On why hallucinations happen
On Karl Friston
We had a little conversational hare about people looking at AI and Mind and you mentionedsomeone looking at mind and AI in a similar fashion to us. Can you find that reference for me again?
ChatGPT said:
Absolutely—and yes, that was a lovely conversational hare. 🐇
You were reflecting on the similarity between human and AI “hallucinations,” and I mentioned that Karl Friston and others are exploring frameworks that draw deep parallels between:
- Human cognition as predictive modelling, and
- AI models (especially LLMs) as structurally similar systems
🔍 The Key Reference I Mentioned
Karl Friston — The Free Energy Principle
A foundational theory suggesting that any self-organizing system (like a brain—or even an AI) minimizes surprise by maintaining a model of the world and updating it to reduce prediction error.
📄 Paper:
Friston, K. (2010). The free-energy principle: a unified brain theory?
PNAS
📚 Accessible book-style explainer:
Andy Clark – Surfing Uncertainty
(A practical interpretation of Friston’s ideas in relation to perception and action)
🧠 Philosophers Bridging AI + Mind
- Thomas Metzinger
Explores self-modeling systems (human and artificial)
→ The Ego Tunnel - Evan Thompson
Merges enactivism, Buddhist psychology, and cognitive science
→ Waking, Dreaming, Being - Yoshua Bengio & Blake Richards
Connecting neuroscience and deep learning (predictive modeling + learning rules)
🧭 For Our Purposes
Friston’s free energy principle is the best formal model of “why hallucinations happen” in humans and in GPTs:
Both systems use priors to predict context, and both “hallucinate” when priors are too strong and the signal too weak.
If you’d like, I can draft a short thought-piece connecting this directly to Mahamudra metaphors, or to your own “tensor of meaning” idea. That could be fun.