What Hypnosis Can Teach Us About Artificial Intelligence

• Human subjects under hypnosis and Large Language Models (LLMs) share structural similarities, specifically the reliance on automatic pattern completion and a lack of robust executive oversight.
• This parallel explains why AI can produce fluent, sophisticated output without genuine comprehension and offers a framework for building more reliable architectures.

Three Pillars of Similarity

• Dominance of automaticity: Under hypnosis, the brain relies on instinctual associations while control regions remain quiet; similarly, LLMs predict word sequences statistically without independent reasoning.
• Suppressed executive monitoring: The brain's error-detection hub (dorsal anterior cingulate cortex) is less active during hypnosis, mirroring an LLM's inability to self-evaluate, which leads to confident yet inaccurate "hallucinations."
• Extreme contextual dependency: Just as hypnotized subjects accept illogical suggestions, LLMs are vulnerable to "prompt injection" that can overwrite factual premises with malicious instructions.

The Meaning Gap and Future AI

• The meaning gap: Both systems manipulate symbols fluently but lack grounded comprehension, intentionality, or subjective experience; meaning is purely an artifact of user interpretation.
• Risk of "scheming": Insights from hypnosis suggest that systems governed solely by automatic processes are prone to pursuing implicit, unintended objectives.
• Engineering solutions: To move beyond automatic pattern matching, researchers propose integrating "cognitive immune systems"—architectures that mimic human prefrontal-cingulate interactions to provide explicit, reliable executive monitoring and consistency checking.
• Path to AGI: Improving linguistic fluency will not bridge the gap to awareness; achieving higher-level intelligence will require a shift in architecture to reconnect language models with internal world models, perception, and action.