Question 10 · What is AI and how does it work?
How do I know whether an AI answer is correct?
Short version
You only know once you check. AI phrases false statements just as convincingly as true ones. Checking means: look at the sources, search for a second independent source for important facts, and look especially closely at numbers.
Long version
Experts call invented facts hallucinations. They arise because a language model is trained to give a plausible answer. In 2025, researchers showed that common training and testing methods reward guessing: a model that makes something up when uncertain scores better in many tests than one that says "I don't know."
From this follow some simple rules. The more specific a piece of information is, for example a quote, a year, or a legal article, the more it is worth checking. Always click on the AI's source links, because sources too can be invented. And feel free to ask the AI: "How sure are you, and how can I check this?" That does not replace checking, but it makes uncertainty visible.
For health, legal, and financial matters, the rule is: use AI to prepare, but decide together with experts.
AI always sounds certain, certainty only comes from checking.
Sources for this question
- Kalai, A. T., Nachum, O., Vempala, S. S., & Zhang, E. (2025). Why language models hallucinate. arXiv. https://doi.org/10.48550/arXiv.2509.04664
More questions on this topic
Your question is missing?
Send it to frag [at] rogerbasler [.] ch. Selected questions will be answered and added here.