AI Can't Steal Human Taste: Models Make the Common, Brains Crave the New
tejaskumarlol · reddit · 2026-10-16
The author responds to the objection that "if taste is just pattern matching, why can't it be trained into AI?" The core argument is timing: a thing has to become common in the training data before a model is likely to generate it, and by then novelty-craving humans have already moved on.
Evidence:
- Human brains pay for novelty: novel images activate the dopaminergic midbrain (Bunzeck & Düzel 2006); monkey dopamine neurons fire for new events and stop after repetition (Ljungberg et al. 1992).
- Models drift toward the typical: the NeurIPS 2025 best paper "Artificial Hivemind" had 25 models each write 50 metaphors about time, yielding only two categories — "time is a river" and "time is a weaver"; RLHF reduces output diversity (ICLR 2024), partly because human raters prefer familiar text (ICML 2026).
- Markets reward "keeping up with the new": among roughly 27,000 songs on the Billboard Hot 100, highly typical songs are less likely to succeed (2017).
He also lists counter-evidence: LLM-generated research ideas were rated more novel than experts' (but only 200 of 4,000 were unique, and humans pulled ahead after execution); non-experts preferred AI poetry in blind tests; FunSearch found new math via an evaluator outside the model.
Related event: IBM Engineer Argues AI Can't Learn Taste Because Humans Crave Novelty(2 posts)→
More from AGI Musings
- Terence Tao warns AI is clearing math's open problems faster than it creates new ones — VoidStateKate · 2026-10-16
- MIT study: AI's accuracy on theoretical science problems jumped from 2% to 35% in two years — mikeflache · 2026-10-16
- Arvind Narayanan: steering AI from an 'automate everything' agenda toward human augmentation — mikeflache · 2026-10-16
- Bindu Reddy: Can you feel the SuperIntelligence? LLMs, robots and self-improving agents — bindureddy · 2026-10-16
- Ben Bajarin: Rooting for an AI bubble burst would only consolidate power among today's giants — BenBajarin · 2026-10-16
- Dean W. Ball on mechanistic interpretability and 'the beauty of something grown' — deanwball · 2026-10-16