Hallucination
Also: AI hallucination, confabulation, fabrication
A hallucination is when an AI model generates output that is fluent and confident but factually wrong, fabricated, or unsupported by its source data.
What It Is
A hallucination occurs when a generative AI model (such as a large language model or image generator) produces content that sounds plausible but is false, invented, or not grounded in any real source. The model is not lying in a human sense; it is predicting likely sequences of tokens, and sometimes the most statistically probable output does not match reality.
Hallucinations can take several forms:
- Factual errors: stating an incorrect date, statistic, or attribution.
- Fabricated sources: inventing citations, URLs, legal cases, or studies that do not exist.
- Logical inconsistency: contradicting earlier statements within the same response.
- Unsupported claims: answering confidently about topics outside the provided context.
Why it matters
Hallucinations are one of the biggest barriers to trustworthy AI deployment. In high stakes fields the consequences are serious:
- Finance: an invented earnings figure or misquoted regulation can drive bad decisions.
- Healthcare and legal: fabricated references can cause real harm or sanctions.
- Marketing: false product claims expose a brand to compliance and reputational risk.
Because models present hallucinations with the same fluent, confident tone as correct answers, users cannot rely on style to detect them. This makes verification a human responsibility.
How it is handled in practice
Teams reduce hallucinations through several techniques:
- Retrieval Augmented Generation (RAG): ground answers in trusted documents so the model cites real context.
- Prompt design: instruct the model to say "I do not know" when uncertain.
- Guardrails and validation: check outputs against databases, schemas, or rules.
- Human review: keep a person in the loop for critical outputs.
- Evaluation: measure hallucination rates with benchmarks before shipping.
Concrete Example
You ask a chatbot: "Which study proved this supplement cures insomnia?" The model replies with a confident answer citing "Journal of Sleep Medicine, 2019, Dr. Lena Park." The citation looks authoritative, but no such article, author, or finding exists. The model fabricated a credible looking reference. A reviewer who checks the source catches the hallucination before it reaches customers.
See also
Frequently asked questions
What is an AI hallucination?
An AI hallucination is output from a generative model that sounds fluent and confident but is false, invented, or unsupported by any real source. The model is not lying: it predicts likely token sequences, and the most statistically probable answer sometimes does not match reality. Common forms include wrong dates or statistics, fabricated citations, self-contradiction within one response, and confident answers about topics outside the provided context.
Why can't I just tell from the tone whether an answer is hallucinated?
Because models produce hallucinations in exactly the same fluent, confident register as correct answers. Style carries no signal about factual accuracy, so a well written paragraph is not evidence that the content is true. That is why verification stays a human responsibility rather than something you can delegate to your reading instinct.
What's the difference between a factual error and a fabricated source?
A factual error is an incorrect date, statistic, or attribution about something that does exist. A fabricated source is an entirely invented citation, URL, legal case, or study that has no real counterpart. Fabricated sources are more dangerous because they look verifiable and often pass unchecked, especially in legal and healthcare contexts where a fake reference can trigger sanctions or real harm.
How do teams actually reduce hallucinations?
Five techniques are used in practice: Retrieval Augmented Generation (RAG) to ground answers in trusted documents, prompt design that instructs the model to say it does not know when uncertain, guardrails that validate output against databases, schemas or rules, human review for critical outputs, and evaluation that measures hallucination rates on benchmarks before shipping. They stack rather than compete, and none of them removes the risk entirely.
Can you give a concrete example of a fabricated citation?
Ask a chatbot which study proved a supplement cures insomnia, and it may answer confidently citing "Journal of Sleep Medicine, 2019, Dr. Lena Park." The reference looks authoritative, but no such article, author, or finding exists. A reviewer who opens the source catches the hallucination before it reaches customers, which is exactly why human review sits in the workflow.