What 'hallucination' actually means.

Hallucination is the AI-industry term for a specific failure mode: a language model producing plausible-sounding false information with apparent confidence. The claim is grammatically correct, contextually appropriate, and totally invented.

Why models hallucinate

Language models predict the next token based on patterns in their training. When you ask a question the model wasn't well-trained on — a specific person's birthday, a specific paper's citation, a version-specific fact — the model doesn't have a "confidence" it can express. It generates the most plausible-looking next tokens.

For most questions this works fine because the plausible answer IS the correct answer. For questions at the edge of training or beyond it, the plausible answer is often wrong — but the model has no internal signal saying "I'm making this up."

Where hallucinations are most common

- **Citations and references.** Asking for a specific study, book, or URL. Model will confidently invent titles, authors, and page numbers that don't exist. - **Recent events.** Anything after the model's training cutoff. Model may invent details. - **Numerical precision.** Specific statistics, market caps, dates. Model rounds and drifts. - **Legal or medical specifics.** Case law, drug interactions, jurisdiction-specific rules. High-stakes accuracy required; model often confident + wrong. - **Made-up people, edge-case history.** Ask about a person the model doesn't know well — you'll get confident biographical detail.

Where hallucinations are rare

- **Common knowledge.** "What's the capital of France." Not going to hallucinate. - **Well-attested facts.** Popular books, famous events, canonical science. The training coverage is thick enough that the model recalls accurately. - **Grammar, style, translation.** Language-shape questions rather than fact-recall. - **Reasoning that doesn't depend on specific facts.** Logic puzzles, math (with caveats), pattern analysis.

What web search fixes

Enabling web search converts hallucinations into either "here's the answer with a citation" or "I couldn't find that." A specific fact-check use case where search-on is the right default.

What web search doesn't fix

Search-based citations can still be miscited (the URL exists but the claim isn't in it). Always click through if the stakes are high.

The mitigation stack

- **Ask for citations** with search on. - **Verify anything you'd stake reputation on**, especially citations and statistics. - **Extended thinking** helps on multi-step reasoning where the model can catch its own errors on scratch paper. - **Ask "how confident are you"** — model self-reports of confidence are unreliable but sometimes catch obvious cases. - **Cross-check with a second source** for anything material.

The right calibration

Modern models (Claude Sonnet 4.6, GPT-5, etc.) hallucinate meaningfully less than models from 2023. But "less" isn't "never." Treating any specific citation, statistic, or biographical detail as verified without checking is the wrong instinct.

The right instinct: use the model for reasoning, drafting, explanation. Verify specifics through primary sources.