ai-hallucination-fact-check-protocol

Use source reconstruction: search Google Scholar for the cited author, journal, and year to confirm the paper exists, then read the abstract to verify it says what the AI claims. The protocol provides step-by-step verification moves for citations, statistics, and expert quotes.

SIFT stands for Stop, Investigate the source, Find better coverage, and Trace claims, developed by Caulfield (2019) from lateral reading research. This protocol adapts it for AI output by replacing source investigation with claim-type identification and source reconstruction.

Common types include fabricated citations with real authors and journals, invented statistics with plausible precision, real studies misattributed to wrong authors or years, and false consensus claims. The taxonomy is based on Ji et al.'s (2023) survey of hallucination in natural language generation.

Yes, the protocol adapts its hallucination taxonomy to the subject area provided, since hallucination patterns differ by discipline. It is less reliable for niche or cutting-edge topics where hallucination rates are higher and verification sources are scarcer.

Full verification takes 3-5 minutes per citation and requires database access, so exhaustive checking is impractical. Subtle errors like real studies applied to wrong populations require reading methods sections, which may exceed students' independent academic reading skills.