·3 min read
A citation is not a source
Models will invent a page, a quote, a case. The product either refuses when the file is empty, or it is guessing in a nicer font.
Photo: Scott Graham on Unsplash
The embarrassing failure is not a clumsy sentence. It is a confident answer with a footnote that does not exist.
OpenAI’s write-up on why models hallucinate is useful because it treats invention as a training and evaluation problem, not a moral failing of the user. The product still has to decide what happens when the model is unsure.
A citation that cannot be opened is decoration. If the system prints “Source: week 4 notes” it should be able to show the paragraph. If it cannot, it should stop.
Eval work exists. The Vectara hallucination leaderboard is one public attempt to measure factual consistency in summaries. Use something like it on your own corpus, not a demo script.
Grounding reduces the space of allowed answers. It does not make the model honest. You still need refusal, quotes, and a human on the classes that can harm someone.
Do not score the assistant on how rarely it says “I don’t know.” That metric trains it to fill silence.
The user-facing test is rude and correct: click the citation. If the page is empty, the product is lying.
The TruFyre Way
A citation is not a source. TruFyre products only speak from the retrieved set and show the passage. If retrieval is empty, they refuse.