Why Employees Aren’t Transparent About Their AI Usage

A 31-year-old physician who uses AI multiple times a day at work told us about a personal prompting template he had built for DoximityGPT, the HIPAA-compliant AI tool his organization had approved. The template, he said, “produces astoundingly good results.” His colleagues, using the same tool, were struggling—they had told him so. He believed his template would help them.

Yet he hadn’t shared it.

This is a scenario playing out at organizations across the globe. A global KPMG and University of Melbourne study of more than 48,000 respondents found that 57% of employees admitted to hiding their use of AI at work. Concealed use is one problem. What people are actually doing and learning in private with their prompt sequences, chained tools, and successful iterations matters more.

To be sure, knowledge hoarding has always existed in organizations. The research on organizational silence—why employees withhold information, concerns, and ideas—is well established. But that work has largely focused on the suppression of problems: bad news, ethical concerns, operational risks. What AI introduces is the suppression of solutions. When individually generated workflow innovations can cut a three-hour task to 20 minutes, and when those innovations are easy to conceal, silence becomes economically consequential in a way it hasn’t been before.

Productivity gains used to scale by default. They were embedded in shared systems, standardized processes, and formal tools that spread improvements structurally. With AI, many of the most valuable gains come from individual experimentation—an employee who discovers the prompt sequence that produces client-ready output in a fraction of the time, or a workaround that bypasses a bottleneck the official process hasn’t caught up to. That knowledge is portable, easy to refine in private—and easy to keep to yourself.

Most organizations have responded to “shadow AI” use with more governance such as usage policies, approved tool lists and monitoring. These are reasonable steps. But they may be targeting the wrong problem. Our research suggests that a less visible driver of AI knowledge hiding is organizational trust—what employees expect will happen to their work, their workload, and their standing when they make their AI methods visible.

When employees don’t trust their leaders or organizations, they withhold information, divert energy from performance to self-protection, and disengage from team learning processes that turn individual insight into collective capability. AI raises the stakes of that dynamic. In one survey, employees reported hiding their AI use for a familiar set of reasons: desire to maintain a competitive edge over peers, worry that disclosed productivity will mean more work, fear their job could be cut once the method is documented, concerns about violating company AI policies, and reluctance to have their ability questioned. These reasons may contribute to a separate finding: in an Anthropic study, 69% of professionals mentioned social stigma around AI use at work. Agentic coding tools like Claude Code and OpenAI’s Codex make it easier than ever to build valuable tools independently.

While recent conversations have focused on easing employee anxiety around AI adoption, the more underappreciated challenge is how AI stress-tests culture. Leaders who focus solely on adoption rates may miss the more important question: do employees feel safe—and incentivized—to share what they’ve discovered on their own?

When Experimentation Looks Like Rule-Breaking

AI knowledge hiding persists partly because organizations are mixing up two types of failure. As Harvard Business School professor Amy Edmondson has argued, one category of failure is blameworthy deviance: people ignoring rules or cutting corners in ways that hurt the organization. At the other end of the spectrum is praiseworthy exploratory testing: people experimenting at the edge of what is known, producing failures that generate valuable learning. When organizations confuse the second for the first, they punish exactly the behavior they most need to encourage.

Employees are using AI to experiment on novel tasks, iterate on prompts, and build idiosyncratic workflows regardless of whether their organization has formal AI policies or approved tools—classic exploratory testing. Whether employees keep the results of that experimentation private depends largely on whether they trust their organization with what they learned. Indeed, a Stanford Digital Economy Lab study of 51 enterprise AI deployments found that 77% of the hardest challenges in AI adoption had nothing to do with the technology itself. One of the factors the researchers named was the challenge of earning trust from skeptical teams. So how can organizations fix this trust gap?

What Our Research Shows

To answer that question empirically, we surveyed 604 U.S.-based employees, all of whom reported using AI at work daily or multiple times per day. We also interviewed professionals, from analysts to CEOs, about how they handle the issue.

Nearly one in three survey respondents (30.3%) said they had intentionally withheld AI-related knowledge, workflows, or techniques from coworkers or employers — lower than the KPMG/University of Melbourne findings, but still a significant minority. At the same time, employees largely understood the collective value of sharing. Nearly four in five agreed that sharing their AI knowledge would improve day-to-day team tasks, help coworkers solve problems, and raise team productivity.

One of the strongest predictors of AI knowledge hiding in our data was organizational trust, which we measured based on employees’ responses to seven statements such as “In general, I believe my employer’s motives and intentions are good” and “My employer is not always honest and truthful.” Employees in the lowest quartile of trust were nearly four times as likely to have withheld AI knowledge as those in the highest quartile (47% versus 14%). A similar pattern held for psychological safety (45% versus 17%).

That does not mean other concerns were irrelevant. Employees were also more likely to hide AI-related knowledge when they felt greater job insecurity and when their workplace was more competitive. But organizational trust remained a strong predictor even after accounting for those factors, along with perceptions of distributive fairness, organizational innovativeness, age, gender, industry, job level, tenure, and whether the organization had an official AI policy and provided sanctioned AI tools. In plain terms, job loss fear and internal competition both mattered, but trust explained something they did not.

The more revealing finding was that the relationship between trust and AI knowledge-hiding weakened considerably after accounting for psychological safety, suggesting that trust may reduce knowledge hiding in part because it creates an environment where employees feel safe sharing how they work, experimenting openly, and discussing AI use without fear of judgment or negative consequences. Trust also mattered more when employees had access to a company-approved set of AI tools. In those environments, employees who trusted their organization were substantially less likely to withhold AI-related knowledge. When employees lacked a shared, sanctioned toolset, trust had a much weaker relationship with knowledge sharing. One interpretation is that trust creates the willingness to share, while a common set of approved tools creates the opportunity. Employees may be reluctant to invest time documenting workflows, prompts, or techniques that coworkers cannot easily access or use themselves.

Neither having an AI policy nor having access to approved AI tools, on its own, predicted whether employees withheld AI knowledge. This is not to dismiss the legitimate concern that some hiding reflects employees using unsanctioned tools, or using sanctioned tools in ways that raise real security and compliance risks—that narrower problem warrants its own response. But the broader pattern of AI knowledge hiding cannot be explained by, or solved through, formal AI infrastructure alone.

The logic behind these numbers came through starkly in open-ended responses. “I don’t trust my boss, and I need to maintain an advantage,” one respondent wrote. Another described their leadership as “foaming at the mouth trying to find ways to use AI to fire everyone,” and said they refused to share what they knew because they didn’t want to give anyone ammunition. These were stories about employees trying not to get burned by their organization.

A meta-analysis of 104 studies (covering roughly 31,800 employees) corroborates the broader pattern we observed: psychological safety is strongly associated with less knowledge hiding across contexts, while abusive supervision, workplace mistreatment, and job insecurity are among the strongest predictors of more hiding. Research on digital technostress similarly finds that technology-related stress increases knowledge hiding—but that this tendency is weaker in teams with supportive leadership and stronger in competitive cultures.

Why Employees Keep Quiet

Our interviews painted a consistent picture. Employees are making a rational calculation about the costs of making their AI workflows visible.

The first cost is reputational. A junior consultant in professional services told us that her colleagues were using AI in the same ways she was, but were not talking about it because they thought it made them seem less capable. At a health consulting firm, an analyst described a colleague who discovered a useful AI note-taking feature and shared it with her team—only to be called out by a senior team member who discredited the work because it was done by “a computer.” The message employees received was clear: the company may say it wants AI innovation, but local norms still punish visible use.

The second cost is workload. In many organizations, efficiency gains are not treated as a dividend to be reinvested thoughtfully. They are treated as spare capacity to be filled. A management consultant put it bluntly: “If I automate A and B, they’re not just gonna let me focus on C. They’re gonna make me do D, E, F.” When faster work is rewarded with more work rather than better work, employees have a reason to keep their best methods private.

The third cost is replaceability. Enterprise AI systems can record prompts, document workflows, and communication patterns, building a detailed map of an employee’s methods that can be handed to someone else or automated entirely. In a recent Wall Street Journal article, Texas A&M business school professor Matthew Call puts it plainly: knowledge that was once accumulated through years of experience can now be extracted, stored, and transferred to a replacement. His advice to employees was to use personal AI tools for their most valuable work rather than those provided by the organization, so that what they learned stays with them when they left. When hiding from your own employer becomes the prudent career strategy, the organization has already lost the trust battle.

These are three faces of one question: What happens to me when I let the organization see how I work?

What Leaders Can Do

The implication for leaders is uncomfortable. None of the concerns workers raised in our research were delusional or overly paranoid. They are reasonable readings of how organizations have historically handled employee productivity gains. If leaders want disclosure, they have to commit credibly to a different deal—one in which what employees share strengthens their position rather than undermining it. Applied to AI, employees need a reason to contribute, an easy path for doing it, and enough skill to turn a private workflow into something a colleague can actually use. The prescriptions below are versions of that commitment.

Earn the disclosure you want.

Employees need clear guidance on what AI use is encouraged, what’s off-limits, and how to handle gray areas. An organization does not have to redesign every workflow around AI to reduce hiding, but it does have to remove the ambiguity that forces employees to manage how their AI use will look instead of explaining how the work is getting done.

Research on electronic knowledge repositories has shown that employees contribute when the system reduces the work of codifying what they know, builds confidence that their knowledge is useful, and makes the social norms around sharing clear. For AI, that means not asking people to turn a working prompt sequence into a long process memo. Use lightweight templates, short demos, and “show me how you built this” sessions to convert private methods into reusable artifacts. Better yet, make asking part of the routine. In a field experiment with salespeople (another field where employees may feel incentivized to hide their reasons for success), randomly paired sales workers were assigned to structured meetings, joint-output incentives, both, or neither. Structured conversations about sales techniques produced average sales gains exceeding 15% that lasted at least 20 weeks after the experiment ended; collective incentives also produced gains, but they were short-lived.

Knowledge repositories may sit empty unless leaders build in recurring, structured conversations where it’s legitimate for one employee to ask another, “What exactly do you do that works?” Attach the contributor’s name to workflows that others adopt. Disclosure should build a person’s standing, not erase it.

Stop taxing efficiency gains.

If employees believe every hour saved by AI will be converted into more undesirable work, they will rationally hide those gains. Leaders need an explicit norm for how saved time will be used, whether for deeper analysis, higher-value projects, professional development, or recovery. A rule like this only works if employees see the upside, not just the extraction. A similar logic appears in supplier relationships: companies often find that suppliers are more willing to identify cost-saving opportunities when they can retain some of the resulting value rather than immediately surrender it through lower prices.

Reward multiplier behavior, not just individual AI productivity.

Employees will hide AI practices if sharing makes them less distinctive while everyone else gets the benefit. Do not rely on a generic AI leaderboard or a one-time bonus for workflow innovations; research suggests that comparison-heavy climates can amplify retaliatory knowledge hiding between coworkers, while learning climates reduce it. In one study of salespeople, team-based compensation generated substantially more knowledge transfer than individual commissions. But collective incentives alone are not enough. In an experiment, group-oriented incentives had only a weak effect on knowledge sharing by themselves; sharing rose mainly in pairs that had also developed pro-sharing norms, such as a shared expectation that people would volunteer what was working and help a struggling colleague without being asked.

Reward reusable workflows, peer adoption, and quality improvements. Possible rewards might include credit in performance reviews for methods that others adopt, protected time to keep experimenting, and a share of the gains once a workflow is in wider use. Then close the loop by telling people where their contribution was used, what improved, and what credit they will receive. This is the difference between asking employees to donate their advantage and letting them become recognized multipliers.

Legitimize AI experimentation, then surface it.

Sanctioned exploration time isn’t new: 3M’s 15% rule produced the Post-it Note, and Google’s 20% time produced AdSense and Google News. Anthropic’s Claude Code team has adapted the idea for the AI era as “side quests”—self-directed experiments engineers, designers, and product managers run outside the official roadmap. The naming matters because it converts AI tinkering from corner-cutting into a sanctioned category of work. Once that legitimacy is in place, the usual surfacing mechanisms—shared workflow repositories, team demos, recognition for adopted methods—can do their job.

Treat disclosure as a contribution.

The above prescriptions are about structures. This one is about what a manager does in the thirty seconds after an employee shows their work. That reaction is the most decisive trust signal an organization sends. Treat the disclosure as corner-cutting, even implicitly, through a skeptical question or visible disappointment, and employees learn to hide. Treat it as something worth understanding, and they learn that disclosure pays. The trap is subtler than open hostility, and more well-meaning: letting one person’s disclosure turn into a standing obligation to bring everyone else up to speed. That converts honesty into unpaid labor and is one of the most reliable ways to ensure the next discovery stays private.

The way out is simple: limit the cost of sharing. The person who found the method demonstrates it once. From there, the company owns the work of documenting, distributing, and supporting it, while the discoverer keeps the credit. Sharing should raise a person’s standing, not become a permanent obligation.

A warning about tooling.

Sanctioned enterprise AI tools log what employees do—the prompts, the iterations, the workflows that worked. That logging is what lets organizations identify whose discovery improved things, credit them, and spread the method to colleagues. It’s also what lets them extract that discovery, document it as a process, and route the work elsewhere. Our data shows approved tools amplify the relationship between trust and hiding: where trust is high, employees with approved tools hide less; where trust is low, they hide more. The mechanism likely runs through what employees expect the organization to do with what the tools can see. In organizations where trust is shaky, an enterprise AI rollout is more likely to entrench hiding than to bring it into the open.

Discover more