Why does AI hallucinate?
Short answer
AI hallucinates because a language model predicts likely words, not verified facts, and training rewards a confident guess over "I don't know". OpenAI's own figures show the gap: on SimpleQA, a model that abstained 1% of the time had a 75% error rate, against 26% for one that abstained 52% of the time (vendor claim, 2025).
Cause 1: the model predicts text, it does not look facts up
A large language model is trained to continue text. Given a question, it produces the sequence of words that best fits the patterns it learned from its training data. That process has no separate step that checks whether a sentence is true.
Most of the time, likely text and true text overlap, because true statements are common in the training data. The two split apart when the model meets a question whose answer it has seen rarely or never. It still produces the most plausible-looking answer: a real-sounding court case, a believable statistic, a citation with the right format. Plausible and true are different things, and the model has no internal signal that tells them apart.
This is why hallucinations sound confident. Tone comes from the style of the training text, not from the model's certainty about a fact. In the Tow Center citation test, ChatGPT Search signalled uncertainty only 15 times in 200 questions and never declined, while getting 134 answers wrong (March 6, 2025).
Cause 2: scoring rewards guessing
OpenAI's research on hallucination argues that the way models are tested makes the problem worse. Most benchmarks score accuracy only: a right answer earns a point, and both a wrong answer and "I don't know" earn nothing. Under that rule, guessing always scores at least as well as abstaining, so models learn to guess.
OpenAI's figures on its SimpleQA test show the effect (Why language models hallucinate, vendor claim, September 5, 2025):
| Model | Accuracy | Abstained | Error rate |
|---|---|---|---|
| gpt-5-thinking-mini | 22% | 52% | 26% |
| o4-mini | 24% | 1% | 75% |
The two models got about the same share of questions right. The one that almost never abstained got far more wrong. Refusing when unsure cut errors without lowering accuracy much.
The independent AA-Omniscience benchmark builds this into its score: its index rewards correct answers and penalizes wrong ones, so a score of 0 means as many right as wrong. When it launched in November 2025, most models gave more wrong than right answers on its hard knowledge questions, and the top model, Claude 4.1 Opus (a model behind Claude), scored 4.8, one of only three above zero.
Cause 3: gaps and rare facts in training data
A model's knowledge comes from its training data, and that data has gaps. Facts about less famous people, niche topics, local events and recent news appear rarely or not at all. When the model is asked about them, it fills the gap with something that fits the pattern.
OpenAI's own PersonQA test, which asks about public facts on people, shows how hard this is even for advanced models. Its o3 reasoning model hallucinated on 33% of questions and o4-mini on 48%, against 16% for the older o1 (o3 and o4-mini System Card, vendor claim, April 16, 2025). A newer model is not automatically a more factual one.
Recent events are the sharpest case. After the July 2024 assassination attempt on Donald Trump, Meta AI told some users it had not happened; Meta called these hallucinations (Meta, July 30, 2024).
Cause 4: web search brings in bad sources
Connecting a chatbot to live web search fixes some gaps but opens another failure: the model can find a false claim online and repeat it as fact. In NewsGuard's August 2025 audit of 10 leading chatbots, the non-answer rate fell from 31% to 0% in a year, while the rate of repeating false news claims rose from 18% to 35%. NewsGuard links the change to live web search: chatbots answered more, and repeated more falsehoods.
Search also adds a sourcing step that can go wrong on its own. Across 8 AI search tools, more than 60% of answers naming a news article's source were wrong in the Tow Center test, and premium tiers gave more confidently wrong answers than free ones. Sourcing problems drove most of Gemini's issues in the EBU/BBC study: 72% of its responses had sourcing problems, and 76% had at least one significant issue overall (October 21, 2025).
Cause 5: summarizing adds what was never there
Even with the right document in front of it, a model can add claims the document does not make. The Vectara Hallucination Leaderboard measures exactly this: how often a model's summary of a short document contains unsupported content. On its board updated October 22, 2026, the best model scored 1.8%, and the models behind the major assistants scored several times higher. GPT-5.5, a model behind ChatGPT, scored 9.3%; Claude Opus 5.5, the model behind Claude, scored 9.9%; and Grok 4 Fast (reasoning), a model behind Grok, scored 20.2%.
Grounding an answer in a source lowers the risk. It does not remove it. Every published rate by model is on AI hallucination rates by model.
Which questions trigger it most
The causes above point to the questions where hallucinations are most likely:
- Requests for sources and citations. The model knows what a citation looks like better than which ones exist. See does ChatGPT make up sources?
- Facts about less famous people, places or organizations. Rare in training data, so the model fills gaps.
- Recent and breaking news. After the training cut-off, or picked up from unreliable pages through search.
- Exact numbers, dates and quotes. Easy to produce in the right format, hard to get exactly right.
- Legal, medical and academic detail. Specialist formats such as case citations are easy to imitate. Courts have addressed AI-hallucinated content in 2,241 decisions in Damien Charlotin's database (checked October 29, 2026).
- Leading questions built on a false premise. A model trained to be helpful tends to answer the question as asked rather than challenge it.
Real cases for each type are on real AI hallucination examples.
What reduces it
On the model side, the evidence points to three changes: rewarding abstention when unsure, grounding answers in retrieved documents, and testing on hard questions that punish wrong answers. On your side, the most reliable protection is to check the parts of an answer that are easiest to invent. The five steps on how to fact-check an AI answer cover claims, sources, numbers, dates and quotes.
Tools also differ. Claude repeated false news claims 10% of the time in NewsGuard's August 2025 audit, the lowest of 10 chatbots, while Perplexity was at 46.67%. A tool-by-tool view is on the most accurate AI assistants.
A language model has no built-in way to know when it is wrong, so the check has to come from you.
Questions
Why does AI make things up instead of saying it doesn't know?
Because training and testing reward answers. Most benchmarks give a point for a right answer and nothing for "I don't know", so guessing scores better. OpenAI's figures show a model that abstained 52% of the time made 26% errors on SimpleQA, against 75% for one that abstained 1% of the time, with similar accuracy (OpenAI, vendor claim).
Do newer AI models hallucinate less?
Not always. OpenAI reported that its o3 reasoning model hallucinated on 33% of PersonQA questions, against 16% for the older o1 (System Card, vendor claim). On the Vectara leaderboard, Gemini 2.5 Pro, a model behind Gemini, scored 7.0% on summaries, while the newer Gemini 3.1 Pro preview scored 10.4%.
Does web search stop AI hallucinations?
No. Search helps with recent facts but brings in bad sources. In NewsGuard's audits, 10 leading chatbots repeated false news claims 35% of the time in August 2025, up from 18% a year earlier, as non-answers fell to 0%. NewsGuard tied the rise to live web search.
Is an AI hallucination a bug that will be fixed?
It is a side effect of how language models work, so it can be reduced but not removed. Even the best model on the Vectara leaderboard added unsupported content to 1.8% of summaries in October 2026. Rewarding abstention and grounding answers in sources lower the rate; checking the answer is still on the user.
Which kinds of questions cause the most hallucinations?
Requests for citations, facts about less famous people, recent news, exact numbers and quotes, and specialist detail such as legal cases. These are rare in training data or easy to imitate in format. Questions built on a false premise also invite a confident wrong answer, because the model tends to answer as asked.
Sources
- Vendor claim
- Vendor claim
- Benchmark
- News audit
- Independent study
- Independent study
- Benchmark
- Vendor claim
