A language model asserts with total conviction that Albert Einstein failed maths at school (false), that an academic article was published in Nature in 2019 (invented), or that the capital of Australia is Sydney (Melbourne is the largest city, but the capital is Canberra). It does so without hesitation, without any signal of uncertainty, in the same tone it would use to state that water boils at 100 degrees.
This is called hallucination, and understanding why it happens is essential for using language models responsibly.
What is an AI hallucination
In the context of language models, a hallucination is any output the model presents as true that is factually incorrect, invented or unverifiable.
Hallucinations can be:
- Incorrect facts (“The French Revolution began in 1793”)
- Invented citations (“As Nietzsche wrote in Thus Spoke Zarathustra: …” + a quote that does not exist)
- False bibliographic references (plausible titles, real authors, existing journals, but the specific article is invented)
- Fabricated statistical data
- Incorrect details about real people
- Wrong dates, places or names in historical contexts
Why it happens structurally
Hallucination is not a bug that can be eliminated with enough engineering. It is a direct consequence of how models work.
Remember: an LLM predicts what the most probable continuation of a text is. “Probable” here means “consistent with statistical patterns learned during training” — not “factually correct.”
The model has no access to a repository of verified facts it consults when generating responses. All its “knowledge” is encoded in the weights — the numerical parameters adjusted during training — and those weights capture statistical patterns of text, not truths about the world.
When asked about something not well represented in its training data, the model does not know what it does not know. It has no internal signal indicating “there is uncertainty here: better flag it.” It simply generates the most statistically plausible text given the context, and that text may not correspond to any real fact.
The problem is that the model is trained to sound confident and coherent — that is how high-quality human text sounds. That property, useful for fluency, is dangerous when it produces incorrect facts with the same tone as correct ones.
Types of hallucination
Plausibly-shaped hallucinations. The most dangerous are those that sound completely reasonable. “The 2018 Harvard study on productivity found that…” — the format is correct, the institution exists, the topic is plausible. Only the study does not exist.
Confusion between similar entities. Mixing data about two people with the same name, confusing similar historical events, applying characteristics of one city to another. The model has seen a lot of text about both entities and mixes them.
Incorrect extrapolation. The model knows facts about X and incorrectly extrapolates them to Y. It knows how monetary policy works in the US and applies that framework to the Eurozone with incorrect details.
Dates and numbers. Specific numbers — statistics, dates, prices, distances — are particularly prone to error. The model learns the context in which numbers appear but not necessarily the exact values.
Warning signs
Not all model assertions carry the same hallucination risk. These signals suggest extra caution:
- Very specific data: exact dates, statistics with decimals, verbatim quotations
- Topics infrequent in training: legislation of small countries, niche research, regional events
- Bibliographic references: titles, authors, journals, specific page numbers
- Recent information: events after the model’s knowledge cutoff
- Assertions about people: biographical details, attributed statements, specific achievements
- Very detailed responses on complex technical topics where error is hard to detect without prior knowledge
Verification strategies
Hallucination does not make language models useless. It makes selective verification necessary.
Ask for sources and verify. If the model mentions a study, a citation or a specific data point, search for that source externally before using it. Do not accept the model’s reference as validation: it may have invented it.
Use the model to check the model. After a response, you can ask: “Are you confident in this information? How certain are you?” The model does not always detect its own errors, but it sometimes acknowledges uncertainty when asked directly.
Triangulate with web search. For important factual information, use the model to identify what to search for, then use a search engine to confirm. Models with real-time web access reduce (but do not eliminate) factual hallucinations.
Distinguish by task type. Hallucinations are a lower risk when you use the model to rephrase, brainstorm, structure or synthesise information you yourself provide. They are a higher risk when you ask for factual information you cannot easily verify.
Calibrated scepticism. Do not distrust everything the model says — that makes it unusable. Distrust proportionally: more when the claim is very specific and hard to verify, less when it is general and corresponds to widely documented knowledge.
Working with language models means working with a powerful but unreliable information source in some contexts. Verification is not optional: it is part of the workflow.
