Why Do Employees Hide Their AI Use at Work?

For transformation leaders, CHROs, CIOs and the managers who keep being told that adoption is going well. Somewhere between a third and two thirds of your people are using AI without telling you. The reason is not confusion about the policy, and a clearer policy will not fix it. Reading time about 10 minutes.

The short answer

Employees hide their AI use because disclosure carries a measurable cost and concealment does not.

That is not a turn of phrase. Thirteen experiments published in Organizational Behavior and Human Decision Processes in 2025 found that people who disclose using AI are trusted less than people who do not. The effect held across writing, analysis, teaching, financial advice and design, and it held whether the person doing the judging was a student, a hiring manager, a client or an investor.

The penalty is specific to AI rather than to getting help at all. In one of those experiments, a job applicant who disclosed that a human career coach had helped write their letter suffered no measurable loss of trust compared with an applicant who disclosed nothing. Disclosing AI in the same position produced a large drop.

So employees who stay quiet are not being evasive. They are reading their environment accurately. That is also why the standard corporate response, a policy requiring disclosure, does not solve the problem. The same researchers tested mandatory and voluntary disclosure regimes side by side, and the trust penalty appeared in both.

How many people are actually hiding it

Estimates run from roughly 39 percent to 66 percent, and most of that spread is a measurement artefact. The surveys are asking different questions.

The first three numbers describe different behaviours. Not volunteering that you used AI on a document is not the same as knowingly working around a policy, and neither is the same as telling a survey you would prefer to keep quiet. Any single headline percentage for this flattens three distinct things.

The fourth number is the one a leadership team should sit with. Six in ten employers believe they have visibility. On the employees' own account, they do not.

One further figure is worth holding onto: only 28 percent of companies require employees to use AI tools, yet 87 percent of employees report using it. Most enterprise AI adoption is not something the organisation caused. It is something the organisation is discovering.

The reason employers assume, and the reason employees give

There is a clean attribution gap in the data, and it matters because the two explanations point at different remedies.

Forty-seven percent of employers believe employees stay quiet because they fear for their jobs. Only 24 percent of employees say that is their reason. What employees report instead is reputational. Thirty-four percent expect to be seen as cutting corners, 27 percent expect to be judged or misunderstood, and 45 percent simply do not think disclosure is required, because no rule says it is.

The PagerDuty survey found a similar shape from a different angle. Thirty-three percent said they would conceal AI use to avoid scrutiny from managers, 30 percent cited restrictive policies or fear of peer judgement, and 29 percent were not sure whether their use was permitted at all.

The fear is not imagined. Among employees who had used AI outside company policy, 53 percent said they received informal feedback or guidance about it, and 48 percent said they faced formal consequences such as a warning or disciplinary action.

An organisation working from the job security theory will run reassurance campaigns about augmentation rather than replacement. That addresses a fear most employees are not reporting. The reported fear is narrower and more immediate: this will make me look worse at my job. There is now direct experimental evidence that the belief is correct.

The evidence most AI policies were written without

Oliver Schilke and Martin Reimann, at the University of Arizona, ran thirteen experiments on what happens to trust when someone says they used AI. The paper, The transparency dilemma: How AI disclosure erodes trust, was published open access in Organizational Behavior and Human Decision Processes in 2025. Four results matter for anyone writing an AI use policy.

Disclosure lowers trust, and the effect is large. In the hiring study, 85 managers with hiring experience read a letter of motivation from an applicant for a senior tax accountant role. When the letter disclosed generative AI, trust in the applicant fell sharply against both the no-disclosure condition and the condition where a human career coach was credited. The two human conditions were statistically indistinguishable from each other. Likelihood of hiring moved with it.

The mechanism is legitimacy, not accuracy. Across several study designs the drop in trust was explained by reduced perceptions of legitimacy, meaning the extent to which the person's conduct is seen as proper by the group. This is not simple distrust of algorithms. The authors tested that and found the disclosure effect operating above and beyond it.

Making disclosure mandatory does not remove the penalty. One study manipulated the regulatory regime, telling participants that legislatures were moving either to require or to permit voluntary disclosure. Trust in a designer who disclosed AI fell in both conditions, and by a similar margin. The penalty is social, so a rule that compels the disclosure does not license it.

Being exposed by someone else is worse than disclosing. In the final study, an adviser exposed by a third party for undisclosed AI use was trusted less than one who disclosed voluntarily. Concealment is not a free option. It is a bet that carries a larger loss if it fails.

There is a fifth result that quietly undermines the most common reason given for doing nothing. A meta-analysis across the thirteen experiments found the penalty was reduced, though not eliminated, among evaluators with favourable attitudes to technology and high confidence in AI accuracy. It found no evidence of any reduction among evaluators who used AI themselves or were highly familiar with it. Familiarity did not dissolve the penalty. Waiting for AI to become normal is not, on this evidence, a strategy.

Why requiring disclosure does not fix it

A disclosure mandate moves a cost onto the individual and calls it integrity. It asks an employee to accept a hit to how colleagues and managers see them, so that the organisation can hold an audit trail. Stated that way, the compliance pattern is predictable. People disclose where the chance of being caught is high and the work is visible, and stay quiet everywhere else.

This is also why tightening the policy tends to make the picture worse rather than better. Thirty percent of concealment in the PagerDuty data was attributed to restrictive policies or fear of peer judgement, and 29 percent to genuine uncertainty about what was allowed. A longer policy document addresses the third group and adds pressure to the first two.

Detection is not pointless. Because exposure is worse than disclosure, an organisation with a credible ability to see AI use changes the arithmetic an employee is doing, and the researchers note exactly this. But detection on its own produces fear rather than candour. It raises the cost of concealment without lowering the cost of disclosure, and the employee is still choosing between two losses.

What actually reduces the penalty

One intervention in the research did reduce it, and it operates at the level of the group rather than the individual.

In the collective validity study, some participants were first shown a published excerpt from a credible institutional source reporting that generative AI improves performance on ideation and content creation tasks. Those participants penalised the discloser less. When AI use is visibly treated as appropriate by people whose judgement carries weight, disclosing it stops reading as a confession.

That is the practical shape of the fix. Legitimacy has to be supplied from above rather than demanded from below. Concretely, and this is MASSIVUE's reading of what the mechanism implies rather than something the experiments tested directly, four things follow.

1. Sanction specific uses, not AI in general. "Use AI responsibly" leaves the employee to justify their own choice. "Approved tool X for first drafts of Y, with review at Z" means the person disclosing is following an instruction rather than admitting a preference. The legitimacy comes from the institution.
2. Have senior people disclose first, repeatedly, on real work. A base rate that is visible from above is the closest operational equivalent of the collective validity manipulation. A single leadership email does not create one.
3. Attach disclosure to the artefact and the review step, not to the person. A field on a document template asking which parts were AI assisted is a process record. The same information volunteered in a one to one becomes evidence about the individual, which is precisely where the penalty lands.
4. Never let a disclosure surface in a performance conversation as a fact about capability. If it does once, in one team, the news travels faster than any policy, and you have taught the organisation that disclosure is expensive.

None of this is culture work in the soft sense. It is a change to what the organisation formally credits and records. That is a change management problem with an operating model attached, which is the same problem enterprises face when they try to manage change during an enterprise AI transformation more broadly.

Stop measuring self-reported AI adoption

If disclosure is costly, every number you collect by asking people about their AI use is biased in one direction, and you cannot correct for it because you do not know the size of the bias in your own organisation.

Self-report is unreliable even when nobody has a reason to shade the truth. In a randomised trial published by METR in July 2025, 16 experienced open source developers completed 246 real tasks on repositories they knew well. Allowing AI tools made them 19 percent slower. Afterwards, they estimated that AI had made them 20 percent faster. METR now labels the result historical, since the tools and workflows have moved on, and the sample was small. The point that survives is narrower and still useful: people are poor witnesses to their own productivity, so a self-estimated time saving is not evidence.

Volume is no better as a proxy. In research by BetterUp Labs with the Stanford Social Media Lab, a survey of 1,150 US full-time desk workers in September 2025, 40 percent said they had received AI generated work that looked finished but lacked substance in the previous month. Each incident took about two hours to resolve, which the researchers costed at roughly $186 per employee per month, or around $9 million a year across a 10,000 person organisation. More AI output is not the same as more value, and the cost of the difference lands on the recipient rather than the sender.

The last row is the one that most often gets cut for being unenforceable. It is the only line in the table that tells you what is happening in the part of the organisation you cannot see, which is the part this article is about.

What a manager can do this week

None of the above requires a programme to begin.

• Stop asking "are you using AI?" Given the penalty, the answer carries almost no information. Ask instead which part of a task the person would hand to a machine tomorrow if they were allowed to, and what would have to be true for that to be safe.
• Say what you use it for, specifically, before you ask anything. Not that you are supportive of AI. Which tool, on which task, last week, and what you had to fix afterwards.
• Name one sanctioned tool and one sanctioned use for your team's most common task. One is enough to change the default from private judgement to shared rule.
• Watch rework rather than output. If work is being sent back more often, you have a quality problem that more output will worsen and that no adoption metric will show you.

There is an older skill underneath all of this, and it is the reason the question is harder than it looks. What people tell you is filtered by what it costs them to tell you. A manager who cannot estimate that cost will keep mistaking silence for agreement, and a clean adoption dashboard for a working rollout. The related question of whether you should require AI use in the first place is covered in Should AI Use Be Mandatory at Work?, and the question of what a human is genuinely adding when they review machine output is covered in When Should a Human Override an AI Recommendation?

Where MASSIVUE fits

MASSIVUE works with enterprises on AI workforce transformation: skills assessment and gap analysis, role design, and the capability programmes that make sanctioned AI use specific enough to be disclosable.

If the immediate problem is that your adoption reporting does not match what people are actually doing, the closest fit in MASSIVUE Academy is the AI Change Management: Upskilling & Reskilling micro-credential, which covers diagnosing skills gaps, designing upskilling pathways, managing resistance honestly and proving the return. For leaders setting the operating model that decides which uses are sanctioned in the first place, what an AI operating model is is the prior question.

Frequently asked questions

Why do employees hide their AI use at work?

Because disclosing it costs them something and concealing it does not. Thirteen experiments published in Organizational Behavior and Human Decision Processes in 2025 found that people who disclose AI use are trusted less than people who do not, and that the effect is driven by perceived legitimacy rather than by doubts about accuracy. Surveys add the reasons employees give directly: expecting to be seen as cutting corners, expecting to be judged, and not knowing whether their use is permitted.

How many employees hide their AI use?

Between roughly 39 and 66 percent, depending on the question asked. Slingshot's Digital Work Trends Report in January 2026 found 45 percent hide or avoid disclosing AI use. PagerDuty's 2026 survey of 1,250 office professionals found 39 percent would rather use AI without telling anyone, and 66 percent had used AI at work despite believing it was not permitted. The same PagerDuty data found 60 percent of employers believe their people are being honest with them.

Does requiring employees to disclose AI use solve the problem?

No, on the available evidence. One of the thirteen experiments manipulated whether disclosure was voluntary or mandated by regulation and found the trust penalty in both conditions, at a similar magnitude. A rule can compel the disclosure but it does not make the disclosure socially safe, and the person disclosing still absorbs the cost.

Is hidden AI use a security problem or a culture problem?

Both, and the security exposure is the more urgent of the two. Employees working around sanctioned tooling are frequently putting business information into consumer services with no enterprise agreement behind them. But treating it only as a security problem leads to tighter restriction, which the survey data links to more concealment rather than less. Provision usable sanctioned tools and fix the disclosure incentive at the same time.

Will the trust penalty fade as AI use becomes normal?

There is no evidence yet that it will fade on its own. The meta-analysis across the thirteen experiments found the penalty reduced among evaluators with favourable technology attitudes and high confidence in AI accuracy, but found no reduction among evaluators who used AI themselves or were highly familiar with it. Personal familiarity did not translate into leniency towards someone else who disclosed.

What should we measure instead of self-reported AI adoption?

Telemetry from sanctioned platforms reported at team level, weekly active use on the specific workflows your business case named, cycle time on a defined task measured before and after, and the rate at which work is sent back for rework. Add an anonymous channel for reporting unsanctioned tool use with no consequence attached, because it is the only one of these that sees the activity happening outside your tooling.

Sources

• Oliver Schilke and Martin Reimann, The transparency dilemma: How AI disclosure erodes trust, Organizational Behavior and Human Decision Processes, volume 188, 2025, article 104405. Thirteen experiments, open access. https://doi.org/10.1016/j.obhdp.2025.104405
• PagerDuty, Shadow AI Survey 2026. Conducted by Wakefield Research among 1,250 office professionals at companies with minimum annual revenue of $500 million, excluding IT and technology roles, in Australia, Japan, the United Kingdom and the United States, 9 to 20 April 2026. Vendor-commissioned. https://www.pagerduty.com/blog/ai/shadow-ai-workplace-survey-2026/
• Slingshot, Digital Work Trends Report, Part 1, released 15 January 2026. Vendor survey; sample size not disclosed in the release. https://www.globenewswire.com/news-release/2026/01/15/3219591/0/en/nothing-to-see-here-nearly-half-of-employees-hide-their-ai-use-at-work.html
• METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, 10 July 2025. 16 developers, 246 tasks. Labelled by METR as a historical result. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
• BetterUp Labs with the Stanford Social Media Lab, Workslop: The Hidden Cost of AI-Generated Busywork. Survey of 1,150 US full-time desk workers, September 2025. https://www.betterup.com/workslop
• Kate Niederhoffer, Gabriella Rosen Kellerman, Angela Lee, Alex Liebscher, Kristina Rapuano and Jeffrey T. Hancock, AI-Generated "Workslop" Is Destroying Productivity, Harvard Business Review, 22 September 2025. https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity