only sometimes or rarely when raised. The number is high because AI is now embedded in routine workflows, so most workers encounter at least one tool that occasionally produces wrong or inconsistent results in a typical week.

Q: Why are employees afraid to speak up about AI mistakes?

A: The Radical Candor research found 60% of employees are afraid to speak up about workplace concerns generally, and AI specifically carries extra risk because leadership has often publicly tied AI adoption to growth or cost goals. Workers reading those signals correctly conclude that flagging AI errors looks like resistance to a corporate priority. Add the fact that 70% of managers were never trained to receive feedback well, and the silence makes rational sense from the worker's side of the table.

Q: How does psychological safety affect AI adoption?

A: Psychological safety is the single biggest variable in whether AI rollouts succeed or quietly fail. Teams where workers feel safe raising errors catch problems early, refine the tools, and retain employees. Teams without it absorb AI failures silently until they surface as missed targets, regulatory issues, or attrition spikes. The HR Dive coverage frames it directly: untrained managers punish the messenger, which destroys the safety needed to make AI work in practice.

Future-proof your career. Find roles where workers' voices actually matter. Start your job search on Metaintro.