Plain factual errors
The most ordinary failure is the everyday factual error: a wrong date, a misattributed quote, a chemical that does not react the way the model claims. These slip out most easily on the long tail of knowledge — obscure people, recent events, niche technical details — where the training data was thin or contradictory and the model never formed a reliable memory. On popular, heavily-repeated facts the model is usually solid; the danger zone is the specific and the rare.
A subtle trap is that errors cluster around things that look memorizable. People assume a model that writes flawless code must surely know a simple birth year — but coding skill and trivia recall are different capacities, and fluency in one says nothing about accuracy in the other.
Fabricated citations: the most seductive failure
Of all hallucinations, fabricated citations do the most damage to trust, because a citation is supposed to be the antidote to making things up. Ask for sources and a model may produce a perfectly formatted reference: real-sounding authors, a plausible journal, a clean volume and page range, even a DOI — all invented. The reason is the same next-token logic: the model has learned the form of a citation far more thoroughly than the contents of any particular paper, so it generates a convincing shell with nothing inside.
The structural fix is to stop asking the model to recall sources from memory and instead make it work from real documents you supply — the idea behind grounding the answer in retrieved text and attaching citations that point to that text. Grounding reduces fabrication sharply, but it does not eliminate it: a model can still misread or over-claim from a real document, so the citation must point to text you can actually check.
A pipeline that retrieves documents and feeds them into the model before it generates an answer.
Knowledge that has gone stale
Even a perfectly accurate model is frozen in time. Its weights encode the world as of its training data, so anything that happened after its knowledge cutoff is invisible to it — a new law, a product release, an election result, the current price of anything. This is knowledge staleness, and it is dangerous precisely because the model will answer questions about 'now' in the same confident tone it uses for settled history.
- Ask the model when its knowledge ends — but treat the answer cautiously, since the stated cutoff is itself sometimes wrong.
- For anything time-sensitive, prefer a setup that retrieves fresh information rather than relying on the model's memory.
- Be especially wary of 'as of today' phrasing in the model's output — it has no clock and is guessing at the present.
Why you cannot read confidence off the tone
All three failures share the confident-wrongness signature from the last guide: nothing in the wording tells you which sentences are solid and which are scaffolding. There are partial signals — you can inspect the model's token-level probabilities, where a low-probability burst sometimes flags a shaky stretch — but this is noisy and easy to over-trust. A more dependable instinct is categorical: facts, figures, names, and references are the kinds of output most likely to be confabulated, so verify them by default.
Every token is drawn from a softmax probability — fluent wording reflects a high token probability, not factual truth, so the confident tone tells you nothing.