Scientists increasingly use AI to inspect data, draft code, identify patterns and formulate questions. Those tools may save time or make certain work possible. But scientific research depends on something not produced by a model: public legitimacy. Taxpayers finance laboratories, volunteers provide data and patients or communities accept the consequences of applied findings. A new UK survey suggests that the public distinguishes science done with AI from science done without it—and that the distinction can materially affect its willingness to support the work.[1]
Researchers Christian Bretter, Boryana Todorova, Caitlin Curtis and Matthew Hornsey report a preregistered, nationally representative survey of 1,465 UK respondents in Research Policy. In the reported results, 77% supported public funding for scientific research in general. Support was 42% when generative AI was part of the research description. This 35-percentage-point difference is an attitude measured under the study’s survey conditions, not evidence that public science budgets will fall by 35%.[1]
Another result complicates the first. Between 83% and 85% of respondents supported scientists using generative AI when substantial human input remained part of the process. The response was not an unconditional rejection of computational assistance. It appears to turn partly on the role assigned to the technology and the amount of visible human responsibility.[1]
The legitimacy problem is not the same as the capability problem
A tool may generate a scientifically useful result while the public worries about who is accountable for it. If a model proposes a molecule, a human chemist still needs to establish what was tested, which controls were used and how the analysis can be repeated. The survey’s patterns are compatible with a preference for that human chain of responsibility, although the published abstract cannot establish each respondent’s reasoning in detail.[1]
The researchers also report that 34%–40% said they would trust scientific findings and scientists less if generative AI was used in the research process. They identify stronger funding penalties in some respondent groups, including younger and more scientifically literate participants. These are associations observed in a survey, not a demonstration that age or literacy causes mistrust.[1]
How public acceptance differs by context
| Survey question or measure | Reported UK result |
|---|---|
| Support for publicly funded scientific research | 77% in the general research conditionSource 1 |
| Support when generative AI is involved | 42% under the described survey conditionSource 1 |
| Less trust in research or scientists if AI was used | 34%–40%, depending on the trust questionSource 1 |
| Support with substantial human contribution | 83%–85%, depending on wordingSource 1 |
| Survey population | 1,465 respondents in a preregistered nationally representative UK sampleSource 1 |
One possible interpretation is that some people treat AI assistance as a sign that public funds are being used to replace judgment with automation. Another is that they imagine less effort, weaker originality or more fabricated findings. There may also be concern about ownership of training data and who benefits from commercial models built on public research. The survey result alone does not decide which explanation dominates; that requires experimental work and detailed qualitative evidence.[1]
It is tempting for scientific institutions to respond by minimizing disclosures of AI use. That would likely solve the wrong problem. If the public values substantial human input, the credible answer is to make that input demonstrable—show what was generated, what was independently checked and which decisions remained with scientists. A generic statement that 'AI was used' tells a reader almost nothing about the quality of the work.[1]
Disclosure requires more than a binary label
Consider three activities: a model checks code for simple errors; a model proposes hypotheses for a scientist to test; a model generates an entire manuscript without conducting the experiments it claims. A one-line disclosure could classify all three as 'AI-assisted research', while their implications for evidence are radically different.[1]
More useful reporting would distinguish assistance with wording, analysis of existing data, experimental design, autonomous execution and verification. Publications can also make data, code and evaluation protocols available where ethically and legally appropriate. These practices help qualified outsiders reproduce findings. They are worthwhile even if no survey had detected a trust penalty.[1]
There are limitations to the present evidence. The findings concern stated attitudes among survey respondents, not actual allocation decisions by grant panels or legislators. The published abstract does not support a causal claim about which institutional transparency measures would restore trust. The journal record carries a December 2026 issue date; the publicly accessible study abstract is available as of this reporting, but an issue label is not the date on which participants were surveyed.[1]
The difference between using AI and handing science over to it
The apparent contradiction—lower general support for AI-involved research but substantial support for human-guided AI research—may be the most important result. It suggests that people are reacting not merely to a technology label but to an imagined distribution of labour, expertise and responsibility.[1]
A useful policy debate should therefore ask which AI tasks genuinely expand scientific capacity and which weaken the ability to challenge a finding. If researchers can explain that boundary, trust can be argued for on evidence. If they cannot, stronger models and faster papers may leave a more fundamental question unanswered: who is answerable when the science turns out to be wrong?[1]
Sources & further reading
- Public views on the use of AI in scientific research Bretter, Todorova, Curtis and Hornsey, Research Policy · Accessed 2 Nov 2026