Why AI hallucinates, in plain English
AI chatbots hallucinate because of how they learn and how they are graded. They learn by predicting the next word, which can't pin down arbitrary facts, and many test scores reward a guess over admitting uncertainty. That is the argument of a paper by Adam Tauman Kalai, Ofir Nachum, Santosh Vempala and Edwin Zhang, posted on arXiv on September 4, 2025, and explained in an OpenAI post a day later.
What is an AI hallucination?
A hallucination is a statement that sounds right but is false, produced by a language model with no hint of doubt. A research survey led by Ziwei Ji, first posted in February 2022 and published in ACM Computing Surveys, describes the same problem for text generation in general. It says systems built with deep learning tend to produce text nobody intended, and that this falls short of what users expect. The survey covers tasks such as summarization, dialogue, question answering and translation.
The OpenAI post gives a plain example. Researchers asked a widely used chatbot for the title of Kalai's PhD dissertation. It gave three different titles, and none was correct. They asked for his birthday and got three different dates, all wrong.
Why does predicting the next word produce false facts?
Language models start with pretraining: they read huge amounts of text and learn to predict the next word. No sentence in that text comes stamped true or false. The model sees only examples of fluent language, so it has little to learn from about what a wrong statement looks like.
The authors add an analogy. Suppose you trained an image classifier to label pet photos, not as cat or dog, but by each pet's birthday. Birthdays are essentially random, so no algorithm could get that right reliably. Some facts are like that: rare, arbitrary and not predictable from patterns in language.
Spelling and matching parentheses are different. They follow consistent patterns, so the authors say those errors fade as models scale. Obscure facts have no pattern to learn.
What they found
Hallucinations come from ordinary statistics. The paper argues they originate as errors in a classification problem: if a model can't tell a false statement from a fact, mistakes arise through natural statistical pressure. In the authors' view, they are not a mysterious glitch.
Scoring rewards guessing. Think of a multiple-choice test where a blank answer earns zero and a wild guess might earn a point. Most AI benchmarks work this way, ranking models by accuracy alone. If a model guesses a birthday, it has a 1-in-365 chance of being right. Saying it doesn't know earns nothing. Over thousands of questions, the guesser can look better on the leaderboard than a careful model.
OpenAI shows the effect with a test called SimpleQA, using numbers from the system card for GPT-5. It compares a newer model, gpt-5-thinking-mini, with an older one, OpenAI o4-mini. The older model scored a little higher on accuracy. It also answered wrong in 75% of cases, against 26% for the newer one, because it almost never declined to answer.
SimpleQA results for two OpenAI models
| Metric | gpt-5-thinking-mini | OpenAI o4-mini |
|---|
The table shows the trade in numbers. The newer model gave up on more than half of the questions, and that made it look slightly worse on an accuracy-only scoreboard. You can compare other systems on our list of AI models.
Perfect accuracy isn't coming. Some real-world questions can't be answered. The information may be unavailable, a small model may lack the ability, or the question may be ambiguous. The authors say no model size, search tool or reasoning ability changes that, so a model that is never wrong is not a realistic target.
Hallucinations aren't inevitable, though. A model can decline to answer when it is unsure. The authors also say this doesn't require a giant model. A small model that knows no Māori can simply say so. A model that knows some Māori has to judge how confident it is. Being calibrated, they write, takes much less computation than being accurate.
How could the scoring change?
The proposed fix is simple. Penalize confident errors more than uncertainty, and give partial credit when a model signals that it isn't sure. Some standardized tests for people already do this with negative marking for wrong answers.
The authors say adding a few new honesty tests isn't enough. The main benchmarks that dominate leaderboards and model cards need their scoring changed. If those keep paying for lucky guesses, they argue, models will keep learning to guess.
OpenAI says GPT-5 has significantly fewer hallucinations, especially when it reasons, but that they still occur.
Why does this matter beyond the lab?
The finding shifts part of the blame from the model's inner workings to the scoreboard. A model can be well built and still be pushed toward bluffing by how it is judged. It also explains why a chatbot's confident tone tells you little about whether an answer is right. To see how researchers study what a model does internally, read what happens inside an AI when it answers you.
Ask your AI
Ask your AI for the birthday of a person you have never heard of, such as a name you make up on the spot. A good answer admits it can't know. If it gives you a date, you have just watched the behavior this research describes. Then ask it what it would need to see before it could answer with confidence.
FAQ
Will AI ever stop hallucinating completely?
The paper's authors say accuracy will never reach 100% because some questions can't be answered. They also say hallucinations aren't inevitable, since models can abstain when unsure.
Is a hallucination the same as a spelling mistake?
No. According to OpenAI's post, spelling and parenthesis errors follow patterns and fade with scale, while arbitrary low-frequency facts can't be predicted from patterns and lead to hallucinations.
Does a bigger model always hallucinate less?
Not by size alone. OpenAI says GPT-5 hallucinates significantly less, especially when reasoning, but still does. It also says a small model can find it easier to know its limits.
Sources
Cite this
Why AI hallucinates, in plain English, New Mind Museum, November 18, 2026. https://newmindmuseum.com/interpretability/2026/11/why-ai-hallucinates-explained/
New Mind Museum. (2026, November 18). Why AI hallucinates, in plain English. https://newmindmuseum.com/interpretability/2026/11/why-ai-hallucinates-explained/
“Why AI hallucinates, in plain English.” New Mind Museum, 18 Nov. 2026, https://newmindmuseum.com/interpretability/2026/11/why-ai-hallucinates-explained/.
@misc{new-mind-museum-2026-why-ai-hallucinates-in-plain-english,
author = {{New Mind Museum}},
title = {Why AI hallucinates, in plain English},
year = {2026},
url = {https://newmindmuseum.com/interpretability/2026/11/why-ai-hallucinates-explained/}
}Free to quote and cite with a link to this page.