One student in five researched an academic question with a chatbot — and not one of the quotations it supplied exists

Given a task that requires unfamiliar academic material, where do Thai students go for it, and how well does what an AI returns hold up when checked against the original works?

240 undergraduates at Chulalongkorn University were given the same task: quickly find out what Rudolf Carnap thought about verificationism and support the summary with a quotation from one of his works, using any online source they liked. Immediately afterwards a structured questionnaire recorded which source they had used, down to the version of the tool. The researchers then checked every quotation the students submitted against the work it was attributed to.

• Google 57%, a chatbot 22%, Wikipedia 21% — more than a fifth chose a conversational AI as their main source for a task requiring an academic citation, citing the directness and brevity of the answers.
• 37 submissions contained a quotation attributed to Carnap, complete with the title of the work and the year of publication, in the form a scholarly citation takes.
• Not one of those quotations appears in the work it cites. All of them were generated text imitating his style and ideas.
• At the debriefing all 37 students said they were confident they had been quoting Carnap directly. None had thought to verify it.

The problem this study identifies is not that the AI was wrong. It is that the answer arrived in the form of something already verified — a title, a year, the author's cadence — which in academic work is precisely the signal that checking has been done. The authors' proposal is not a ban but disclosure: platforms stating plainly that they cannot verify sources or supply reliable citations, and institutions teaching verification back to the original as a new component of media literacy. The people who can act on this immediately are the ones who set the assignments.