AI hallucination is when a language model generates a response that sounds confident and coherent — but is factually wrong, invented, or unsupported by evidence. The term is borrowed from medicine, where a hallucination is a perception of something that isn’t there. In AI, it describes output that seems plausible but cannot be traced back to any real source.

Why it happens

Large language models are trained to predict the most likely next word given the words before it — not to look up facts in a database. When a model reaches a detail it doesn’t have firm knowledge of, it doesn’t stop; it picks the statistically most fitting continuation. This can mean inventing a plausible-sounding author, paper title, or date.

Part of the problem is the incentive structure during training: models are often rewarded for producing confident, fluent responses, rather than for saying “I’m not sure.” Hallucinations are therefore not a bug in the traditional sense — they are a consequence of how language models work.

What it looks like in practice

Hallucinations come in two main varieties:

  • Factuality errors — the model states a wrong fact: a fabricated quote, a wrong year, a made-up statistic.
  • Faithfulness errors — the model misrepresents what it was actually told or shown, such as summarising a document incorrectly or “correcting” correct information the user provided.

Real examples researchers and journalists have documented include:

  • ChatGPT inventing academic papers — complete with plausible authors, journals, and publication years — that do not exist.
  • AI legal tools generating fictional court cases and precedents, which ended up cited in real legal filings.
  • Chatbots producing false biographical details about living people.

How to reduce hallucinations

You cannot eliminate hallucinations entirely — they are inherent to how current language models work. But several practical habits meaningfully reduce the risk:

Ask for sources. Tell the model to cite specific documents, links, or passages for any factual claim. If it can’t, treat that as a signal to verify externally. Requiring citations has been shown to reduce hallucination rates in structured tasks by roughly 40%.

Use “I don’t know” prompting. Include something like: “If you’re not sure, say so rather than guessing.” Models respond to this instruction and are less likely to confabulate when given explicit permission to express uncertainty.

Narrow the task. The more focused the question, the less room for invention. Instead of “tell me about X,” try “summarise only what this document says about X” and paste in the document.

Verify important claims independently. Treat any specific fact — a statistic, a name, a date, a legal reference — as a starting point, not a final answer. Cross-check against authoritative sources before relying on it.

Use retrieval-grounded tools. Some AI assistants use retrieval-augmented generation (RAG) — they search real documents before composing an answer. Systems built this way hallucinate significantly less on factual queries because the response is anchored to actual retrieved text.

Are models getting better?

Newer models hallucinate less than earlier ones, and researchers are actively working on the problem. But as of mid-2026, no language model is hallucination-free. The research consensus is to aim for calibrated uncertainty — models that flag when they’re unsure rather than projecting false confidence — rather than a target of zero errors.

FAQ

Are AI hallucinations intentional? No. Models are not “lying.” They generate the statistically most likely continuation without an internal fact-checker — and sometimes that produces content that is wrong.

Which AI model hallucinates the least? Benchmark results vary by task. No model is hallucination-free, but models with built-in search grounding (RAG) and those trained with uncertainty-aware objectives tend to perform better on factual queries.

Should I stop using AI because of hallucinations? Not necessarily. AI is a useful starting point for research, drafting, and brainstorming — just verify any specific fact before relying on it, especially in legal, medical, or financial contexts.

Is hallucination the same as AI bias? No. Bias refers to systematic skew in outputs (such as demographic stereotypes). Hallucination refers specifically to fabricated or factually incorrect content. Both are reliability concerns, but they have different causes and different fixes.

Sources: Hallucination (artificial intelligence) — Wikipedia · Lakera Guide to LLM Hallucinations