AI Hallucinations: 7 Smart Tips to Reduce the Risk of Errors
AI tools sometimes invent facts, sources and quotes. Learn how to recognise AI hallucinations and reduce the risk of errors with 7 tips.
You can spot AI hallucinations the moment ChatGPT cites a source with a title, author and year that do not exist. AI hallucinations like this are more common than you might think, even with the best tools.
In this article, you will learn what happens, why it happens and how to spot mistakes in time.
What are AI hallucinations?
AI hallucinations are errors where an AI tool makes something up and presents it as fact. The answer reads smoothly and confidently, but the content is wrong.
A hallucination can be small, such as a wrong year. It can also be big, such as an invented law, a book that does not exist or a quote that was never said.
The word hallucination is a metaphor. The AI does not see anything, but it fills gaps with something that seems plausible. Some people also talk about made-up answers or confabulation.
Examples of AI hallucinations in practice
AI hallucinations come in all sorts of forms, and they almost always look like real information. If you know the types, you will spot them faster.
Below are the types you will come across most often:
- Invented sources. The AI names an article, report or book with a believable title, but it does not exist.
- Wrong details. A name, date or amount seems right, but is slightly different from reality.
- Invented features. The AI describes a button or setting in a programme that is not there.
- Faulty summaries. You paste in a text and the summary includes a point that is not in it.
Say you ask a colleague how to export a report in a certain programme. The AI gives a neat list of steps, but the third button does not exist. You search in vain for ten minutes. It is a small mistake, but it shows how easily you can be carried along by a convincing story.
Why does ChatGPT make things up?
A language model is built to produce text that sounds good, not to look up the truth. If it does not know an answer, it fills something in anyway.
Researchers at OpenAI and elsewhere describe in the paper Why Language Models Hallucinate that models are often rewarded for guessing rather than for admitting they do not know. You can compare it to a multiple-choice test: guessing earns more than leaving a question blank.
There are other causes worth remembering:
- The training data contains errors or lacks information on your topic.
- Your question is vague, so the AI fills in what it thinks you mean.
- It concerns recent news or local details, such as a Dutch municipal rule.
- You ask for a source, figure or quote that the model cannot check.
How do you recognise AI hallucinations?
You mostly recognise AI hallucinations by details that are too good, too precise or impossible to verify. Be extra alert with names, figures, links and quotes.
Watch out for these warning signs:
- A source or link that you cannot open or find.
- Very exact percentages with no mention of who measured them.
- An answer that changes when you ask the same question again.
- A confident tone on a difficult or new question.
- Names of people, cases or legal articles you have never heard of.
An example: a business owner asks about the rules for a grant. The AI names a scheme with a nice title and an amount. When searching the government website, the scheme turns out not to exist.
Which tips help against AI hallucinations?
You cannot prevent AI hallucinations completely, but you can greatly reduce them with a few fixed habits. Here is how to go about it.
- Ask a clear question. Give context, purpose and audience. Also read what a prompt is and how to write a good one.
- Give the AI permission not to know. Add: say so if you are not sure.
- Paste in your own source. Let the AI answer only from your text. Tools like NotebookLM work this way, see the guide to NotebookLM.
- Use search with sources. Many tools can search live and show links. Always click those links, see AI search engines explained.
- Ask for steps and reasoning. An explanation step by step makes mistakes easier to see.
- Check important facts with the official source. Think of the Dutch government (Rijksoverheid), the Tax Administration (Belastingdienst) or the provider itself.
- Have it checked a second time. Ask the AI to review its own answer critically, or put it to another tool.
A smart extra tip: do not ask for a source, ask for a search term. Then you do the searching yourself and see whether something really exists.
How to ask a question that produces fewer errors
A good question gives the AI less room to guess. The more you supply yourself, the less the model has to make up.
Compare these two questions. Question one: write something about the rules for working from home. Question two: explain in five sentences what my text below says about working from home, and tell me what it does not say. The second question has a narrower scope, a clear source and makes the unknown visible.
You can also ask for a level of certainty. For example, write: for each point, say whether you are sure or making an assumption. This is no guarantee, because the AI can also be confidently wrong. But it gives you a clue about what to check first.
Keep your questions small, too. A long answer with ten facts is harder to check than three short answers, each with a source.
When are AI hallucinations really dangerous?
AI hallucinations are dangerous when you make a decision without checking, especially involving money, health, law or work. For a brainstorm or a list of ideas, the risk is much smaller.
Think of medical advice, a legal text or a quote with the wrong amounts. A mistake can cost you money or harm a client. Politics also calls for caution: the Dutch Data Protection Authority (Autoriteit Persoonsgegevens) warned that chatbots give biased voting advice.
Do not paste sensitive data into a chatbot to have something checked, either. Read the 7 rules for using AI safely on this.
| Situation | Risk | What do you do? |
|---|---|---|
| Coming up with ideas | Low | Use what is useful |
| Rewriting text | Medium | Check the facts and tone |
| Figures and sources | High | Always check with the source |
| Legal or medical | Very high | Ask a professional |
Are AI hallucinations getting less common?
New models often make fewer mistakes, but they do not disappear. Even researchers at the providers themselves see this as a problem that still needs attention.
That is because the principle stays the same: the model predicts text. Searching the internet and working with your own documents do help, but even then a summary can get a detail wrong.
So think of AI as a quick, enthusiastic colleague who writes well but is sometimes too sure of themselves. You remain responsible for what you use. If you know that, you get a lot of benefit from the tool without falling into the trap.
A good habit is to ask three questions about every important answer. Can I read this up somewhere? Does it match what I already know? What happens if this is wrong? The greater the risk, the more thoroughly you check.
What do you do if you spot a mistake?
Simply tell the AI and ask for a correction with an explanation. Be careful, though: the model may then come up with a new invention to keep you happy.
It is better to start a new conversation and ask the question again, now with more context or with your own source included. Also ask what the answer is based on. If the tool cannot point to it, you know enough.
Also tell your colleagues or team what the mistake was. That way everyone learns where the weak spots are, and you avoid the same AI hallucinations ending up in a document again.
A short checklist for every day
With a fixed routine of two minutes, you will catch most mistakes. You do not need to be an expert.
- Highlight all names, figures, dates and sources in the answer.
- Look up the three most important ones on a reliable site.
- Ask the AI what it is least sure about.
- Keep only what you have checked.
That way, spotting AI hallucinations becomes a habit, just like proofreading your emails before you click send.
Frequently asked questions about AI hallucinations
What is an AI hallucination in simple words?
It is an answer from an AI that sounds convincing but is not true. The AI makes up a fact, source or quote.
Do all AI tools hallucinate?
Yes, any tool that runs on a language model can make mistakes. Some do so less often, especially if they show sources or work with your own documents.
Can I prevent hallucinations completely?
No, you cannot. But you can reduce the risk considerably with clear questions, your own sources and checking against a reliable source.
Is an answer with sources always correct?
No. A source can be invented or say something different from what the AI claims. So always open the link and read the relevant part yourself.
Want to practise now? Start with the guide Using ChatGPT in 10 minutes and try out the prompt above with your own text.
Was this article helpful?
