The invented reference, and why it is so plausible
Lesson 1 of 6 · Citations that survive
1 · Learn the move · Constraints + negative instructions
A model generates a citation the way it generates any sentence - by what is plausible next. Authors who publish in your field, a journal that exists, a volume number in the right range, a DOI with a real registrant prefix. Every element is drawn from the distribution of real citations, which is exactly why an invented one reads as real. The defense is not sharper eyes; it is a rule. No reference exists in your notes without a verification state: VERIFIED, UNVERIFIED, or FAILED. Every reference a model produces is born UNVERIFIED, and an ambiguous check fails toward UNVERIFIED, never toward VERIFIED. A citation is a claim you are making in print. The state records whether you have actually checked it.
Suggest candidate references for [claim or topic]. Hard rules: (1) label every reference UNVERIFIED - you cannot confirm that any of them exist, (2) for each one, list the elements I must check: DOI resolution, exact title, full author list, journal, year, (3) never present a reference as confirmed, and if I ask whether one is real, say you cannot know and keep it UNVERIFIED.
2 · Your turn. You write the prompt
You are drafting the introduction for a paper on detergent selection in membrane protein purification and asked a model for supporting references. It returned five, with authors you recognize, plausible titles, and well-formed DOIs. They look better than the ones you found yourself, and your co-author wants to paste them into the draft tonight.
Remember: the AI sees only your prompt, not this page. If the situation isn't in your prompt, it doesn't exist.
Optional. These shape the output when you run your prompt below, not your score.