Hallucinations and Confident Errors

Hallucination is the most discussed failure mode in large language models — and the one evaluators must catch most vigilantly. A hallucination is content the model generates that is false, fabricated, or unsupported by its inputs, presented with the same fluency and confidence as accurate information.

Why Models Hallucinate

LLMs are trained to produce plausible text, not verified text. When asked a factual question, the model does not retrieve a confirmed answer — it generates the sequence of tokens most likely to follow the question. If the training data contained similar question-answer patterns, the output is often correct. If not, the model still generates a fluent answer — because stopping or saying "I don't know" is statistically less common in its training than providing an answer.

Types of Hallucinations

Factual hallucinations. Invented dates, nonexistent people, wrong statistics, incorrect citations. "The treaty was signed in 1847" when no such treaty exists.

Citation hallucinations. Fabricated references — plausible author names, journal titles, and DOIs that do not exist. Particularly dangerous in academic and medical contexts.

Logical hallucinations. Conclusions that do not follow from premises. Coherent-sounding reasoning with invalid steps.

Context hallucinations. Claims about provided documents that misrepresent or invent content not present in the source material.

Self-contradiction. The model states a fact in one sentence and contradicts it later, without acknowledgment.

Confident Errors vs. Uncertainty

Well-calibrated models express uncertainty when they should. Many LLMs do not — they state false information with the same authoritative tone as true information. Tone is not a reliability signal. Evaluators must verify claims independently, especially for:

• Specific numbers, dates, and statistics
• Names, titles, and affiliations
• Legal, medical, and financial information
• Quotes and attributions

How to Document Hallucinations

Effective hallucination reports include:

1. The model's exact claim (quoted)
2. Why it is false (correct information or source)
3. Severity (mild inaccuracy vs. dangerous misinformation)
4. Whether the error was preventable (model should have declined to answer)

High-Risk Domains

Hallucinations in creative writing are annoying. Hallucinations in medicine, law, finance, or crisis situations can cause real harm. When evaluating outputs in high-stakes domains, apply extra scrutiny to every factual claim. A confident wrong dosage, legal citation, or emergency instruction is a critical failure even if the surrounding prose is eloquent. Always note the domain when reporting severity.

Key Takeaways

• Hallucinations are fluent falsehoods, not random gibberish
• Models hallucinate because they optimize for plausibility, not truth
• Confident tone does not indicate accuracy — always verify factual claims
• Document hallucinations with exact quotes and corrections for maximum impact