Hallucination
When a language model states something false or made up, fluently and confidently, as if it were fact.
Also confabulation · factual errors · grounding
A student who has to answer every question, even ones they do not know. Instead of saying 'I don't know', they invent an answer that sounds right.
Where it fits
+All relationships as a list
Is a kind of / part of
Siblings
- AI Agents and Tool Use
- BERT
- Direct Preference Optimization
- Embedding
- Fine-Tuning
- Google Translate
- GPT
- In-Context Learning
- Keyboard Autocomplete
- Large Language Model
- LLM Evaluation and Benchmarks
- Perplexity
- Pretraining
- Prompting
- Retrieval-Augmented Generation
- RLHF
- Sampling and Temperature
- Scaling Laws
- Tokenization
- Vector Database
- word2vec
Applies to
How it works
A large language model is trained to produce plausible text, and plausible is not the same as true. When the model’s knowledge is thin or missing, the most plausible continuation can be a confident fabrication: a citation that looks right, a statistic in the right range, a biography that mixes up two people.
Why it happens:
- Training goal: predicting the next token rewards fluent guesses; the model is rarely trained to say “I don’t know”.
- Lossy memory: facts are stored only approximately in the weights; rare facts are blurred.
- No built-in checking: generation does not look anything up unless the system is designed to.
What helps:
- Grounding: RAG supplies real documents and asks the model to answer from them, ideally with citations you can check.
- Training for honesty: post-training (such as RLHF) can reward admitting uncertainty.
- Verification: checking claims against sources, or with tools like calculators and search.
Rule of thumb: the more a wrong answer would cost, the more you should verify.
Everyday use cases
- Case 01
Invented citations
Models have produced references to papers or court cases that do not exist.
- Case 02
Wrong details in summaries
A summary can slip in a date or number that was not in the original document.
Strengths, weaknesses and when to use it
Strengths
- The same generative ability enables creativity and fluent writing.
Weaknesses
- Errors look exactly as confident as correct answers.
When to use itAssume it can happen; verify important facts and prefer grounded answers with sources.
Compare with
Check yourself
- Q01
Which technique most directly reduces hallucinated facts about a company's policies?
Show answer
A: Retrieval-augmented generation over the actual policy documents