Why AI Makes Things Up: Hallucinations Explained

You ask an AI a simple question and it answers instantly, in clean sentences, with a date, a name and a source. Then you check, and the source doesn’t exist. Why would a tool this capable make things up?

Here’s the short answer. A chatbot writes by predicting likely words, not by looking facts up, and the way models are trained and scored has rewarded confident guesses over “I don’t know.” This guide explains how that works in plain English, the five kinds of made-up answers you’ll run into, when the risk is highest, and the three habits that cut it down the most.

Why AI makes things up: the 30-second version

• It predicts, it doesn’t look up. A language model builds each answer one small chunk of text at a time, picking what is most likely to come next. There’s no built-in fact database it checks against.
• Some facts can’t be predicted. Well-known facts form strong patterns. Obscure ones (a small company’s founding year, a person’s birthday, the exact title of a paper) don’t, so the model fills the gap with something that sounds right.
• Training rewarded guessing. OpenAI’s own researchers argue that the tests used to score models give points for lucky guesses and nothing for “I don’t know,” so models learned to guess.

The industry name for a confident wrong answer is a hallucination. OpenAI defines it as “instances where a model confidently generates an answer that isn’t true.” It’s not a glitch in one app. Every chatbot built on a large language model does it.

How a chatbot actually writes an answer

When you type a question, the model doesn’t pull the right card from a filing cabinet. It predicts the next chunk of text, then the next, until the answer is done. It learned to do this from an enormous amount of text during training.

That’s why it writes so well: grammar, tone and the usual shape of an answer are strong patterns. It’s also why it can be fluent and wrong at once. Sounding right and being right are different skills, and prediction mostly trains the first. (OpenAI’s researchers note that big models rarely misspell words, for exactly this reason.) For a friendly walkthrough of the “guess the next word” idea, read how AI chatbots really work.

Five reasons prediction produces confident wrong answers

1. Some facts have no pattern to learn

OpenAI’s explanation of why language models hallucinate uses a good analogy. Train a system on millions of pet photos labeled “cat” or “dog” and it can learn to tell them apart. Label the same photos with each pet’s birthday instead and it will always make errors, because birthdays are essentially random. Nothing in the photo tells you the answer.

Language models hit the same wall with what OpenAI calls “arbitrary low-frequency facts”: details that rarely appear in the training text and can’t be worked out from anything else. Asked for the title of one co-author’s PhD dissertation, a widely used chatbot confidently gave three different answers. None were correct.

2. The scoring rewarded guessing

Think of a multiple-choice test: a blank gets zero, a guess might get lucky. OpenAI argues most AI benchmarks grade the same way, so a model that guesses looks better on the leaderboard than one that admits uncertainty.

Their own published numbers show the trade-off. On one factual quiz called SimpleQA, an older model that almost never declined to answer got 24% right and 75% wrong. A newer reasoning model that declined more than half the time got 22% right and only 26% wrong. Nearly the same accuracy, far fewer confident errors.

3. Recognizing a name isn’t the same as knowing about it

Anthropic’s researchers looked inside their Claude model and found that its default behavior is to say it doesn’t have enough information. When it recognizes something familiar, like a famous athlete’s name, that default gets switched off so it can answer.

Hallucinations can happen when that switch misfires: a name feels familiar, the “I don’t know” brake comes off, but the model doesn’t actually know anything about it. Out comes a plausible, invented answer. Half-familiar names, niche products and minor historical figures are prime territory.

4. Its knowledge stops at a date

A model learns from text collected up to a cutoff. Ask about anything after that (a new price, a law that changed, an app update) and it may confidently describe the world as it used to be. Anthropic’s help center says it plainly: Claude “might not have been trained on the most up-to-date information.” Unless the tool is searching the web, it isn’t checking.

5. It’s tuned to agree with you

Anthropic’s research on sycophancy found that the AI assistants it tested consistently leaned toward answers that matched the user’s views, partly because people rate agreeable answers higher. So a question with a false assumption baked in invites the model to play along and explain something that never happened. The same habit makes it fold when you push back, which is the story behind the AI that lied about the rain, twice.

The five kinds of made-up answers you’ll actually see

Anthropic’s own support page warns that Claude can “display quotes that may look authoritative or sound convincing, but are not grounded in fact.”

When the risk is highest

• Obscure questions: small companies, local rules, less-famous people, specialist topics.
• Exact specifics: numbers, dates, names, page numbers, URLs, case citations, version numbers.
• Anything that changes: prices, plans, laws, software menus, who holds a job.
• Requests for lists of sources, especially “give me ten studies that prove…”
• Questions that assume something false.
• Long answers, where one invented detail hides in paragraph six.
• High-stakes topics like health, law and money, where a single wrong detail costs the most.

The lowest-risk jobs are rewriting your own text, summarizing a document you pasted in, and brainstorming. The more an answer depends on material in front of the model, the less room it has to invent.

What actually reduces AI hallucinations

You can’t switch hallucinations off. Anthropic’s guide to reducing hallucinations says its techniques “significantly reduce hallucinations” but “don’t eliminate them entirely.” Three habits do most of the work, though.

Ground it in a document

Paste or upload the source (the contract, the policy, the report) and ask your question about that text. Better still, ask it to quote the exact sentences it’s relying on before it answers. Anthropic recommends this for long documents: pull word-for-word quotes first, then do the task using only those quotes.

Turn on search and ask for sources

For anything current, make the tool search the web and link its sources, then click at least one. A link only helps if the page says what the AI claims.

Give it permission to say “I don’t know”

If models learned to guess because guessing was rewarded, change the rules for your conversation. Anthropic’s documentation calls this a simple technique that “can drastically reduce false information.” Paste this before any question where facts matter:

Answer the question below using only facts you are confident about.

Rules:
- If you don't know or can't verify something, say "I'm not sure" instead of guessing.
- Separate what you know well from what you're inferring.
- If my question assumes something that isn't true, tell me before you answer.
- End with a short list of any names, numbers, dates or quotes I should double-check.

Question: [paste your question]

Answer the question below using only facts you are confident about.

Rules:
- If you don't know or can't verify something, say "I'm not sure" instead of guessing.
- Separate what you know well from what you're inferring.
- If my question assumes something that isn't true, tell me before you answer.
- End with a short list of any names, numbers, dates or quotes I should double-check.

Question: [paste your question]

Reasoning models help too, but they don’t fix the problem. OpenAI says GPT-5 has “significantly fewer hallucinations especially when reasoning, but they still occur.” For the full set of habits, see how to improve AI accuracy: 7 fixes. If ChatGPT is your main tool, these ChatGPT settings and prompts build the same guardrails in.

Four myths about AI hallucinations

• “It’s lying to you.” Lying needs intent, and the model has none. Treat it like a sharp colleague who sometimes misremembers with total confidence.
• “The next update will fix it.” OpenAI’s researchers say accuracy will never reach 100%, because some questions can’t be answered. The real fix is a model that declines when unsure, and that’s still a work in progress.
• “A paid plan makes it accurate.” Paying buys more usage and stronger models, which can help. No plan makes a model right about facts it never learned.
• “If it cites a source, it’s true.” Citations can be invented, or real but misread. OpenAI’s help center says search results and citations “can be incomplete, outdated, or incorrect,” and tells you to open the source and check it.

Try it now: catch one in the wild

Pick a slightly obscure topic you know cold: your town’s history, a niche hobby, the details of your own job. Ask an AI three specific questions about it with web search off. Note which answers are right, vague, or confidently invented. Then ask again with the “I’m not sure” prompt above and compare.

Once you’ve watched it happen on a subject you know, you’ll spot it on subjects you don’t. When you need to check an answer quickly, run the 3-minute AI fact-check workflow or paste the answer into the free hallucination checker prompt.

Related guides

• Your AI Is Lying to You: 3 Prompts to Catch Hallucinations
• I Asked AI If It Was Going to Rain. It Lied. Twice.
• How to Improve AI Accuracy: 7 Fixes That Actually Work