AI glossary Models & Architecture
What is an AI hallucination?
← All glossary termsWhen a language model produces information that sounds plausible but is false, invented or not supported by its sources.
Explained
An AI hallucination is an output from a large language model that reads as confident and fluent but is wrong: an invented fact, a citation that does not exist, a number that is not in the document or a claim that contradicts the context the model was given. It happens because these models generate the most likely next words, not verified facts, so a gap in knowledge turns into a plausible guess instead of an admission. Hallucinations cannot be removed entirely, but grounding the model in sources (for example with RAG), clear instructions, citations and confidence checks reduce them in production systems.
Why models hallucinate
- They are trained to produce plausible text, and plausible is not the same as true.
- Their knowledge stops at the training data, which can be incomplete, outdated or contradictory.
- They are rarely rewarded for saying "I don't know", so they tend to fill gaps.
- Long or noisy context can lead them to mix up sources or misattribute details.
Types of hallucination
Researchers usually separate two kinds. A factuality hallucination contradicts verifiable facts about the world, such as a wrong date or an invented law. A faithfulness hallucination contradicts or goes beyond the input the model was given, such as a summary that adds a point the document never made.
In enterprise use the second kind matters most, because the model is usually asked to work from the organisation's own documents and data.
How to reduce them
- Ground answers in retrieved sources and ask the model to quote or cite them.
- Allow, and ask for, "I don't know" when the sources do not cover the question.
- Constrain the task: structured outputs, narrower prompts, smaller steps.
- Check outputs automatically (consistency checks, a second model as judge) and with people where the risk is high.
- Measure the hallucination rate on a fixed set of real questions before and after each change.
Frequently asked questions
Why do AI chatbots make things up?
Because a language model predicts likely text rather than looking up verified facts. When it lacks the information, the most likely continuation is often a confident guess. Grounding it in sources and letting it admit uncertainty reduces this.
Can AI hallucinations be eliminated completely?
Not with current models. They can be made much less frequent and easier to catch, through retrieval, citations, constrained tasks and evaluation. For high-stakes decisions a human review step remains advisable.
Does RAG stop hallucinations?
It reduces them, because the model answers from relevant documents instead of memory. It does not stop them: the model can still misread the passage, or the retrieval can miss the right one. RAG needs evaluation like any other component.
How do you detect a hallucination?
Compare the answer with its sources: every claim should be traceable to a passage. Automated checks include asking the model several times and comparing answers, or using a second model to verify claims against the context. For critical outputs, people review a sample.
What is the difference between a hallucination and a bias?
A hallucination is a specific false or unsupported output. A bias is a systematic tendency in the model's outputs, for example favouring one group, that may appear even when each individual statement is accurate. Both need testing, but with different methods.
Decision guides that use this term
How to put it to work
Where it shows up at Thinkia
Further reading
Related terms
Sources
- Huang et al. · A survey on hallucination in large language models (arXiv)
- Anthropic · Reduce hallucinations (Claude documentation)
- OpenAI · Kalai et al., Why language models hallucinate
Updated: 2026-10-21