How Do You Build an AI-Ready Culture? Start With What Your People Are Hiding

The short answer

You build an AI-ready culture by measuring the gap between how much AI your people actually use and how much they are willing to admit to, then closing that gap by changing what disclosure costs them. Culture is not what your values statement says. It is what people do when they believe nobody is checking, and AI has made that unusually easy to measure.

Executive summary

Most enterprises still approach culture the way they did a decade ago. Run a survey, hold an offsite, publish a refreshed set of values, appoint champions, expect behaviour to follow. The evidence that this works has always been thin. MIT Sloan Management Review compared the official values of 562 large companies against 1.2 million employee reviews and found no correlation between the values a company promotes and how employees rate it on those same values. Four of the nine values studied correlated negatively.

AI has changed the diagnostic situation. For the first time there is a behaviour that is widespread, consequential, easy to conceal and directly measurable: whether people tell you how they use AI. In the largest global study on the subject, KPMG and the University of Melbourne found that 57 percent of employees hide their AI use and present AI-generated work as their own. That is not a training problem. It is a reading of the culture.

It is also rational behaviour. Thirteen experiments published in Organizational Behavior and Human Decision Processes found that people who disclose their AI use are trusted less than those who do not, and that the effect survived every mitigation the researchers tried, including making disclosure mandatory. Your people are not concealing AI use because they lack integrity. They are concealing it because disclosure carries a cost and silence does not.

This guide sets out what culture actually is, why declaring values does not change it, how to measure your own disclosure gap with a single anonymous survey, the five conditions that close it, a 90 day sequence, and the strongest argument against this entire framing. Where a recommendation is MASSIVUE practitioner judgement rather than a research finding, it says so at the point of use.

Key takeaways

1. Culture is observable behaviour, not stated intent. The distance between the two is measurable, and in most organisations it is wide.
2. Published values do not predict lived culture. Across 562 companies the correlation was effectively zero, and negative for collaboration, customer orientation, execution and diversity.
3. The disclosure gap is the sharpest culture signal available in 2026. It is the difference between the AI use that happens and the AI use people will admit to.
4. Concealment is a rational response to a real penalty. Disclosing AI use reduces how much others trust you. Exhortation does not defeat an incentive.
5. Managers move this more than executives do. When managers visibly use AI themselves, employees report materially higher confidence in it and better critical thinking about it.
6. A disclosure rule without cover makes things worse. Requiring disclosure before removing the penalty converts a culture problem into a compliance problem, and the gap widens.
7. Only 13 percent of workers say they are rewarded for reinventing work when results miss targets. Until that changes, staying quiet is the sensible choice for a competent employee.

What is organisational culture?

Organisational culture is the sum total of behaviours and interactions among individuals, teams and groups of people. You observe it as a set of patterns in day to day work. It is not the poster in reception, and it is not the engagement score.

That definition, which MASSIVUE has used with client executive teams for years, has one practical consequence that most culture programmes ignore. If culture is a pattern of behaviour, then it can only be assessed by observing behaviour. Asking people to describe their culture measures their opinion of it, which is a different and much softer quantity.

Culture guides decisions that nobody is watching. It determines what people do when there is no instruction, no policy and no supervisor. It lives in assumptions people rarely articulate, which is precisely why it resists change by announcement. Any real cultural change first requires surfacing those assumptions, then changing the conditions that produced them.

Culture, climate and engagement are not the same thing

Why values statements do not change culture

This is worth stating plainly because it invalidates the most common first step in culture programmes, including the one MASSIVUE itself recommended in earlier versions of this article.

Donald Sull, Stefano Turconi and Charles Sull analysed the published value statements of 562 companies and compared them with 1.2 million employee reviews, measuring how prominently each company promoted a given value against how positively employees discussed that same value. They found no correlation between the cultural values a company emphasises in its published statements and how well it lives up to them in the eyes of employees. All correlations were very weak. Four of the nine values studied, collaboration, customer orientation, execution and diversity, were negatively correlated. The strongest relationship they found, agility at 0.22, still describes a very weak link between a company's public commitment to flexibility and employees' assessment of how agile it actually is.

The finding is not that values are worthless. It is that publishing them changes nothing on its own. A value statement is a claim. Culture is the evidence. Where the two diverge, employees believe the evidence, and they are right to.

What is the disclosure gap?

The disclosure gap is the difference between the AI use that is actually happening in your organisation and the AI use your people are willing to tell you about.

It matters for three separate reasons, and they belong to three different executives.

• It is a trust measure. People conceal what they believe will be punished. A wide gap tells you what your organisation actually rewards, regardless of what it says.
• It is a control measure. Work you cannot see is work you cannot review. In the KPMG and University of Melbourne study, 66 percent of employees using AI said they rely on its output without evaluating accuracy, and 56 percent reported making mistakes in their work because of AI. Concealed use means those errors reach decisions unreviewed.
• It is a capability measure. If people hide how they use AI, the organisation cannot learn from the people who are good at it. Your best practice stays private, and every team solves the same problem alone.

How wide is it, typically?

The best available global figure comes from KPMG and the University of Melbourne, who surveyed more than 48,000 people across 47 countries between November 2024 and January 2025. In that study, 57 percent of employees said they hide their use of AI and present AI-generated work as their own. In the same population, only 47 percent had received any AI training and only 40 percent said their workplace had a policy or guidance on generative AI use.

Those three numbers belong together. A majority conceal their use. A majority have no formal guidance telling them what is permitted. In the absence of a rule, people assume the strictest one they can imagine and act accordingly, which means quietly.

Regional data points the same way. Slack's Workforce Index, which surveyed 17,372 desk workers including 1,008 in Singapore in August 2024, found that 45 percent of Singapore desk workers were uncomfortable admitting AI use to their manager. The top reasons given were fear of being seen as less competent, fear of being seen as lazy, and a feeling that using AI is cheating.

Why hiding AI use is rational, not deviant

This is the part that most culture advice gets wrong, and it is the reason communication campaigns fail here.

Oliver Schilke and Martin Reimann ran thirteen experiments, published in Organizational Behavior and Human Decision Processes in 2025, testing what happens to trust when someone discloses that they used AI. The result was consistent across contexts including education, investment advice, job applications, creative work and routine corporate communication: people who disclose AI use are trusted less than those who do not. The researchers attribute the effect to perceived legitimacy. Disclosure reads as an admission that the work was not entirely the person's own.

They then tried to break the effect. Framing the disclosure differently did not prevent it. Evaluators already knowing about the AI use did not prevent it. Making disclosure mandatory rather than voluntary did not prevent it.

Read that alongside the concealment data and the picture is uncomfortable but clear. Your employees are responding correctly to their environment. Disclosure costs them credibility. Concealment costs them nothing unless they are caught, and in most organisations there is no mechanism to catch them.

Why mandatory disclosure alone backfires

The obvious response is to require disclosure. Debevoise and Plimpton set out a sensible version of this in February 2026, recommending that disclosure be required when a substantial portion of a document was generated by AI, when the work product may be relied on for decisions, and when errors in it could affect those decisions. That is a well drafted trigger, and it is a reasonable thing to adopt.

The authors acknowledge the objection that disclosure could stigmatise AI use, and describe that concern as overstated. On the experimental evidence, we think the concern is real rather than overstated. The trust penalty exists, it is measurable, and it does not disappear when disclosure is made mandatory.

That does not mean skip the rule. It means the rule is necessary and not sufficient. If you introduce a disclosure requirement into an environment where disclosure is still penalised, you have created a compliance obligation that people will meet in the narrowest possible way while continuing to conceal everything that falls outside the trigger. The measured gap will appear to close while the real one widens.

How to measure your disclosure gap

One anonymous survey, one population, three questions. The distance between the answers is the measurement. This instrument is MASSIVUE's, built from the research described above, and you are welcome to run it yourself.

The bar widths above are illustrative of the pattern the global research describes, not a benchmark. Your own numbers are the point.

Four rules for running it credibly

1. Anonymity has to be real and demonstrable. If people suspect the survey is attributable, question A returns a fiction and the whole exercise is worthless. Report at team level with a minimum cell size, and say what that minimum is before you field it.
2. Do not run it in the same instrument as a performance or engagement survey. Proximity to anything that feels evaluative depresses the honest answer to A.
3. Ask about the last 30 days, not in general. General questions invite people to answer about their intentions rather than their behaviour.
4. Add one free text question. "What would have to be true for you to tell your manager exactly how you use AI?" In our experience this single question produces more usable material than the quantitative results, because it surfaces the specific local penalty rather than the general one.

How to read the result

Five conditions that close the gap

These five conditions are MASSIVUE's, carried forward from the success factors we identified working with senior executives across transformation programmes, and rewritten for the specific problem of AI disclosure. Each names one owner and one observable. If you cannot point at the observable, the condition is not in place, whatever the values statement says.

Why sequence matters more than completeness

Conditions 1 and 2 are prerequisites. Running condition 3 without them converts a disclosure rule into a compliance trap, for the reasons set out earlier. This is our judgement rather than a measured finding, but it is consistent with the experimental evidence that mandatory disclosure does not remove the trust penalty.

Condition 1 has the strongest evidence behind it

Microsoft's 2026 Work Trend Index surveyed 20,000 knowledge workers across ten markets between February and April 2026. When managers actively modelled AI use, employees reported a 17 point lift in perceived AI value, a 22 point lift in critical thinking about AI use, and a 30 point lift in trust in agentic AI. Where managers created psychological safety around experimentation, employees showed up to 20 points higher AI readiness and were 1.4 times more likely to be high frequency agentic AI users.

The same study found that only 26 percent of AI users say their leadership is clearly and consistently aligned on AI. The lever with the best evidence behind it is also the one most organisations have not pulled.

Condition 5 is the one most organisations skip

Microsoft describes what it calls the Transformation Paradox: employees are ready to reinvent how they work, but the system around them, the metrics, incentives and norms, continues to reinforce the old way. The supporting numbers are stark. 65 percent fear falling behind if they do not adapt to AI. 45 percent say redesigning their work feels riskier than hitting their current goals. Only 13 percent report being rewarded for reinventing work with AI when the results miss targets.

If 45 percent of your people believe that changing how they work is riskier than protecting this quarter's number, no amount of culture communication will move them. That is not a belief problem. It is an accurate reading of your performance management system.

A 90 day sequence

This sequence is MASSIVUE's recommended order of work. It assumes an enterprise of a few thousand people with AI tools already in the environment, whether sanctioned or not.

Re-run the diagnostic at six months, not at three. Disclosure behaviour lags the conditions that produce it, and an early re-measure will tempt you to declare victory on a survey effect.

Is this culture, or just a bad rollout?

There is a serious argument against everything above, and it deserves a fair hearing rather than a straw man.

Salesforce, with YouGov, surveyed more than 1,500 desk workers across 14 markets between December 2025 and January 2026. Singapore workers turned out to be among the least sceptical about AI globally, at 29 percent identifying as sceptics against a global average of 37 percent, yet only 6 percent said AI was a core part of their daily work, against a global average of 11 percent. Among Singapore workers who had experienced an unsuccessful AI pilot, 40 percent blamed generic outputs, 38 percent low trust in the outputs, and 30 percent a lack of business context. Salesforce's conclusion is direct: the barrier is a delivery gap, not cultural reluctance.

That reading is coherent and partly correct. If the tool produces generic output that is wrong for the job, people will not use it, and no amount of psychological safety will change that. Anyone selling culture work as the answer to a bad deployment is selling the wrong thing.

Two qualifications. First, that survey covered just over 1,500 workers across 14 markets, so the Singapore subsample is small and the market level figures should be treated as indicative rather than precise. Second, and more importantly, the two explanations are not competing. A delivery gap explains why people stop using a tool. It does not explain why 57 percent of people who are using AI decline to say so. Concealment is a separate behaviour from non-adoption, and it has its own separate cause, which the experimental evidence identifies as a trust penalty.

What this looks like in Singapore and APAC

The behavioural mechanics are the same everywhere. The context around them is not.

Singapore has an unusually active institutional layer. IMDA has committed to training tech professionals under its National AI Impact Programme, building on a national commitment to equip 100,000 non-tech workers with AI capabilities by 2029. That matters for condition 4, because protected practice time and funded training are easier to justify when public programmes are already carrying part of the cost.

Two local specifics are worth naming. First, the Slack Workforce Index figure of 45 percent of Singapore desk workers uncomfortable admitting AI use to a manager sits close to the global concealment picture, so this is not a market where the problem is milder. Second, the reasons given in Singapore, being seen as less competent or lazy, map onto a working culture where visible effort has traditionally been part of how contribution is judged. Where effort is the proxy for contribution, a tool that reduces effort is a threat to how you are evaluated. Naming that dynamic explicitly, rather than treating it as a general resistance to change, is in our experience what unlocks the conversation with APAC leadership teams.

For regulated institutions the disclosure question also becomes a supervisory one, since concealed AI involvement in a decision is a documentation gap as well as a cultural one. Our guide to end to end change management in Singapore covers that regulatory layer in more depth.

Limitations of this guide

Stated plainly, because an article that claims no limits should not be trusted.

• The headline concealment figure predates the agentic era. The KPMG and University of Melbourne fieldwork ran from November 2024 to January 2025. Attitudes may have shifted since, in either direction. No later wave of that study was available at the time of writing.
• The trust penalty research is experimental, not field based. Thirteen experiments establish the effect robustly under controlled conditions. They do not tell you the size of the penalty inside your specific organisation.
• The Microsoft sample skews Western. Ten markets, of which only Australia, India and Japan sit in Asia Pacific. Singapore is not among them.
• The MIT Sloan values research is based on US companies and Glassdoor reviews. Both introduce selection effects. The finding is strong enough to act on and not strong enough to treat as universal.
• The diagnostic and the five conditions are MASSIVUE's design judgement. They are built on the research cited here, but the instrument itself has not been independently validated, and we would rather say so than imply otherwise.

Frequently asked questions

What is an AI-ready culture?

An AI-ready culture is one where people can use AI on real work, say so accurately, and be judged on the outcome rather than on how much visible effort the work took. The practical test is not whether people are enthusiastic about AI. It is whether the AI use that happens is the AI use leadership can see.

Why do employees hide their AI use?

Because disclosure costs them. Experimental research published in 2025 found across thirteen studies that people who disclose AI use are trusted less than those who do not, an effect the researchers link to perceived legitimacy. Employees also report fearing that they will look less competent or lazy, or that using AI counts as cheating. In most organisations no written rule contradicts those fears.

How do you measure organisational culture?

By measuring behaviour rather than opinion. For AI specifically, ask one population three anonymous questions: whether they used AI for work in the last 30 days, whether their manager knows how they use it, and whether they can name the written rule their use sits inside. The distances between those answers are the measurement.

Should we make AI disclosure mandatory?

Yes, with a clear trigger, and not first. A defensible trigger is that disclosure is required when a substantial portion of the work was AI generated, when it may be relied on for decisions, and when errors could affect those decisions. But introduce it only after leaders are visibly using AI themselves and the employment position on jobs has been published, or you will produce narrow compliance and deeper concealment.

Do company values statements change culture?

The evidence says no, on their own. Research across 562 companies and 1.2 million employee reviews found no correlation between the values a company publishes and how employees rate it on those values, with four of nine values correlating negatively.

Is low AI adoption a culture problem or a technology problem?

Both exist and they look different. Low actual use points to a delivery problem: generic outputs, poor workflow fit, missing business context. High actual use with low disclosure points to a culture problem. The diagnostic separates them, and treating one as the other wastes a year.

Who owns AI culture change?

Ownership splits by condition. Visible leader practice sits with the CEO and executives, the employment deal with the CEO and CHRO, the disclosure rule with risk and legal alongside the line business, protected practice with line managers, and the reward system with the CHRO and finance. Assigning the whole thing to HR or to a communications team is the most common way it fails.

How long does it take to close a disclosure gap?

Expect to see movement at six months rather than three, and re-measure on that cycle. Disclosure behaviour lags the conditions that produce it, so an early re-measure mostly captures a survey effect rather than a change in behaviour.

How MASSIVUE works on this

MASSIVUE is an enterprise transformation firm based in Singapore. We operationalise target operating models with client teams rather than handing over a recommendation, and the culture work described here sits inside that. Our approach on this topic combines a cultural diagnostic, an invitation to co-create rather than a cascade, and a systemic roadmap to rollout, which is the structure the 90 day sequence above follows.

The relevant service lines are AI Workforce Transformation, which covers AI skills assessment, role design and capability building, and Enterprise Transformation. Where the underlying issue is that nobody holds the decision rights over AI, Protum, our AI operating model framework, is the relevant starting point instead.

On the capability side, the MASSIVUE Academy microcredential AI Change Management: Upskilling and Reskilling covers diagnosing rational, emotional and structural resistance separately, psychological safety during transition, and designing an AI champions network. It sits at Associate level within the Certified Associate in AI and Digital Transformation certification, and it is the most directly useful next step for the manager layer this article identifies as the main lever.

Related reading

• How Do You Manage Change During Enterprise AI Transformation? covers the wider change programme this culture work sits inside, including skills, resistance and measurement.
• What Is an AI Operating Model? sets out who holds decision rights over AI, which is the authority structure conditions 2 and 3 depend on.
• High-Performing Teams in the Age of AI covers accountability for agent output at team level.
• End-to-End Change Management in Singapore covers the Singapore regulatory and funding layer in more depth.

Sources

Each figure cited above was checked against the publisher's own material at the last review of this article. Where a source is small sample, narrowly scoped or lagging, that is stated at the point of use.

• Nicole Gillespie, Steve Lockey and others, Trust, attitudes and use of artificial intelligence: A global study 2025, University of Melbourne and KPMG, April 2025 (more than 48,000 people, 47 countries, fieldwork November 2024 to January 2025). https://kpmg.com/xx/en/our-insights/ai-and-technology/trust-attitudes-and-use-of-ai.html
• KPMG and University of Melbourne, Global study reveals trust of AI remains a critical challenge, press release, April 2025. https://kpmg.com/xx/en/media/press-releases/2025/04/trust-of-ai-remains-a-critical-challenge.html
• 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). https://www.sciencedirect.com/science/article/pii/S0749597825000172
• Donald Sull, Stefano Turconi and Charles Sull, When It Comes to Culture, Does Your Company Walk the Talk?, MIT Sloan Management Review, 21 July 2020 (562 companies, 1.2 million Glassdoor reviews). https://sloanreview.mit.edu/article/when-it-comes-to-culture-does-your-company-walk-the-talk/
• Microsoft, 2026 Work Trend Index Annual Report, 5 May 2026 (20,000 knowledge workers, 10 markets, fieldwork 18 February to 7 April 2026, conducted by Edelman Data x Intelligence). https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization
• Slack, Half of Workers in Singapore Feel Uncomfortable Admitting AI Usage at Work, Says New Slack Workforce Index, 13 December 2024 (17,372 desk workers including 1,008 in Singapore, fielded 2 to 30 August 2024, administered by Qualtrics). https://www.salesforce.com/ap/news/press-releases/2024/12/13/half-of-workers-in-singapore-feel-uncomfortable-admitting-ai-usage-at-work-says-new-slack-workforce-index/
• Salesforce and YouGov, Singapore workers among world's least AI-sceptical, yet lowest in daily workplace adoption, 8 July 2026 (more than 1,500 desk workers across 14 markets, fieldwork December 2025 to January 2026). https://www.salesforce.com/ap/news/press-releases/2026/07/08/singapore-workers-among-worlds-least-ai-sceptical-yet-lowest-in-daily-workplace-adoption/
• Charu A. Chandrasekhar, Avi Gesser, Karen Levy and William Sadd, Why Companies Should Consider Requiring Internal Disclosure of AI Use, Debevoise Data Blog, 22 February 2026. https://www.debevoisedatablog.com/2026/02/22/why-companies-should-consider-requiring-internal-disclosure-of-ai-use/