In artificial intelligence, a hallucination is an output, usually from a generative model, that is fluent and confident but false, fabricated or unsupported by the source it was given. A chatbot may invent a citation, state a wrong date as fact or describe an event that never happened, all in a calm and convincing tone. The term is borrowed from psychology, though many researchers prefer words such as fabrication or confabulation, because a model does not perceive anything. Hallucination is among the most important reliability problems for large language models and other generative systems.
Types of hallucination
Researchers usually distinguish two broad kinds.
- Factual hallucination: the output contradicts known facts or invents them, such as a nonexistent book, court case or scientific paper.
- Faithfulness hallucination: the output departs from the input or instructions, for example a summary that adds claims the document never made or an answer that ignores the provided context.
Other categories include logical errors, invented tool results in AI agents and, in image or video generation, objects and details that do not match the prompt or physics, such as hands with extra fingers.
Causes
Hallucinations arise from how models are built and trained.
- Objective: a language model is trained to predict plausible next tokens, not to verify truth, so fluent text can be produced without a fact behind it.
- Data gaps and errors: training data contains mistakes, outdated facts and rare topics with little coverage, and models guess where knowledge is thin.
- Reward for guessing: training and many benchmarks give credit for a correct guess and none for saying that the answer is unknown, which encourages confident answers.
- Decoding: sampling choices such as high temperature can raise the chance of unlikely tokens.
- Context problems: long or conflicting inputs, and prompts that assume false premises, can lead models astray.
- Compounding: once a model writes an error, later text may build on it.
Notable cases
In 2023, a New York lawyer submitted a brief containing cases that a chatbot had invented, and the court sanctioned the lawyers. Chatbots have attributed false claims to real people, produced fake references in academic work and given incorrect medical or legal information. In customer service, a company was held responsible by a tribunal in 2024 for misleading statements that its chatbot made about a refund policy. These cases show that responsibility stays with the people and organizations that deploy the system.
Detection
Detection methods include checking claims against trusted sources, asking a model to cite and then verifying the citations, comparing several sampled answers for consistency, measuring the model’s uncertainty and using separate verifier models. Benchmarks such as TruthfulQA and a range of faithfulness tests measure how often models produce false statements, though results vary by task.
Mitigation
- Grounding: supplying documents and requiring answers to rely on them, as in retrieval-augmented generation and grounding.
- Better training: reinforcement learning from human feedback and rewards that value calibrated uncertainty and honest refusals.
- Prompting: clear instructions, permission to say that the answer is unknown and requests for evidence.
- Tools: calculators, search and code execution for tasks where models are weak.
- Human review: checks by experts for high-stakes content.
- Reasoning: reasoning models check their work and reduce some errors, though not all.
No method removes hallucination completely, so risk management matters as much as model improvement.
Implications
Hallucinations undermine trust, can spread misinformation and create legal, medical and financial risks. They matter most where users cannot easily verify answers. Responsible deployment includes disclosures, logging, citations, human oversight and limits on use in critical decisions. Some creative uses, such as brainstorming fiction, benefit from the same ability to generate novel text, so the goal is controlling when it occurs rather than eliminating generation altogether.
A worked example
A student asks a chatbot for three academic papers on sleep and memory. The chatbot returns three titles with authors, journals and years. Two exist and one does not: the title sounds right and the authors are real researchers, but the paper was never written. The student who searches for each title in a library database discovers the problem. Had the student used a search-connected tool that returns links, the error would have been less likely.
Calibration and uncertainty
A well-calibrated model signals how sure it is, so that high confidence usually goes with correct answers and low confidence with doubtful ones. Language models are often reasonably calibrated internally but express answers in a uniform, assured tone. Research therefore studies ways to read uncertainty from token probabilities, from agreement among several sampled answers and from the model’s own statements, and to convert it into useful behavior, such as refusing, asking a clarifying question or adding a caution.
Hallucination in other media
The problem is not limited to text. Image generators may draw impossible anatomy or garbled lettering, speech systems may transcribe words that were never spoken, and video models can break physical rules. Medical imaging models that reconstruct or enhance scans can add features that are not in the patient, which is dangerous when the output is trusted. Each field develops its own checks, such as comparing a reconstruction with measured data.
Design practices for products
Product teams reduce harm by showing sources next to claims, separating generated text from verified facts, limiting the model to approved documents for sensitive domains, logging answers for audit and offering easy ways for users to report errors. Testing with adversarial questions, including false premises and requests about obscure facts, reveals weaknesses before launch. Ongoing monitoring matters because behavior changes when models or data are updated.
Measuring hallucination rates
Reported rates vary widely because they depend on the task, the domain and how a response is judged. Summarization tests count claims that the source does not support, while open question tests count false statements about facts. Rates tend to be higher for obscure subjects, recent events and requests for citations, and lower when the model has the relevant document in front of it. Because rates change with every model update, organizations that depend on a system measure them on their own data rather than relying on published numbers.
Related concepts
Hallucination connects to prompt engineering, to evaluation of language models and to the broader field of AI safety.
FAQs
Is a hallucination the same as a lie? No. A lie involves intent to deceive, while a model has no beliefs or intent. It generates text that is plausible but wrong.
Can hallucinations be eliminated? Not at present. They can be reduced with grounding, better training and verification.
How can I spot one? Verify names, numbers, quotations and citations independently, and be wary of very specific details that you cannot confirm.
More in Large Language Models
Key facts
- A 2023 survey by Ji and colleagues reviewed hallucination in natural language generation and separated faithfulness errors from factual errors.1
- In 2023, a New York lawyer filed a court brief that cited cases invented by a chatbot, and a federal judge sanctioned the lawyers involved.2
- A 2025 OpenAI paper argued that standard training and evaluation reward guessing over admitting uncertainty, which helps explain why hallucinations persist.3
Related concepts
Quick checkWhy do language models hallucinate?Show answer
They generate the most plausible continuation of text rather than checking facts, and training and evaluation often reward confident guesses over admitting uncertainty.
Sources
- Ji et al., Survey of Hallucination in Natural Language Generation, https://arxiv.org/abs/2202.03629
- Reuters, New York lawyers sanctioned for using fake ChatGPT cases in legal brief, https://www.reuters.com/legal/new-york-lawyers-sanctioned-using-fake-chatgpt-cases-legal-brief-2023-06-22/
- Kalai et al., Why Language Models Hallucinate, https://arxiv.org/abs/2509.04664
Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue
