An AI answer cites quit-smoking research: reconstruct the source before using it

01

Separate the answer's sentence from its bibliography

Copy the public claim and its attached citation into two separate notes. Search the title and author independently in a publisher record or relevant scholarly catalogue, then compare title, authors, year and identifier. A DOI that opens a different paper is a mismatch, not verification. If a title is translated, try the supplied original title as well.

A failed search may reflect a typo, incomplete record or access limit. Say ‘not located with these searches’ rather than ‘proved nonexistent’. Asking the same AI to supply another reference may help produce a search lead, but it cannot replace opening and reading the work.

02

An invented answer with three different problems

The following answer and all three source situations are fictional; there are no real titles, authors or DOI values. The answer says: ‘A digital diary reduces return to smoking after six months,’ followed by three references. This is the assertion to investigate, not a finding supplied by this page.

Reference A cannot be located from its title or author. Retain its unverified status and do not use it as supporting evidence. Reference B leads to a one-week diary evaluation: participants rated ease of use, with no comparison group and no six-month smoking outcome. You can describe that evaluation, but cannot turn usability into reduced return to smoking.

Reference C opens a public agency's general page about getting cessation support. It does not compare the diary or report the claimed outcome. The agency's standing does not make that specific diary claim its recommendation. Three citations cannot be pooled into one missing result: an unlocated record, a different outcome and an unrelated page leave this assertion unsupported by the supplied references.

03

Keep the useful part without preserving the unsupported effect

ICMJE requires human review of AI output. NIH's research-integrity reminder specifically warns about nonexistent AI-generated references.

For the invented answer, retain the question you want to investigate and the limited usability report, if accurately described. Remove the six-month effect claim; merely adding ‘may’ would still attribute an unmeasured effect to these sources. This does not establish that a diary can never be useful or that all AI answers are false.

A replacement note can say: ‘The supplied references do not establish the six-month claim. One is unlocated, one reports one-week usability without a comparator, and one gives general support information. I have not found the claimed smoking outcome in them.’ New evidence requires a new reading, not a refreshed AI confidence score.

[2][3]

04

Verify public evidence without entering personal health details

You can trace a public citation with its bibliographic details alone. Do not upload private medical records, smoking diaries, names or a friend's messages to get the model to judge them. If a source is inaccessible, state that limit instead of supplying personal detail to fill the gap.

For your own support or treatment decisions, take the public question to qualified local health or cessation professionals. A source-reading exercise does not decide a personal treatment or authorize changes to prescribed care.

05

What to keep in mind

• Identify the source separately from the generated answer.
• A located source must still measure the claimed outcome.
• Several mismatched references do not add up to evidence.

06

Common questions