Hallucination rates are down and still not zero, and zero is the only number that matters when you cite something. A citation slot gets filled like any other slot, with the most plausible string, so invented DOIs and author names look immaculate. Grounding on a paper index helps most; audited legal research tools still hallucinated on a third of queries.
checked 2026-09-21 · by dan scutaru
Why not yet: Generation is probabilistic, so a model with nothing to say still emits its most likely next token, and retrieval narrows that risk without ever removing it.
What would change the verdict: This flips to yes when a tool will refuse rather than guess: showing per sentence which source supports it and abstaining when none does, which is a product decision about being boring more than a modelling breakthrough.
People keep asking for this: “No AI is completely hallucination free yet”
the tools that earn the verdict — no pay-to-win