Hallucination: Confident but False or Fabricated Content
CoreRecognize hallucinations, inconsistencies, and bias in Claude's output · Difficulty 2/5
Explanation
What a Hallucination Is
A **Hallucination** is when Claude produces confident, plausible-looking content that is false or fabricated -- an invented statistic, a made-up citation or subsection number, a non-existent source, or a quote no one said. It is stated with the same fluent confidence as true content, which is exactly what makes it dangerous: nothing about *how* it reads distinguishes it from a correct answer.
Where Hallucinations Concentrate
Hallucinations are not randomly distributed across an output. They concentrate in predictable places:
- Specific-looking details: numbers, dates, citations, legal subsections, URLs, quotes -- anything that sounds precise is easy to fabricate and easy to mistake for verified fact.
- The edge of the model's knowledge: very recent events, niche internal facts, anything outside well-covered training data.
- Long outputs: a single fabricated detail can hide among many correct ones, especially in long-form answers reviewers skim rather than check line by line.
Why This Matters in High-Stakes Contexts
Because a Hallucination is delivered with the same confident tone as accurate content, and because it concentrates in exactly the specific-looking details (citations, figures) that high-stakes work depends on, it is most dangerous precisely where the consequences of being wrong are largest -- legal citations, financial figures, medical claims.
Common exam traps
- Believing a fabricated citation is a rare edge case. It is not -- specific-looking details are exactly where hallucinations concentrate, so citations and numbers deserve the most scrutiny, not the least.
- Assuming a Hallucination will "look wrong." It is defined by looking *right* -- confident and plausible -- while being false.
Key Takeaways
- Hallucination = confident, plausible-looking output that is false or fabricated (invented stats, citations, sources, quotes)
- Hallucinations concentrate in specific-looking details, at the edge of the model's knowledge, and inside long outputs
- A hallucination reads exactly as confident as correct content -- tone is not a signal
- Specific-looking citations and numbers deserve more scrutiny, not less, because that's where fabrication concentrates
Glossary Terms
Related Concepts
Inconsistencies, Bias, and Hallucination-Reduction Techniques
Hallucination (fabrication), inconsistency (internal contradiction), and bias (skewed framing) are three distinct failure types requiring different fixes
Capability Hallucination: Claiming an Action It Never Took
Capability hallucination is Claude claiming to have taken an external action (emailed a file, saved a document) it never actually took
Fact-Checking and Validation Scaled to Stakes
Validation is a deliberate confirmation step, distinct from improving an output's polish or formatting