Why I don't think AI will take our taste: we get bored of whatever is common and models make what's common

I don't think we can just train taste into AI agents and be left with nothing. Someone asked me exactly that after a talk I gave on taste in Prague last week (if taste is pattern matching and you can train it, why not?). Then I read the research to check myself. Our brains pay out for surprise. Novel pictures activate the dopamine-producing midbrain in humans (Bunzeck and Düzel, Neuron, 2006). In monkeys, dopamine neurons fire for a new event and stop responding once it repeats (Ljungberg, Apicella and Schultz, 1992). Andrew Huberman's line for it: "when you repeatedly engage in something that you enjoy, your threshold for enjoyment goes up and up and up." Models drift to the typical. The Artificial Hivemind paper (a NeurIPS 2025 best paper) asked 25 models for a metaphor about time 50 times each and got 2 clusters: time is a river and time is a weaver. RLHF, the tuning that makes chatbots helpful, reduces output diversity (Kirk and others, ICLR 2024), partly because human raters favor familiar text (Zhang and others, ICML 2026). Audiences reward the new thing they can still follow. In about 27,000 Billboard Hot 100 songs, the highly typical ones were less likely to succeed (Askin and Mauskapf, American Sociological Review, 2017). The evidence against me. LLM research ideas were judged more novel than experts' ideas (Si, Yang and Hashimoto, ICLR 2025), though only 200 of the 4,000 were unique and the human ideas came out ahead once people actually executed them. Non-experts preferred AI poems in a blind test (Porter and Machery, 2024). FunSearch found new math with a model, with an evaluator outside the model scoring every attempt. So my claim is about timing. Something has to be common in a model's training data before the model is likely to make it, and by then the people who crave the new thing have moved on. Where do you think this breaks?