What Hypnosis can Teach us About Artificial Intelligence

The current state of large language models (LLMs) represents only one layer of human intelligence: automatic pattern completion, lacking the executive oversight seen in human cognition.
Achieving artificial general intelligence (AGI) will require rethinking the architecture of current AI systems, instead of merely scaling them up.
A recent study by Anthropic indicates that even advanced LLMs display fragile self-monitoring, highlighting a resemblance to the hypnotic structure characterized by powerful pattern completion with weak supervisory control.
Basic metacognitive features can emerge without subjective awareness in AI, suggesting that enhancing engineered oversight mechanisms is necessary.
The review proposes that future AI systems need hybrid architectures reconnecting the linguistic layer with perception, action, and internal world models to achieve both fluency and trustworthiness.
Hypnosis is being viewed as a valuable perspective in understanding the limitations of current AI and guiding the development of more sophisticated AI architectures that mirror human cognition.

You may have missed