By Uwe Peters.
Most scientists strive to communicate their findings as accurately as possible. But our latest research suggests that they may sometimes encourage exaggerated interpretations without intending to, not because they overstate their findings, but because their audience understands seemingly ordinary scientific language differently.
Uwe Peters holds an MSc in Psychology and Neuroscience of Mental Health and a PhD in Philosophy from King's College London, UK. He is an Assistant Professor at Utrecht University in the Netherlands, where he teaches philosophy of science and AI. Previously, he was a postdoctoral researcher at the University of Cambridge at the Centre for the Future of Intelligence. His research focuses on meta-science, science communication, and the societal implications of AI.
 |
| Source: Pixabay. |
Most scientists strive to communicate their findings as accurately as possible. But our latest research suggests that they may sometimes encourage exaggerated interpretations without intending to, not because they overstate their findings, but because their audience understands seemingly ordinary scientific language differently.
For instance, if you read sentences such as “Statins reduce cardiovascular events" or "People benefit from a Mediterranean diet,” you would probably understand them immediately. They are short and clear. They also share an important linguistic feature: they make claims about whole categories (statins, people) without saying how many people or situations they apply to.
Linguists call these statements generics. Scientists use them very frequently (e.g., in biomedical research articles), as generics make scientific writing concise and readable. But when scientists write these sentences, do readers understand them in the same way that scientists do?
The hidden ambiguity of generic language
Take the statement “Statins reduce cardiovascular events.” Does this mean statins work for everyone? Most patients? Patients with particular medical conditions? Or does it simply describe the outcome of one clinical trial involving a specific group of participants? The sentence itself doesn’t tell us.
Scientists rarely read these sentences in isolation. Years of scientific training encourage them to interpret them against the background of study populations, methodological limitations, external validity and previous evidence. These qualifications can thus remain implicit because they may seem obvious within the scientific community. However, readers outside the field may not make those same assumptions and so understand the sentences as much broader in scope.
Comparing scientists, the public and AI chatbots
In a recent study, we asked more than 400 people, including laypeople, psychologists, biomedical researchers and other experts, to evaluate scientific conclusions taken from published research. Based on our previous finding that AI chatbots like ChatGPT also extensively used generics in science summaries, we additionally presented the same material to ChatGPT-5 and DeepSeek to see if humans and AI differ in their interpretation of generics.
Participants rated each conclusion on three dimensions: How broadly does this finding apply to people (e.g., only to study participants, to most people, or everyone)? How credible is it? How impactful is it? Some participants saw conclusions written as generics. Others saw equivalent conclusions written without generics, for instance, in the past tense. Because participants evaluated otherwise identical findings presented in different forms, we could isolate the effect of generic language itself.
Our main research question was simple: Do scientists, laypeople, and AI systems actually understand generic statements routinely found in science communication in the same way?
The same sentence, different interpretations
The results revealed a consistent pattern. Scientific experts interpreted generic conclusions more cautiously than laypeople. They judged the same statements as applying much less broadly and, in many cases, as less credible.
This does not mean that laypeople misunderstood the science. Rather, scientists and non-scientists appear to bring different background assumptions to the same sentence. Scientists often mentally supply qualifications that remain unstated, whereas lay readers are less likely to do so. Successful communication depends on speakers and audiences sharing these assumptions, yet our findings suggest this common ground cannot always be taken for granted.
Scientists may believe they are communicating cautiously, while readers interpret their conclusions more broadly than intended. Generic language can thus inadvertently encourage exaggerated interpretations of scientific findings in laypeople.
We expected some differences between scientists and laypeople. What surprised us was that the AI systems went further still. Both ChatGPT-5 and DeepSeek interpreted the generic conclusions as more generalizable, more credible, and more impactful than even the laypeople. If AI systems systematically favor broader interpretations of generics and therefore also generate similarly broad language when summarizing research, they may amplify an interpretive mismatch that already exists between scientists and the public.
What should science communicators do?
Our findings highlight an important challenge in science communication. Scientists may often implicitly assume that readers interpret their generic statements in much the same way they do when this assumption is in fact often unwarranted. Scientists tended to interpret generic conclusions more cautiously than lay readers, while AI systems interpreted them more broadly still. Consequently, the same scientific statement may communicate different messages to different audiences.
Science communication is often discussed in terms of explicit misinformation, trust and scientific literacy. Our findings suggest that another factor deserves attention, namely the subtle linguistic choices that shape how research is interpreted. Generics may invite incorrect interpretations because different audiences infer different levels of generality from the same wording. To reduce miscommunication risks, scientists or science communicators, more generally, should not assume that these inferences are shared and, where appropriate, make the intended scope of a finding more explicit, particularly when communicating beyond specialist audiences.
As AI systems increasingly become intermediaries between researchers and the public, opportunities for overgeneralization may accumulate. Scientific claims may be interpreted more broadly by AI systems than by scientists, and broader still as they reach wider audiences, creating opportunities for exaggerated interpretations even when the original research was communicated in good faith.
Improving public understanding of science may therefore depend not only on communicating accurate evidence, but also on recognizing that different audiences can infer different messages from exactly the same scientific sentence, especially when it involves a generic.
Read the original research: Generics in science communication: Misaligned interpretations across laypeople, scientists, and large language models.
---