Judgment in the Age of AI · Article 2
The more complete a citation looks, the easier it is to trust: author, year, journal, page and link all appear to be present. Yet a language model can combine a real author, a plausible title and invented publication details into a perfectly formatted false citation. Verification begins by leaving the AI answer itself.
Put the citation on a four-level source ladder
- Search the full title, not only the author. Failure to find the title is an immediate warning.
- Open the publisher, journal, university repository or government site and match author, year and title together.
- Find the DOI or another persistent identifier and make sure it resolves to the same work.
- Read the relevant abstract or passage to see whether the source actually supports the claim attributed to it.
Existence is not support
A common second-level error is not an invented paper but an exaggerated conclusion. A study reporting correlation may be presented as proof of causation; a small sample may become a universal consensus. The final check must therefore compare the AI sentence with what the source actually says.
Citation checking answers two questions: Is the source real, and does it support this claim?
If you can find only summaries, bibliography pages or other AI-generated material, mark the citation as unverified. Unverified references should not enter the final version of academic, medical, legal or commercial work.
References
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.