AI Knowledge Hiding

Confidence 0.85 · 4 sources · last confirmed 2026-10-11

AI knowledge hiding (Anicich & Brouwers, HBR 2026) is the deliberate withholding — by employees — of the AI workflows, prompts, and techniques they discover through private experimentation. The authors frame it as “the suppression of solutions”: where the classic organizational-silence literature studied the hiding of problems (bad news, risks, ethical concerns), AI introduces the hiding of solutions.

Why it is newly consequential

Productivity gains “used to scale by default” — embedded in shared systems, standardized processes, and formal tools that spread improvements structurally. With AI, the most valuable gains come from individual experimentation: a prompt sequence that produces client-ready output in a fraction of the time. That knowledge is portable, easy to refine in private, and easy to keep to yourself. When an individually-discovered workflow cuts a three-hour task to twenty minutes and is trivial to conceal, silence becomes economically consequential in a way it was not before. This is why AI knowledge hiding is a distinct failure mode of enterprise-ai-adoption rather than a restatement of generic knowledge hoarding: high individual adoption can coexist with near-zero collective capability gain.

The empirical claim: trust, not governance

The headline finding is that the binding driver is organizational trust, working largely through psychological safety — not formal AI policy or approved tooling.

• KPMG / University of Melbourne (48,000+ respondents): 57% admit hiding their AI use at work. Primary source: Gillespie et al. 2025 (see the section added 2026-10-11 below).
• Anicich & Brouwers’ own survey (604 daily-AI users): 30.3% intentionally withheld AI knowledge; the lowest-trust quartile was ~4× more likely to hide than the highest (47% vs 14%); psychological safety showed the same gap (45% vs 17%).
• Trust survived controls for job insecurity, internal competition, distributive fairness, AI policy, sanctioned tools, age, gender, tenure, and job level. Having an AI policy or approved tools, on its own, predicted nothing. Trust mattered more where a shared sanctioned toolset existed (“trust creates the willingness to share; a common toolset creates the opportunity”).
• Corroboration: a 104-study meta-analysis (~31,800 employees) links psychological safety to less knowledge hiding, and abusive supervision / mistreatment / job insecurity to more; a Stanford 51-deployment study found 77% of the hardest adoption challenges were non-technical.

The three rational costs of disclosure

Employees make a rational calculation about the cost of making their AI workflows visible:

1. Reputational — being judged as less capable, or having the work discredited (“a computer did it”).
2. Workload — efficiency treated as spare capacity to fill: “if I automate A and B, they make me do D, E, F.” This is the micro-productivity-trap at the individual level.
3. Replaceability — enterprise tools log prompts and workflows, so a method can be extracted and routed to a cheaper replacement; hence the (perverse) advice to use personal AI tools so learning “stays with you.”

Underlying all three is Amy Edmondson’s distinction: organizations confuse praiseworthy exploratory testing with blameworthy deviance, and so punish exactly the experimentation they need.

What leaders can do (the source’s prescriptions)

Earn the disclosure you want (remove ambiguity; lightweight demos over process memos; credit the contributor); stop taxing efficiency gains (an explicit norm for reinvesting saved time); reward multiplier behavior (credit for adopted workflows; team incentives + pro-sharing norms, since comparison-heavy climates amplify retaliatory hiding); legitimize experimentation, then surface it (Anthropic’s Claude Code “side quests” ≈ 3M 15% / Google 20% time); treat disclosure as a contribution (the manager’s 30-second reaction is the decisive trust signal; don’t turn one demo into a standing obligation). A standing tooling warning: the same logging that lets you credit a discovery also lets you extract it.

Debates and supersession

• Is this distinct from classic knowledge hoarding? The novel claim is the suppression of solutions (not problems) plus the portability/concealability of AI workflows that makes silence newly costly. Single-source so far; the supporting meta-analysis is about knowledge hiding in general, not AI specifically.
• Single-source caveat. Rests on one HBR article (with its own 604-respondent survey + cited studies). Confidence at the single-source floor pending independent corroboration.

Related concepts

• enterprise-ai-adoption — hiding is why high individual adoption need not yield firm-level gains; trust/culture is the binding constraint, not tooling.
• micro-productivity-trap — the workload “tax” is the shared mechanism; hiding is the individual-level response that keeps task gains from aggregating.
• durable-skills / ai-deskilling — the replaceability fear (your logged method becomes someone else’s, or an automation’s).
• responsible-ai — the surveillance/credit-vs-extract trade-off in sanctioned-tool logging.

Open questions

• Does the trust→disclosure finding replicate outside the U.S. / outside knowledge work?
• Do “side quests” and multiplier-reward schemes measurably raise disclosed (not just total) AI productivity?
• How does mandatory enterprise-tool logging net out — does the credit mechanism or the extraction fear dominate as trust varies?

The formal mechanism (added 2026-09-15)

This page rested on a single survey source, which located the driver in organisational trust working through psychological safety. Yuan, Aseri & Ramasubbu (2026) derive the same behaviour from compensation structure alone, with no trust deficit required — and that is a materially different diagnosis with a different remedy.

Their game-theoretic model splits employees on two independent dimensions: tangible skill, which an AI system can observe, codify and redistribute, and intangible skill (communication, coordination), which it cannot. Under pay-for-performance rank-order competition, deploying an AI that learns from the high performers erodes exactly the advantage those performers are paid for. Output and welfare fall for the high-tangible/low-intangible group; low-tangible employees rise; firm profit can fall.

Read against this page, the implication is uncomfortable: hiding is rational even in a high-trust organisation, provided rewards are relative. An employee who discovers a workflow that triples their output, working under PFP, is not being paranoid when they keep it — disclosure is the mechanism by which their rank is competed away.

Two consequences worth carrying:

• The prescriptions on this page are incomplete. Trust-building and psychological safety address the disclosure cost the survey identified; they do not touch the rank-order incentive. A firm could do everything Anicich & Brouwers recommend and still get hiding, if it pays on a curve.
• The proposed remedy is counter-intuitive and untested. Yuan et al. find that guaranteeing the wages of disadvantaged employees can leave the firm worse off than doing nothing, and that the profit-maximising move is to bound AI capability deliberately — “pursuing the most advanced AI system is not always optimal.”

Treat this as mechanism, not evidence: it is an unrefereed analytical model with no data, and its load-bearing assumption (AI transfers hard skill but not soft) is asserted rather than shown. But it is the first source in the corpus that explains hiding without appealing to culture.

The primary survey, and the effect of bans (added 2026-10-11)

The 57% figure above now has its primary source in the wiki: KPMG), 48,340 respondents in 47 countries, fieldwork November 2024 to January 2025. In the employee section, 57% have used AI non-transparently (presenting AI output as their own, or not saying they used it); 48% have uploaded sensitive company information or copyrighted material into public AI tools; 56% have used AI at work without knowing whether it is allowed. Among students, two-thirds report the same non-transparent use.

Two findings bear on the prescriptions above:

• Bans go with more risky use, not less. Uploading company data to public tools is reported by 67% of employees whose organisation has banned generative AI and 56% where a policy guides it, against 33% where there is no policy. The authors: “outright bans may be ineffective, and … simply having policies does not guarantee compliance.” This is a cross-sectional association; organisations may ban AI because misuse was already high.
• Policy is often invisible to staff. 34% report a policy guiding generative AI use, 6% a ban, 41% none, and 19% do not know.

The authors expect under-reporting, “given social desirability bias”, which is the knowledge-hiding mechanism itself applied to the survey.

Mollick (HBR IdeaCast, Oct 2026) puts the cause in culture rather than rules: with the right structure, culture and rewards, people bring use cases forward; otherwise “everyone’s hiding all their uses … because they’re scared if they tell you they’ll get fired or they won’t get respected.”