A study by OpenAI provides mathematical evidence that hallucinations in AI models cannot be completely eliminated. Language models can produce false statements with high confidence, and according to the research, this behaviour is mathematically unavoidable. Because it can persist even with perfect training data, hallucination is not simply a side effect that can be removed through further improvements to the models.
For users, this means that appropriate safeguards are needed to avoid relying on incorrect information when working with language models. For example, a model is more likely to hallucinate the less frequently it encountered a particular fact during training. When asking research questions for which very little material is available in public sources, both the answer and the sources provided should therefore be checked particularly carefully.
By Maria Kleiner
Source: Telepolis, 30 September 2025, by Uwe Kerkow
