What "Hallucination" Means Here
In AI, a hallucination is output that's fluent and confident-sounding but factually incorrect or entirely fabricated — a citation that doesn't exist, a statistic that was never published, a historical detail that's simply wrong. The term is a bit misleading (the model isn't perceiving anything), but it's become the standard way to describe this specific failure mode.
Why It Happens
As covered in our How LLMs Work page, a model generates text by predicting the most statistically plausible next token, not by retrieving verified facts from a database. When a model doesn't have solid information on a topic, it can still produce fluent, plausible-sounding text — because fluency and factual grounding are fundamentally different things to the underlying prediction process.
When It's Most Likely to Happen
Hallucination risk tends to increase for very specific or obscure facts the model saw rarely (or not at all) during training, requests for exact citations, dates, or statistics, and multi-step reasoning where an early error compounds through the rest of the response.
How to Reduce the Risk in Practice
Ask the model to cite its reasoning or flag uncertainty rather than just stating an answer, use RAG to ground responses in verified source documents rather than relying purely on training data, and independently verify anything factually important, especially specific numbers, quotes, or citations, before relying on it.
Related Pages
Frequently Asked
Can hallucination be completely eliminated?
Not currently — it's an inherent characteristic of how these models generate text; techniques like RAG and careful prompting reduce but don't eliminate the risk.
Do newer, more capable models hallucinate less?
Generally yes, on average, but even the most capable current models still hallucinate under the right conditions, particularly on obscure facts or exact citations.
Is hallucination the same as bias in AI?
No — hallucination is generating false information confidently; bias refers to systematic skew in a model's outputs reflecting patterns (including unwanted ones) in its training data. They're related but distinct issues.
How can I tell if a response is hallucinated?
There's no foolproof way from the output alone — independently verifying specific facts, especially citations, statistics, and quotes, against a reliable source is the most reliable check.