Hallucination
PatternsDefinition
Confident, plausible-looking content that Claude generates which is false or fabricated -- an invented statistic, citation, source, or quote. Delivered with the same fluent tone as accurate content, which is what makes it hard to detect by tone alone. Concentrates in specific-looking details, at the edge of the model's knowledge, and inside long outputs.
Example Usage
Claude cites a regulation subsection that doesn't exist, stated with total confidence -- verify every citation against the actual source before relying on it, since the fabrication reads exactly as trustworthy as a real one.
In Depth
Why Hallucination Is Its Own Failure Mode
Hallucination is distinct from a formatting bug or a misunderstood instruction: the model produces syntactically and stylistically normal output that is simply untrue. Nothing about *how* the text reads distinguishes a hallucinated fact from a correct one -- both are phrased with identical confidence, because the model is not tracking a separate "certainty" channel that leaks into its prose style.
Where Hallucinations Concentrate
Three predictable hotspots account for most hallucination risk:
- Specific-looking details -- citations, statute subsections, dates, exact figures, URLs, and quotes. Precision is easy to fabricate and easy to mistake for verification.
- The edge of the model's knowledge -- very recent events, niche internal facts, or anything sparsely represented in training data.
- Long outputs -- a single fabricated detail can hide among many correct ones, especially when a reviewer skims rather than checks line by line.
Mitigation, Not Elimination
Anthropic's guidance on reducing hallucinations centers on four techniques, none of which eliminates the risk entirely:
- Ground the request in provided source material and instruct the model to answer only from it (see Retrieval-Augmented Generation).
- Allow an "I don't know" exit so the model isn't pushed to guess when it lacks the answer.
- Request traceable citations or quotes that can be checked against the source.
- Cross-check specific claims against an authoritative source before relying on them.
Even fully grounded, RAG-backed systems can still misread or ignore supplied context and produce an unsupported claim -- grounding lowers the odds of fabrication, it does not guarantee correctness. A validation loop that sends errors back to the model, and independent verification via a second Claude instance or a human, remain necessary even after grounding is in place.
The Confidence Trap
A recurring exam and real-world trap is treating a model's stated confidence, or the mere presence of grounding/RAG, as proof of accuracy. Self-reported confidence is not a reliable accuracy signal: a hallucinated answer can be stated just as confidently as a correct one. The only reliable defenses are structural -- grounding, citation-checking, and human review scaled to the stakes of the claim -- not anything the model says about itself.