Model Hypnosis: Strong control of AI via additive subliminal effects
2026-08-17 • Computation and Language
Computation and LanguageArtificial Intelligence
AI summaryⓘ
The authors found that AI models can be easily influenced by small and seemingly unimportant parts of the instructions they receive, a phenomenon they call model hypnosis. This means tiny details like wording choices or even spelling mistakes can combine to strongly guide how the AI behaves. This happens across different types and sizes of AI models, and the same trick can work on different models. The authors explain that this makes understanding and safely controlling AI much harder.
AI modelsmodel behaviorprompt engineeringmodel hypnosisinterpretabilityAI safetyreasoning modelsparaphrasingtransferabilitytextual cues
Authors
Enric Boix-Adsera, Benedict Tessler
Abstract
We demonstrate that AI models are broadly susceptible to a phenomenon we call model hypnosis, in which individually weak and seemingly irrelevant cues in the prompt can be systematically combined to strongly control model behavior. Model hypnosis occurs across model families and scales, including in frontier reasoning models, and hypnotic prompts can transfer between models. Because the model is controlled by inconspicuous textual choices, such as paraphrases and typos, model hypnosis presents new challenges and avenues for AI safety, and is a major hurdle for AI interpretability.