⚡ Hallucination is when an AI model generates confident, fluent, plausible-sounding content that is factually wrong or entirely fabricated. The model does not know it is wrong — it generates the most statistically likely text, and sometimes that text is false. A lawyer submitted a legal brief with six ChatGPT-fabricated case citations. None existed. He was sanctioned by the judge.
Category: Foundational Concepts · Difficulty: Beginner · Last updated: 15 May 2026 · 5 min read
What is Hallucination?
The word is borrowed from psychology — a perception of something that does not exist. In AI, hallucination is the generation of content that sounds real but is not. The model cites a paper that was never written. It describes a historical event that never happened. It gives you a person’s CV with fabricated job titles and dates. And it does all of this with the same fluent, confident tone it uses when it is completely correct.
This is what makes hallucination dangerous. The problem is not that the AI produces obvious nonsense — that would be easy to spot. The problem is that hallucinated content is often indistinguishable from accurate content without independent verification. The fabricated legal case sounds like a real legal case. The invented statistic sounds like a real statistic. The wrong drug dosage sounds like the right drug dosage.
The root cause is architectural: language models do not store facts and retrieve them. They learn statistical patterns from text and generate new text that fits those patterns. When the pattern strongly suggests that “the capital of France is” should be followed by “Paris,” the model gets it right — Paris appears in that context millions of times in training. When the pattern is less clear — “Professor X published a paper on Y in 19—” — the model fills in what plausibly follows, regardless of whether it is true.
How Hallucination works
- LLMs are next-token predictors — they generate the most statistically likely continuation of the text, not the most factually accurate one.
- No truth-checking mechanism exists — the model cannot verify its own outputs against reality.
- Plausibility ≠ accuracy — a hallucinated fact often appears in context where such a fact would be plausible, making it convincing.
- Training data contains errors — models trained on the internet absorb misinformation alongside accurate information.
- Gaps in training coverage — for topics not well-covered in training data, the model generates plausible-sounding text to fill the gap.
Real-world examples
Not theory — what real teams actually shipped using this technique.
- Mata v. Avianca (2023) — a New York attorney submitted a legal brief citing six precedent cases generated by ChatGPT. All six were entirely fabricated — fake case names, fake courts, fake decisions. The judge sanctioned the attorney and ordered sanctions on the firm.
- Google Bard’s launch (2023) — in the first public demonstration, Bard incorrectly stated that the James Webb Space Telescope took the first image of an exoplanet outside our solar system (it did not). Alphabet’s market cap fell by approximately $100 billion in the following trading session.
- A Belgian man died by suicide after extended conversations with an AI chatbot that, according to his widow, encouraged and reinforced his suicidal ideation — a devastating example of hallucinated or irresponsible AI responses in a high-stakes mental health context.
How to Reduce Hallucination
Grounding and RAG — retrieve relevant documents and instruct the model to answer only from those sources. If the answer is not in the documents, say so.
Lower temperature — reduces randomness in generation, making outputs more conservative and less likely to confabulate.
Citation requirements — require the model to cite specific sources for every claim. Unsupported claims become visible.
Verification layers — use a second model or external fact-checker to validate the first model’s outputs.
Human review — for high-stakes outputs (legal, medical, financial), require human verification before acting on AI outputs.
Calibration training — train models to express uncertainty appropriately rather than generating confident text regardless of knowledge quality.
Common pitfalls
- Grounding reduces but does not eliminate hallucination — models can still misinterpret retrieved documents.
- Low temperature reduces creativity — the tradeoff for less hallucination is less varied, sometimes more generic outputs.
- Users trust confident outputs — even trained users fail to catch hallucinations in fluent, professional-sounding text. Awareness alone is not sufficient mitigation.
- Hallucination varies by domain — models hallucinate much more on obscure topics, recent events, specific statistics, and long-tail knowledge than on well-covered topics.
Frequently asked questions
QUESTION 1 What is AI hallucination in simple terms?
ANSWER 1 When an AI makes something up and presents it as fact with complete confidence. It is not lying intentionally — it generates the most statistically plausible text, and sometimes that text is wrong.
QUESTION 2 Why do AI models hallucinate?
ANSWER 2 They are next-token predictors, not fact retrievers. When they do not know something well, they generate plausible-sounding text rather than admitting ignorance.
QUESTION 3 What are real examples of harmful hallucinations?
ANSWER 3 A lawyer sanctioned for submitting six fabricated ChatGPT case citations. Google losing $100 billion in market cap from a hallucinated fact in a live demo. A chatbot reinforcing a user’s suicidal ideation.
QUESTION 4 How do you reduce hallucination?
ANSWER 4 Grounding and RAG, lower temperature, citation requirements, verification layers, and human review for high-stakes outputs.
📬 Get one concept + one use case every Tuesday. Join the newsletter →