Checking AI answers: How to spot and avoid hallucinations

What is an AI hallucination and why does it happen?

We speak of a hallucination when an AI claims something false with complete conviction. That can be a wrong year, a made-up quote or a book that was never published. The AI isn't lying on purpose. It simply doesn't notice that it's wrong.

The reason lies in how it works. A large language model has learned from huge amounts of text which word is most likely to come next. It doesn't look things up in a database of facts. It learns recurring patterns such as grammar reliably, but not rare individual facts such as the birthday of a little-known person. It then fills the gap with something that sounds plausible.

On top of that, according to research by OpenAI, the usual training and testing methods reward guessing more than an honest “I don't know”. It's like a multiple-choice test where a guessed answer can earn points but a blank one never does. In addition, models without web search only know the state of their training data, so current events are missing.

How can you tell that an AI answer might be wrong?

At first glance, hallucinations look like any other answer. These warning signs should make you suspicious, though:

• Very precise figures, percentages or dates without a source.
• Word-for-word quotes, especially from well-known people.
• Sources, studies, links or court rulings that you can't find with a search engine.
• The answer changes when you ask the same question again.
• Contradictions within the answer or with what you know for certain.
• Questions about very recent events, niche topics or little-known people and companies.

How serious this can get is shown by a case from the USA: lawyers there had to pay a penalty because they had cited rulings in court that ChatGPT had completely made up. So a confident tone is no proof of accuracy.

How do you check an AI answer step by step?

Not every answer needs a fact check. You can take a brainstorm for ideas lightly. But as soon as you pass something on, publish it or base a decision on it, this routine is worth it:

1. Mark the key statements: Which facts, figures or names would be embarrassing or costly if they were wrong?
2. Ask for the source: Ask where the information comes from, ideally with a link.
3. Open the source yourself: Check whether the page exists and whether it really says what the AI claims.
4. Cross-check: Look for a second, independent source, for example the official website of a public authority or a manufacturer, or a specialist portal.
5. Recalculate figures and check how up to date they are: Do totals and percentages add up, and how old is the data?

Which tools help you back up AI answers with sources?

It's easiest to check answers when the AI provides its sources straight away. Among our research tools you'll find suitable helpers for this, many of them free:

• Perplexity: an AI search engine that backs up its answers with sources. Free with 3 Pro searches a day.
• ChatGPT Search: web search with sources right in the chat, free with limits.
• Google Gemini (Deep Research): creates multi-page reports with sources, with limits even without a subscription. Our glossary explains how Deep Research works.
• Gemini Notebook: you upload your own documents and ask questions about them. The answers are based on your sources rather than on the model's general knowledge.
• Elicit and Consensus: specifically for scientific studies, each with a limited free version.

If you need this regularly, you can automate it: the research agent delivers a briefing with sources on a topic, and a knowledge base for your own documents answers questions from your files with source references. Important: source references can be wrong too. Sometimes the AI summarises a real source incorrectly, which is why step three of the routine always applies.

Which prompts help you avoid hallucinations?

You can't prevent mistakes entirely, but clear instructions make them less frequent and easier to spot. These phrasings have proven useful:

• “If you're not sure about something, say so explicitly instead of guessing.”
• “Give the source with a link for every figure and every fact.”
• “Answer only on the basis of the following text. If the answer isn't in it, write: The text doesn't say anything about this.”
• “In your answer, separate established facts from assessments and assumptions.”
• “Which parts of your answer should I definitely check myself?”

The third prompt is particularly helpful: if you give the AI the text itself, such as an article or a PDF, it doesn't have to fill anything in from memory. You'll find more basics in the guide Writing prompts for beginners.

Which topics call for particular caution?

With some topics, a mistake can cause real harm. Here you use AI only for preparation, never as the final authority:

• Health: Always clarify symptoms, medication and dosages with your doctor or pharmacist.
• Law: Laws differ from country to country and change over time. For contracts, deadlines and disputes, a law firm or an advice centre can help.
• Finance and tax: Allowances, tax rules and terms change frequently. Check them with official bodies or a tax adviser.
• Figures and statistics: AI makes calculation errors and sometimes uses outdated data. Get statistics from the original source, for example a national statistics office.
• Quotes: Word-for-word quotes are a typical source of errors. Only use what you've found in the original source.

Even so, AI can be genuinely helpful here: it explains technical terms in plain language and helps you prepare good questions for a doctor's appointment or advice session. Pay attention to what personal data you enter. You'll find tips on this under Data protection with AI tools.