What is an AI hallucination and where do they come from?
TL;DR, the essentials
- An AI hallucination is a factually false response produced with an appearance of certainty: a made-up number, a source that does not exist, a wrong date, all stated in a perfectly confident tone.
- This is not an isolated bug but a logical consequence of how a language model works, calculating the most probable sequence of words, not the most truthful one.
- Root cause identified in an OpenAI paper published in Nature (April 2026): evaluation benchmarks reward guessing and penalize admissions of ignorance, training models to bluff.
- You can reduce it significantly (RAG, web search, reasoning mode, human review), but no technique eliminates it entirely. Verification remains essential for high-stakes use.
You ask an AI assistant for a legal reference, a market figure or a company executive’s biography, and it answers with confidence, except the information is completely made up. This phenomenon is called hallucination, and it is the primary limitation of generative AI systems in 2026. Here is a clear definition, the real technical causes (beyond “AI makes mistakes”), the actual measured frequency and concrete ways to protect yourself.
What exactly is an AI hallucination?
A hallucination is a response generated by an AI that appears credible and coherent, but is factually false, made-up or inconsistent with actual facts or provided information. The term is misleading: the model is not “seeing” anything; it simply produces a plausible output without having the actual fact and without signaling any doubt.
The key nuance is this: unlike a hesitant human error, hallucinations come with an assured tone. The model affirms a contract clause, cites a study, gives a precise date, with the same apparent confidence as a true statement. That is what makes it dangerous: nothing in the form signals that the content is false.
In one sentence
A hallucination is when an AI fills a gap in knowledge with a plausible invention, presented with the same confidence as a verified fact.
What are the different types of hallucinations?
Not all cases are alike. Two main families are generally distinguished, useful to know where to focus vigilance.
- Intrinsic hallucinations (or fidelity): the response contradicts information you provided to the model. For example, you paste a document and the AI distorts its content or invents a passage. The problem comes from processing the source, not a knowledge gap.
- Extrinsic hallucinations (or factual): the model invents a fact with no verifiable basis, outside any supplied source. These are the most common and most risky in business, because nothing lets you cross-check the answer in the moment.
In addition, typical cases include: false citations (articles or court rulings that do not exist), invented numbers, wrong dates, or incorrect attributions (a quote put in the wrong person’s mouth).
Where do AI hallucinations come from?
This is the core of the matter, and the answer runs deeper than “AI is not ready yet”. Four structural causes combine.
Probabilistic reasoning
A language model (LLM) predicts the most probable next word, one word at a time. It optimizes for likelihood, not truth. Nothing in its design verifies that a claim corresponds to real fact. A small drift early in a sentence can compound into a cascade later on.
Benchmarks that reward guessing
This is the most striking discovery of 2026. According to an OpenAI paper published in Nature (April 2026), tests used to evaluate models score a correct answer positively and penalize a wrong answer the same way as no answer at all. Result: statistically, guessing pays more than saying “I do not know”. It is a multiple-choice test with no penalty for wrong answers, where random guessing is always profitable. Models learn to fill in their gaps instead of admitting them.
Knowledge cutoff date
A model is frozen at the end of its training date. Any event after that escapes it, and facing a question about that period, it tends to reconstruct a plausible answer instead of signaling it does not know.
Training data quality
The web that models learn from contains errors, contradictions and outdated content. A model can produce false information simply because it was present, and repeated, in its training data.
It is not a malfunction, it is a feature
The counterintuitive point: hallucination is not a bug to fix, but a direct consequence of how these systems are built and evaluated. That is why we speak of reducing it, never eliminating it with a single patch.
Want an AI tool that hallucinates less?
Our comparison ranks AI software 2026 on reliability, cited sources and privacy.
How often does AI actually hallucinate?
Numbers vary widely depending on the model, domain and how you phrase the question. A few benchmarks measured in 2026, to take as rough orders of magnitude.
- General use: a factual hallucination rate of around 15 to 20% is observed on open-ended questions with no safeguards (indicative figures, 2026).
- Legal domain: according to multiple 2026 studies, the error rate climbs to between 58 and 88% on pointed legal questions, hence the extreme caution required.
- Medical domain: summaries of clinical cases showed up to roughly 64% of incorrect content without control measures in place.
Not all models are equal
According to the AA-Omniscience benchmark from Artificial Analysis (April 2026), which measures the propensity to fabricate an answer when the model does not know the right one: Claude Opus 4.7 invents in roughly 36% of cases, Gemini 3 Pro in 50%, and GPT-5.5 in 86%. The same prompt can be significantly more risky depending on which tool you pick (indicative data, April 2026).
Quick quiz
The root cause of a hallucination is…
When are hallucinations most dangerous?
Risk is not the same across all uses. A hallucination in a recipe idea is trivial; in a client file it can be costly. High-risk domains where every output must be verified by a human:
- Legal: citing a court ruling or law article that does not exist can lead to real professional sanctions, already documented.
- Medical and health: an invented dosage or diagnosis puts people’s safety at risk.
- Financial and accounting: a wrong number in analysis or reporting skews a decision.
- Research and academia: false bibliography references are a classic trap.
- Code generation: AI can invent a nonexistent function or library, with security gaps as a result.
You remain responsible
Legally, the company or professional who publishes or acts on AI output remains liable. Handing off writing to a model does not transfer responsibility for the result.
How to reduce AI hallucinations
Good news: several levers slash the hallucination rate, sometimes dramatically. Combine them based on stakes.
- RAG (retrieval augmented generation): hook the model to your documents so it answers from retrieved sources, not memory alone. On well-anchored summaries, error rates often drop below 2%. It is the most effective lever for document-based work. We detail how it works in our AI hub.
- Web search: letting the model consult live sources sharply cuts fabrication. In some 2026 tests, an assistant drops from roughly 60% to 30% invention thanks to search.
- Reasoning mode: models that break down their logic before answering hallucinate less, around half less on some tasks (indicative figure).
- Sourced prompt: explicitly asking for citations and “I do not know” when in doubt reduces bluffing.
- Low temperature: setting creativity to minimum (0 to 0.3) on factual tasks limits drift.
- Human review: on any critical output, sign-off by a qualified person remains the last line of defense, non-negotiable.
The golden rule
Never ask an AI to supply you a fact it cannot source. Use it to synthesize documents you give it, rephrase or structure, and keep the fact-checking on your side.
How to spot a hallucination
A few signals let you smell a suspect answer before running with it:
- Too-perfect precision: numbers to two decimal places, exact dates, word-for-word quotes on niche topics deserve a fact-check.
- Unverifiable sources: a study title or link that goes nowhere is a classic red flag.
- An answer that shifts: rephrase the same question. If the answer varies on a supposedly stable fact, be cautious.
- A recent or specialist domain: the more specialized the subject or past the knowledge cutoff, the higher the risk.
Which reliable AI tool to pick in 2026?
We compared AI software on reliability, price and privacy, with real limitations spelled out.
What comes next
Want an AI that invents as little as possible on your data? Check out our comparison of the best AI software 2026, or explore all our guides from the AI hub.
Frequently asked questions
What is an AI hallucination?
It is a response generated by an AI that appears credible but is factually false, made-up or inconsistent. The AI fills a gap in knowledge with plausible information, presented with the same confidence as a verified fact, without signaling any doubt. This can take the form of a made-up number, a false citation or a wrong date.
Why do AIs hallucinate?
Four reasons combine. First, a language model predicts the most probable word, not the most truthful. Second, according to an OpenAI paper published in Nature in April 2026, evaluation benchmarks reward guessing and penalize admissions of ignorance, training models to bluff. Third, there is a knowledge cutoff date. Fourth, training data contains errors and outdated content.
Can you stop hallucinations entirely?
No. Hallucination is a consequence of probabilistic reasoning in models, not a simple bug. You can reduce it sharply through RAG, web search, reasoning mode or a low temperature setting, but no technique eliminates it at 100%. Human review remains essential for any high-stakes use.
Which AI hallucinates least in 2026?
It depends on the model. According to the AA-Omniscience benchmark from Artificial Analysis (April 2026), Claude Opus 4.7 fabricates an answer in roughly 36% of cases where it does not know the right one, against 50% for Gemini 3 Pro and 86% for GPT-5.5. The gaps are significant. We compare tool reliability in our comparison of the best AI software 2026 (indicative data, worth rechecking as models evolve fast).
Does RAG eliminate hallucinations?
It reduces them sharply, down to below 2% on well-anchored summaries, but does not eliminate them. If the document base is incomplete, poorly prepared or out of date, the model can still invent to fill gaps. RAG quality depends directly on your document quality.