Research Brief: Workers are less willing to share expertise with AI

In a study of 61 professionals, people were less willing to let AI reuse their expertise than colleagues. They cited lost control and replacement.

Research · Luminary Research Brief · 12 October 2026 · 3 min read

Workers are less willing to let an AI system reuse their expertise than to let a colleague reuse it. That is the central result of a new preprint from researchers at the National University of Singapore and Emory University. If you plan to turn your best people’s know-how into reusable AI “agent skills” (packaged practices and decision rules that an AI draws on when doing similar work), the study suggests the first obstacle may be consent, not technology.

What the researchers did

The team ran a mixed-methods study (numerical measures plus open-ended answers) with 61 professionals. Participants completed simulated work tasks in occupational scenarios, using a purpose-built system with a large language model behind it. In doing so they created materials that contained their own domain knowledge.

Each participant then made a real decision: whether to authorise reuse of those materials, either by human professionals or by an AI system. Participants in the two conditions also answered questions about the consequences they expected and about how they would want sharing to be governed. The researchers examined who authorised reuse, which kinds of material they were willing to share, and how people explained their choices.

What they found

Participants were less likely to authorise AI reuse than human reuse. The authors describe this as the clearest difference: it concerned whether people allowed reuse at all. Evidence of differences in what they chose to share, at the level of material categories, was inconclusive.

The reasoning differed. Workers in both conditions worried about their knowledge being taken out of context. Those in the AI condition expressed greater concern about losing interpretive control, meaning they could no longer shape how their knowledge is read and applied. They also raised possible professional replacement. The abstract adds that, with humans, reasoning leaned on prosocial considerations such as helping colleagues, while with AI it shifted towards control, replacement and downstream governance.

Participants also wanted ongoing control over how their knowledge is used once shared. Preferences on credit were mixed: some wanted their name attached, others preferred anonymity. The paper also reports a theme of protection against adverse employment consequences.

What this means for your business

Do not assume goodwill transfers. A team that happily writes handbooks and mentors new starters may still hold back when the audience is an AI. The paper’s evidence is that the willingness to share, not the content, is what changes.

Treat consent as a continuing process. The authors recommend that knowledge systems preserve context and provenance, support review, correction, restrictions and later withdrawal, and accommodate different attribution preferences. For a small firm, that means practical steps: say which materials are going into an AI tool, let people see how their contribution is used, let them pull it back, and ask whether they want credit.

Address replacement directly. Workers raised fear of being replaced. Silence on this is likely to be read as confirmation. If your intention is to free people from repetitive work rather than reduce headcount, put that in writing and be prepared to be held to it.

Expect wariness about data generally. The ONS’s UK Business Data Survey 2025-26 found that among businesses handling digitised data, 73% would feel uncomfortable with their data being used to train external AI models, and 18% comfortable. That concerns company data, not employee expertise, so it is a different question from the one in this paper. It does show that unease about feeding AI is common at firm level as well as among individual staff.

You are probably early. The ONS Management and Expectations Survey reported AI adoption by only 9% of UK firms in 2023, with 39% citing difficulty identifying business use cases as a barrier. If you are still working out use cases, settle the rules on staff knowledge before you build anything on it.

Limits worth knowing

This is a preprint, so it has not been peer reviewed. Sixty-one participants is a small sample, and the paper’s own conclusion calls the category-level evidence inconclusive, so the finding on what people share should carry little weight. The tasks were simulated scenarios built for the study, not a live deployment inside a real employer, where pay, promotion and job security are at stake. Real stakes might make workers more cautious, or might make them more compliant. The authorisation decision was also a single choice, not the repeated negotiation a workplace involves.

I have worked from the abstract, introduction and conclusion, not the full results tables, so I cannot give the authorisation rates for each condition.

The direction of the finding is plausible and consistent with the wider evidence on data comfort. It deserves attention as a prompt to design consent into any knowledge-capture project. It does not give you a measured size of the effect for your own staff, which you should find out by asking them.

Source: Tianqi Song, Zicheng Zhu, Hancheng Cao, Yi-Chieh Lee, “When My Skill Becomes Agent Skill: How Knowledge Workers Share Their Expertise with AI Systems”, arXiv. This paper is a preprint and has not yet been peer reviewed.