Confidence is not evidence
Language models generate text one token at a time based on statistical patterns, and that process produces the same fluent, assured tone whether the underlying claim is well-supported or completely fabricated. There is no built-in 'uncertainty voice' — a model stating a fake court case citation and a model stating a real one sound identical. This means the confidence of the delivery tells you nothing about the reliability of the content, and treating fluency as a proxy for accuracy is the single most common way people get burned by these tools.
Discernment starts with noticing which claims in a response are checkable and checking the ones that matter. Not every sentence needs a fact-check — if a model suggests a phrasing for a sentence, there's no external fact to verify. But specific numbers, quotes, citations, API behaviors, legal or medical claims, and anything you plan to act on should be treated as unverified until you've confirmed them against a source you trust, independent of the model that generated them.