A readership survey found researchers split on chatbot use

The writer's problem

A researcher deciding whether, and how, to say a chatbot helped write a manuscript wants to know how peers actually use these tools, not just what a journal policy permits. A Nature news feature by Richard Van Noorden and Jeffrey M. Perkel reports a survey built to answer exactly that, including a specific question about writing research manuscripts.

What the documents show

Nature's own headline states the survey covered 'more than 1,600 researchers around the world,' and its own published survey instrument asks respondents what they use generative AI tools for, listing 'To help write research manuscripts' as one selectable answer alongside reviewing manuscripts, coding and literature reviews. The survey's own methodology document, released as supplementary material, states that Nature journalists emailed more than 40,000 scientists who had published papers in the last four months of 2022, plus readers of the Nature Briefing; about 5 percent responded, yielding 2,728 responses, 1,801 complete, and 1,659 active-research respondents after further filtering, the basis for the rounded '1,600' headline. That same document states plainly that researchers interested in AI were more likely to respond, so the sample is not representative of scientists generally, a self-selected readership rather than a randomly sampled population.

The editorial choice

A journal or a writer citing this survey should say so as a readership survey with a roughly 5 percent response rate among an AI-interested skew, not as a representative measure of how researchers in general use AI in writing, a distinction the source document itself draws. Editorially, this survey is better used to describe what a self-selected, engaged slice of the research community reports doing than to estimate how common manuscript-writing AI use is across all of science.

What stays with the author

The survey reports what respondents say they do, not what any given manuscript's text actually contains, and it does not resolve what disclosure a specific journal should require. A researcher's own account of how a chatbot was used in drafting, and their standing behind the resulting text, is not something a readership survey can verify or replace.

• Does a self-selected, AI-interested sample overstate or understate how common manuscript-assistance use really is?
• What would a randomly sampled survey of the same question find instead?
• Does listing 'help write manuscripts' as one answer among many understate how central that use is for some respondents?

The feature and its released methodology together describe a large but self-selected readership survey, useful for what it documents about respondents' own reported habits, not a representative census of scientific writing practice.