08.05.2026
News article
Recommend Haaga-Helia to a friend outside of EU/EEA – earn 500 €
Can an AI be prompted not to hallucinate?
Thursday 2 April is International Fact-Checking Day. The day reminds us of the importance of critical thinking and verifying information. Most of us certainly know that generative artificial intelligences based on language models can hallucinate, that is, come up with incorrect or fully invented answers. Language models have evolved at a wild pace and do not produce as glaring errors as they did just a short while ago. However, the hallucination problem has not gone away.
A solution may be sought in prompt formatting or tuning the AI tool settings. Why not write a prompt for the AI tool: "Don't hallucinate, rather say you don't know", "Only give answers that are based on verifiable facts", or something along those lines.
Body
Why can't we rely on the "do not hallucinate" prompt?
"Don't hallucinate" is just one rule governing artificial intelligence among countless others. It does not change the basic operating logic of language models. They are trained with huge masses of data to identify statistical probabilities. Their answers are not based on truthfulness, but on what kind of text is statistically probable. That is why artificial intelligence produces, say, incorrect references. It has learned that when talking about academic writing and sources, the combination of a person's name, year and publication is statistically a highly probable string. Even though the source may not exist, it looks right.
Convincing is not the same as correct
The existence of a single source is easy to verify. Assessing the accuracy of the longer output is already more demanding. Users have been found to overestimate how accurate AI responses are. The longer the answer, the more convincing the answer may seem to us. It takes solid expertise of the subject matter to be able to identify errors, gaps, and biases in a persuasive-sounding text.
It depends on the context how important it is to verify information. Let’s say you’re preparing dinner. A quick glance at a recipe given by artificial intelligence may be enough. If the list of ingredients doesn’t contain anything harmful, then perhaps you can go ahead and cook the dish. Evaluation and verification will take place at the dinner table. It pays to take data verification more seriously when the stakes are higher: developing your own expertise in the form of a course assignment or thesis, or preparing an assignment for a client, for example.
Start reading laterally
It is advisable to use a lateral reading method to evaluate the text produced by artificial intelligence. Don’t read the text from start to finish and evaluate it as a whole. Instead, do this:
- Identify the first claim made by the AI tool.
- Open another browser window.
- Check the information from a reliable source.
- Return to the AI response and repeat steps 1–4 for all claims separately.
If you want to prompt AI to tell you the truth, you can continue doing so. However, maybe you should think of the prompt more as a wish or suggestion. Perhaps the language model should be prompted in this way:
End all your replies with the following reminder: “When you evaluate this answer, keep in mind the limits of your expertise. Be sure to check the facts."
Sources
Haaga-Helia Library and Information Services 2026. How to search for information: Use reliable sources. URL: https://libguides.haaga-helia.fi/how-to-search-for-information/use-reliable-sources. Accessed: 31 March 2026.
Kalai, A.T., Nachum, O., Vempala, S.S. & Zhang, E. 2025. Why language models hallucinate. URL: https://openai.com/fi-FI/index/why-language-models-hallucinate/. Accessed: 31 March 2026.
Latest news
