One principle explains both AI vision attacks and language hallucinations

Researchers found that two seemingly unrelated AI failures — adversarial attacks on image systems and hallucinations in language models — actually stem from the same mathematical constraint: the more you compress information in an AI system, the more sensitive it becomes to small input changes. In practice, this means you can detect when a language model is about to hallucinate before it generates an answer, and you can make vision systems more robust to attacks without expensive adversarial training.

If this unifying principle holds up in practice, it offers a single diagnostic lens for two major categories of AI failure that teams currently treat as separate problems with separate fixes — potentially letting engineers catch unreliability earlier and cheaper across different types of AI systems.