A Third of Workers Are Hiding What They Know From the AI They Were Told to Train

There is a standoff happening quietly inside a lot of companies right now, and Fast Company named it plainly when it reported that more than a third of workers say they are hoarding expertise because they fear being replaced by the AI agents they are being asked to train. If you have been asked to document your process, narrate your decisions or sit through a workflow recording, you are inside that story. The short answer to what is going on is that your uncodified knowledge has become an input to a system your employer owns, and workers have started to notice. At Metaintro, we follow this because it changes the thing your job security has always quietly rested on, which is knowing something your employer cannot easily replace.

What is actually being collected from workers?

Start with the mechanics, because they are less abstract than the debate around them. This spring Meta told thousands of its US employees, according to Fast Company, that software on their work computers would begin capturing mouse movements, clicks and keystrokes, and said it needed the data to train its AI agents on real examples of how people work. That is not a survey or a knowledge base. It is a recording of the act of working.

Nor is it unusual. Any company trying to automate workflows needs access to how humans actually perform those tasks. McKinsey built its internal generative AI, Lilli, by drawing on expertise developed across the firm's history, and the director of design for the tool made the logic explicit when she said that for most of McKinsey's history its knowledge sat with its experts. Lilli was the attempt to take that knowledge out of the experts and make it an organisation wide resource. Said that way it sounds like good knowledge management. Said from the expert's chair it sounds like something else. The same bargain is already running openly where companies ask workers to train the AI that replaces them, and it runs hardest in the low paid jobs where refusing is not an option.

Why do workers think the fear is justified?

Because of what happened next at the company that started the recording. In May, a month after news of the click tracking broke, Meta laid off 8,000 people and gave AI as the reason. More broadly, more than 175,000 people have been laid off in tech in 2026, and the main reason given has been the rapidly improving capability of AI.

You do not need a conspiracy to explain the resulting behaviour. You need a sequence that employees can see. Expertise is captured, capability improves, headcount falls. Whether or not any individual worker's recorded clicks contributed to any individual redundancy, the pattern is legible enough to change how people behave. The Fast Company piece put the incentive problem in one line, describing the request as being like asking a turkey to vote for Christmas. The same sequence played out in Oracle's restructuring and in Heineken's decision to cut roles once AI had absorbed the work.

How widespread is the recording actually?

Wide enough that most workers will meet it, and spreading fastest exactly where the expertise is expensive. The US Census Bureau puts the national AI use rate among businesses at 19.8 percent as of 3 May 2026, with adoption running between 17 and 20 percent since December 2025. The figure comes from the Business Trends and Outlook Survey, which was revised in November 2025 to ask about AI use in any business function rather than only in production. The averages hide the part that matters. Among firms with at least 250 employees the rate is 37 percent, and in the information sector it reaches 39.7 percent against 33.9 percent in finance and insurance. Those are precisely the large, knowledge intensive employers where a career is built on knowing things other people do not.

The Federal Reserve reaches the same conclusion from the worker's side, estimating that about 54 percent of the workforce is employed at firms that use large language models, and that work related use by workers stands at about 41 percent. Put those two findings together and the picture is clear. Roughly half of employees already work somewhere that runs these systems, so the question of what your employer is allowed to record from you is not hypothetical and is not confined to technology companies.

Is there real evidence that workers hold back, or is this just a mood?

This is the part that lifts the story above sentiment, and it is worth knowing the difference. Researchers at the Harvard Business School AI Institute tested it directly in a working paper titled Labor as Capital, AI and the Ownership of Expertise, co written by HBS AI Institute Associate Zoe Cullen. Rather than asking people how they felt, the authors ran a randomised experiment alongside survey evidence and formal economic theory.

The design was simple. Among 971 participants recruited through the research platform Prolific, half were shown a video explaining that AI systems can be trained on records of worker behaviour, including their own survey responses. The other half watched an equivalent video that said nothing about AI training. Then both groups were offered a 10 dollar bonus to share their past survey data. Among those who had been told how the data could be used, 41 percent refused, against 25 percent in the control group. The same informed workers were also significantly less willing to take part in future surveys at their existing wage.

Read that carefully, because the finding is not that workers dislike AI. It is that telling workers the truth about what their data does changes what they will sell it for, and by how much. Awareness alone moved refusal by sixteen percentage points and pushed up the price of participation.

What kind of knowledge is actually at stake?

Not the documented kind. The research describes workers holding large amounts of valuable uncodified knowledge, meaning tacit know how, judgment, communication skill and context specific understanding of customers, projects and processes that is not fully captured in manuals or company wikis. That is the material that never made it into the handbook because nobody could write it down, and it is precisely what behavioural recording is designed to capture.

The authors also found that workers believe they have substantial control over how much of that knowledge becomes visible, whether by documenting more or less carefully, by moving conversations on or off monitored channels, or by changing how they participate in surveilled workflows. That perceived control is the mechanism behind the hoarding. If you cannot stop the recording, you can still decide how much of your reasoning shows up inside it. That calculation gets sharper where AI usage already feeds into performance reviews, and sharper still where managers are pasting employee data into public AI tools with no policy behind them.

Does hoarding actually protect you?

Individually, for a while. Collectively, no, and the research is unusually blunt about the trade. In the authors' model, workers recognise that the knowledge they reveal today can improve the firm by helping build AI that substitutes for their expertise, and anticipating weaker future bargaining power they withhold knowledge in the present. That withholding is individually rational but collectively costly, because it reduces productivity and limits the quality of the AI systems firms can build.

There is a further trap in the obvious solution. Workers surveyed favoured individual ownership of work data, meaning the right to control and sell their own. The authors show why that instinct backfires. Because each worker's knowledge can substitute for another's, every individual sale strengthens the firm's bargaining position against everyone else who has not sold. Collective ownership removes the problem, since when workers bargain jointly and their knowledge is bundled, one person's contribution no longer undermines another's position. The paper's conclusion is that collective bargaining over work data eliminates the externality and can achieve both efficient knowledge sharing and a more equitable division of the gains from AI.

What does this mean for your career?

Three practical moves, and none of them is quiet sabotage.

First, understand that refusing alone mostly costs you. The research is explicit that individual withholding does not fix the bargaining problem, it just lowers output. If your workplace is capturing process data, the leverage sits in what your team or your representative body negotiates, not in how vague you make your documentation. Where a union or works council exists, work data is now a bargaining subject.

Second, move your value to the places recording cannot reach. The PwC data makes the case for where. Analysing more than one billion job advertisements across 27 countries, PwC found AI increasing the emphasis on human skills such as judgement, creativity and leadership, with the average wage premium for AI skills rising to 62 percent. Its two track finding is the useful one. Professionalised roles, where AI automates routine tasks so human judgement is emphasised, are seeing twice the growth in available jobs and 42 percent faster salary growth than roles democratised by AI, where the job itself becomes easier for non experts to do. A recorded click stream captures the second kind of work far better than the first. Those are the same capabilities in the set of human skills AI cannot replace, which is where the practical response starts.

Third, price the ask honestly if it comes to you directly. The experiment showed informed workers demanded more to participate at all. If you are being asked to do genuine knowledge transfer, that is project work with a business value attached, and it is reasonable to have it recognised in scope, title or pay rather than absorbed as an invisible extra duty. Before you accept a role built around this kind of work, there are specific questions worth asking about an AI job that are worth putting in writing.

Where does the standoff go next?

Two directions, and they are already visible. The first is regulation of the recording itself. Legislators have started treating workplace monitoring as a distinct category rather than a subset of data protection. Colorado moved to outlaw wage setting based on AI surveillance, and there is a parallel push against emotion detection in white collar workplaces. Where those rules land determines how much of the pipeline described in the research is even lawful.

The second is worker response that is neither quiet hoarding nor acceptance. Resistance to mandated AI is already reshaping hiring and retention, visible in the backlash against forced adoption at work and in flashpoints such as the response to performance monitoring at Duolingo. The research gives that response an economic argument it previously lacked. Withholding is not framed as technophobia in the Harvard paper, it is framed as a predictable answer to a bargaining problem.

For workers in the middle of this, the honest read is that the standoff is unresolved and your employer probably has not thought it through either. The paper's own framing is that under the current default, worker awareness slows adoption not because people dislike AI but because workers have reason to protect themselves. That is a solvable problem, and it gets solved through terms rather than through silence. If your organisation is capturing process data with no stated position on what happens to the resulting model, that gap is worth naming early, while it is still a policy question rather than a redundancy consultation. The roles furthest into this are the high paying knowledge jobs AI is disrupting first, and what an employer announces rarely says so plainly, which is why the corporate euphemism decoder is worth keeping to hand.

What should employers take from the research?

That the cheapest input in their AI programme is the one they are least likely to secure by mandate. The finding that awareness alone raised refusal from 25 percent to 41 percent means the transfer depends on consent that can be withdrawn quietly and invisibly, through vagueness rather than refusal. You cannot audit for reasoning somebody chose not to explain.

The governance answer in the paper is uncomfortable but coherent. If knowledge sharing is what makes workplace AI work, and workers withhold when they expect to lose from it, then the systems that share the gains are also the systems that produce the better model. That reframes data bargaining as an engineering input rather than an employee relations concession. While this gets worked out, the durable roles are the ones on the list of jobs AI cannot replace and what each pays, and the evidence you carry between employers is moving toward a career passport rather than a resume.

Related Articles

• Most Big Employers Already Watch You and That Is the Real AI Job Risk
• AI Workplace Surveillance Is Quietly Breaking How We Work
• AI Agents Are Not Your New Coworkers and Research Maps the Failure Modes
• CEOs Say AI Use Is Mandatory But Employees Do Not Agree
• The AI Comfort Trap and How Automation Erodes the Skills That Make You Hireable
• The Performance Divide Between AI Power Users and Newcomers
• AI Monitoring Software Flags Short Breaks and Whether That Goes Too Far
• Companies Cannot Track Their Own AI Agents and the Jobs That Gap Creates
• The Skills You Need to Thrive Alongside an AI Coworker
• Why Most AI Rollouts Fail Before They Ship
• AI Agent Projects Are Failing and Companies Are Rehiring Graduates
• Can Your Boss Legally Dock Your Pay When You Show Up Late

People Also Asked

Q: Can my employer legally record how I work to train AI?

A: In most jurisdictions employers have wide latitude to monitor activity on company owned equipment, which is how the practice described by Fast Company operates, with software capturing mouse movements, clicks and keystrokes on work computers. The Harvard Business School research treats the legal permission as settled and argues the live question is governance, specifically whether the resulting data is owned individually or bargained over collectively. Check your own contract and local rules, since notice requirements vary widely by country.

Q: Is withholding my expertise a good way to protect my job?

A: The research suggests it protects you less than it feels like it does. The authors describe withholding as individually rational but collectively costly, reducing productivity and limiting the quality of the systems firms build, without resolving the underlying bargaining problem. They point to collective bargaining over work data as the mechanism that achieves both efficient knowledge sharing and a fairer split of the gains.

Q: Which skills are hardest for this kind of recording to capture?

A: Judgement and context. The research describes the valuable material as uncodified knowledge, meaning tacit know how, judgment, communication skill and context specific understanding of customers and projects that is not captured in manuals or wikis. PwC's analysis points the same way, finding AI increasing the emphasis on human skills such as judgement, creativity and leadership, with professionalised roles seeing twice the job growth of democratised ones.

Future proof your career by understanding where your value actually sits before someone else codifies it. Metaintro tracks how AI adoption, monitoring and pay are reshaping real roles, so you can see which way your job is moving while you still have options. Create your free Metaintro profile and get matched to roles that pay for judgement rather than throughput.