AI sources are not citations: how to treat model outputs responsibly

AI tools can explain ideas quickly, but they are not a primary source. Treat an AI answer as a lead to investigate, not evidence to cite.

In research workflows, “citation” is shorthand for accountability: the reader should be able to open the referenced work, evaluate it, and see whether it supports your claim. Model outputs do not provide that chain of accountability. They are generated text, not an archival record.

A simple rule: cite what you can open and verify

• If a claim matters, locate the underlying paper, dataset, spec, or official documentation.
• Prefer stable identifiers (DOI/PMID/ISBN) over “a link someone pasted.”
• If you cannot locate the source reliably, treat the claim as unverified.

A clean workflow: use AI for direction, not authority

• Step 1: Ask for a search plan (keywords, venues, authors), not “the answer.”
• Step 2: Collect candidate sources using identifiers and official indexes.
• Step 3: Verify the source exists and the metadata matches what you expect.
• Step 4: Read the relevant section and confirm the claim is actually supported.
• Step 5: Cite the primary work. If needed, cite a review paper for context, not as evidence for a specific result.

Common failure modes (and how to spot them)

Where AI goes wrong (even when it sounds confident)

• Plausible citations that do not exist (hallucinated identifiers).
• Real papers attached to the wrong claim (citation drift / overclaiming).
• Outdated information presented without context or timestamps.

• “Title looks right, authors are off”: often a mash-up of multiple real papers.
• “Journal name is almost right”: look for venue/series confusion or fabricated issue numbers.
• “The DOI resolves but the claim doesn’t”: the identifier is real, but the conclusion is not in the paper.
• “The paper exists but not at that URL”: prefer DOI/PMID over arbitrary links.

How to write responsibly when evidence is incomplete

If you can’t verify a claim quickly, you can still be accurate by changing the sentence. Replace certainty with a bounded statement: what you checked, what you did not, and what remains unknown.

• Instead of “X causes Y,” write “Some reports suggest X may be associated with Y; we did not verify the primary studies.”
• Instead of “The paper says…,” write “We could not confirm the paper; treat this as unverified.”
• Record timestamps and versions for fast-moving topics (policies, software, security guidance).