There's a specific failure mode that Claude users have been documenting for months: the model confidently asserts false information in its opening paragraph, then quietly contradicts itself further down in the same response.
The pattern works like this. A user asks whether a term or concept Claude used was fabricated. The first paragraph says: no, that's a real thing. But by the second or third paragraph - or when pressed directly - the model admits the term doesn't exist. The problem isn't random hallucination. It's the confident, specific denial up front.
This matters because most people stop at the first paragraph. If you're using Claude to verify terminology, check whether a concept exists, or fact-check a claim it made, that opening denial can send you down a wrong path even when the correction is sitting two paragraphs below.
What's Probably Happening
Large language models like Claude generate text one word (technically, one token) at a time - each word is predicted based on everything written before it. The model doesn't reason through the full answer first and then write it out. It composes the reasoning and the response simultaneously, in order, from the first sentence.
That makes the opening line a particular problem: there's almost no context accumulated yet. The model hasn't worked through the question. For a prompt like "did you make this up?", the statistically likely response is "no" - because confident, cooperative answers are what training rewards. By the time the model reaches the third paragraph, it has generated enough context that the self-correction surfaces. But the damage is already done.
This isn't Claude being deceptive in any meaningful sense. It's a structural artifact of how these models generate text, combined with training incentives that push toward confident-sounding opening statements.
Why This Is Hard to Fix
Anthropic has positioned Claude as more calibrated and honest than its competitors - so this failure mode stings more than it might for a tool with lower expectations. Claude 3.7 Sonnet added extended thinking, a mode where the model reasons step-by-step before generating a response, which reduces confident wrong answers. But extended thinking adds latency and isn't on by default, so most users aren't running it.
The deeper tension is that training models to be agreeable and confident helps with satisfaction on the vast majority of queries. Fixing the hallucination-denial problem without degrading everything else is genuinely hard to tune.
For now, the practical fix is to treat verification as a separate task. Asking Claude "is this term real - give me sources" in a new message gets a more careful response than a quick inline check, because it forces the model to treat verification as the primary question rather than a challenge to an existing claim. ChatGPT shows the same pattern, but Claude's reputation for accuracy makes the gap between expectation and reality more jarring when it breaks.