Why AI Can't Replace Human Taste and Creativity

AI can’t replace human taste and creativity because human beings crave novelty. A model makes whatever was most common in what it learned from. By the time something new is common enough for a model to make it, we’re bored of it and already chasing the next new thing.

That’s more or less what I told somebody on 6 October in Prague, after my talk at FrontKon (FrontKon with a K!). He asked me a really good question and my answer got kinda controversial. So here’s all of it written down properly, with the research I didn’t have in my pocket on the night.

“Can’t we just train taste into AI?”

The talk was about taste. It’s the same one I gave as The Critical Advantage With AI at Agent Conf in Warsaw in September (at FrontKon it ran in the Frontend After AI: The New UX slot). In it I define taste as a trainable skill where you tacitly judge quality against external standards. And I spend most of the talk on how you train it: immersion, valid cues and corrective feedback. The long version with every source is in What Is Good Taste in Software?.

So afterwards a guy came up to me and asked, roughly:

Hey, but if taste is just pattern matching, and if taste is a trainable skill, then can we not just train taste into AI agents? And then they take taste too, and we’re left with nothing?

And I said “No.”

He asked “Well, how come?” which is fair!! He was taking my own definition seriously. If taste is judgment you train and judgment is pattern matching, then training pattern matchers is literally what AI labs do all day.

My answer had 2 parts. The first one is biology: we love novelty and we can’t switch that off. The second one is the controversial part and it involves Hitler. I’ll get to it.

Why do humans crave novelty?

Humans crave novelty because our brains give new things a bonus: novel things activate the dopamine-producing part of the midbrain and that response fades once the new thing is familiar.

In Prague I called this our dopamine schedule. Andrew Huberman talks about “so-called dopamine schedules” in his episode on dopamine. He puts it better than I did: “This is why when you repeatedly engage in something that you enjoy, your threshold for enjoyment goes up and up and up.” How good something feels “depends on the height of that peak relative to the baseline.” His example is scrolling Instagram: you see a post you love, then the next one is “Not that interesting”, even though “had you arrived at that second thing first, you might think that it was really interesting.” That’s pretty much what I told the guy in Prague!!

3 findings show what’s going on underneath:

1. New things light up the midbrain. In 2006 two neuroscientists at University College London (UCL) scanned people’s brains while they looked at pictures and found the dopamine areas of the midbrain “were indeed driven by stimulus novelty rather than other forms of stimulus salience” like rareness or emotion. Babies do this before they can talk: in a 1964 study in Science, infants 2 to 6 months old looked at the same pattern less and less over 10 one-minute showings, while new patterns kept their attention.
2. Novelty steers our choices even when it shouldn’t. In a 2008 study in Neuron, another team at UCL found that “even when the degree of perceptual familiarity of an option is unrelated to choice outcome, novelty nevertheless drives choice behavior”. They called it “a newly identified and quantifiable source of human irrationality”, which is hilarious… it’s literally us. The brain starts every new option off with an optimistic value so that we go and try it.
3. It wears off. In 1992 Wolfram Schultz’s lab recorded dopamine neurons in monkeys while an empty box opened in front of them. The neurons fired the first few times. Then “Neuronal and behavioral responses disappeared on repeated stimulus presentation.” Schultz’s later work is why we talk about dopamine as a prediction error: dopamine neurons fire for what’s better than predicted and stay flat for what’s exactly as good as predicted.

So the brain pays out for surprise, not for repeats. (Dopamine’s job is mostly wanting, by the way. Kent Berridge’s work says the dopamine system is necessary for wanting rewards but not for liking them. A 2015 review he co-wrote even says it “may not generate pleasure after all”. Its role in pleasure is still argued over: in a 2019 study, changing people’s dopamine levels changed how much they enjoyed music.)

That machinery is also behind a lot of compulsive behavior. I named porn and sex in Prague and there’s research on exactly that. Show someone the same erotic clip again and again and their arousal drops: a 1985 study controlled for fatigue to show it’s habituation. In 2016 Paula Banca and colleagues scanned men with compulsive sexual behavior (what’s usually called porn addiction) and found they habituated faster to repeated sexual images. And the faster they habituated, the more they preferred new ones. Their own words: “the Internet provides a broad range of novel and potentially rewarding stimuli.” Wild.

It’s biology, but it isn’t the same amount in everyone. Novelty seeking is a personality trait that varies from person to person and it tracks dopamine receptors in the midbrain. We don’t want everything new either: a 2010 study in PNAS found people prefer familiar faces and new scenery. But nobody gets to opt out of the bonus.

Why movies keep getting more unhinged

I’ve said this on stage before. On the accessibility panel at Future Frontend in June I said we “got bored of the movie formula” (the Marvel one especially) and that the movies doing well are the weird ones (recording at 2:09): “by weird I mean they’re novel. And as human beings we’re we’re born and raised novelty seeking.”

The one everybody’s talking about right now is You Can See Everything, A24’s documentary about Elizabeth Holmes of Theranos, directed by Nathan Fielder and Lance Oppenheim. It doesn’t open until 16 October and its official trailer already had 15.4 million views on A24’s channel when I checked on 6 October. It’s 100% on Rotten Tomatoes from 17 reviews. Owen Gleiberman called it “a one-of-a-kind reality-TV-as-hangout-movie psychodrama circus” in Variety. The trailer is unhinged as fuck. It makes your skin crawl:

Before that it was Everything Everywhere All at Once and its universe where everybody has hot dogs for fingers. It made about $145 million worldwide on a $25 million budget (The Numbers) and won 7 Oscars in March 2023, Best Picture included. It was A24’s biggest film until this year, when Backrooms, directed by Kane Parsons, a YouTuber who was 20 when it came out, became “officially A24’s highest-grossing worldwide release” (Variety, 9 June 2026):

Backrooms (2026)

$400.7M

Marty Supreme (2025)

$192.4M

Everything Everywhere All at Once (2022)

$145M

Franchises still top the charts overall. But the surprises keep being the strange ones: Sinners was the only original story among the 10 biggest films of the first half of 2025 (Variety). And the record summer of 2026 had Obsession and Backrooms as its sleeper hits while live-action Moana and Supergirl flopped.

This is an old idea in psychology. Colin Martindale spent his career on it, and in a 1986 study of Italian painting he put it like this: “habituation pressures successive artists within any style to increase the impact value or arousal potential of their works.” When a style can’t get any more intense, it changes and the process starts again. In Prague I said that after unhinged we might crave something even more unhinged, or we might get so fed up with it that we crave hinged stuff again. There’s a lab version of that too: in a 2009 study, people who’d been shown a huge number of complex images then found simple ones more beautiful, and the other way round.

But we don’t want maximum new. The industrial designer Raymond Loewy called the sweet spot “most advanced, yet acceptable” (MAYA). A 2003 study found people “prefer novel designs as long as the novelty does not affect typicality”. When Noah Askin and Michael Mauskapf analyzed about 27,000 songs from the Billboard Hot 100, songs “sounding too much like previous and contemporaneous productions” were less likely to succeed and songs with “optimal differentiation” were more likely to reach the top. (The big datasets show change coming in bursts rather than a steady climb: US pop changed fastest around 1964, 1983 and 1991.) We crave the new thing we can still follow.

Can AI replace human creativity?

Not on its own. Left to itself, a model drifts toward the most typical answer, while what feels new to people is whatever isn’t typical yet.

Here’s why models drift there. A language model learns how likely each next word is, given everything it read. Always picking the likeliest one gives you text that’s “bland and strangely repetitive”. And the tuning that makes models helpful narrows them further: reinforcement learning from human feedback (RLHF) “significantly reduces output diversity”. Part of the reason is us! A 2026 paper found “annotators systematically favor familiar text” so the models learn that familiar wins.

You can see the result. The Artificial Hivemind paper, which won a best paper award at NeurIPS 2025, asked 25 models for “a metaphor about time” 50 times each. They mostly landed in 2 clusters: time is a river and time is a weaver. People using AI drift together too. In Wharton’s brainstorming experiments, published in Nature Human Behaviour in 2025, 9 people using ChatGPT independently named their toy “Build-a-Breeze Castle”. And in a 2024 study in Science Advances with 293 writers, stories written with AI ideas were rated more creative and were “more similar to each other than stories by humans alone”. The most creative writers barely got anything out of the AI at all.

There’s good evidence on the other side too.

• In a study of over 100 NLP researchers, ideas a model generated were judged more novel than experts’ ideas. But in that same study only 200 of the model’s 4,000 ideas weren’t duplicates, and when experts spent 100+ hours actually executing the ideas, the human ideas came out ahead.
• In a 2024 study with more than 1,600 people, readers couldn’t tell AI poems from poems by famous poets and rated the AI ones higher. The authors think it’s because the AI poems were simpler and easier to follow. So slop can absolutely win with a casual audience.
• Google DeepMind’s FunSearch found genuinely new mathematics with a language model. It did that with an evaluator outside the model scoring every attempt, which is basically the job taste does.

So my actual argument is about timing. Something has to be common in a model’s training data before the model is likely to make it. Models can catch up to any style once it exists and enough people have made and written about it (a 2025 preprint fine-tuned models on an author’s books and even trained writers preferred the imitations). They can’t be the ones who make it exist first. And by the time they’ve caught up, the hot dog fingers are old.

That’s also what happened to AI movies so far. OpenAI’s animated Critterz was meant to premiere at Cannes in 2026 and missed it after OpenAI shut down Sora. Copán: La Leyenda, an AI-generated feature from Honduras, came out in May and got panned. Cartoon Brew wrote that “viewers seemed less interested in the novelty of the technology than in whether the film itself demonstrated craft, artistic intention, and human creativity.” Meanwhile films where people made the choices already use AI as a tool, like The Brutalist for some of its Hungarian dialogue and Netflix’s The Eternaut for a building collapse. The line is who made the choices.

The controversial part

I told him the reason I’m not worried about AI models overtaking us on taste is the same restlessness that helped Hitler come to power in 1933 after World War I, that helped Donald Trump win twice and that has the Alternative for Germany (AfD) coming first in German elections. And that restlessness is our biology: we love novelty.

The Nazis went from 2.6% of the vote in 1928 to the biggest party in the Reichstag with 37.3% in July 1932. And on 30 January 1933 Hitler was appointed chancellor. That came right after a crisis: Nazi votes surged where failing banks hit hardest in 1931 and where austerity cut deepest. It’s a pattern across history too. After financial crises the far right’s vote share grows by about 30% on average (Funke, Schularick and Trebesch), which doesn’t happen after normal recessions. And even without a crisis, incumbents “experience an erosion of support over time” across more than 500 elections. We get tired of whoever’s in charge.

In 2016, 39% of American voters said the quality that mattered most was a candidate who “can bring change”. In 2024, about 3 in 10 voters said they wanted “total upheaval” in how the country is run. The AfD came second in Germany’s February 2025 federal election with 20.8%, then first in Saxony-Anhalt last month with 43.8%. And it leads the national polls right now. 83% of its supporters in Saxony-Anhalt told pollsters they want fundamental change, while 85% of them worry that the way people live in Germany is changing too much. They want something new in charge so that things around them stop changing.

The novelty seeking psychologists measure as a personality trait actually leans liberal, and in Germany openness went with intending to vote for nearly every party rather than the AfD. What these elections share is people being tired of the status quo and turning to something new, sometimes something dangerous. It’s the same pull that makes us bored of a movie formula.

Slop is typical, craft is new

This ties into what makes slop slop and craft craft. In Please Stop the Slop I drew the line at whether somebody cared enough to make the choices and check what came back. Novelty is why that line matters: the choices nobody made default to the most probable ones. And the most probable thing is exactly what we’re already bored of. Craft is a person choosing something that wasn’t the most probable thing, on purpose, and checking it against a standard.

Prague also gave me half an answer to the question I ended my long taste post on, whether taste is just the aggregation of the majority. The majority’s taste is what a model can learn and it’s always a little behind the people who are bored of it.

Where does this end?

He had one more question: “Where does this end?”

I said ideally it ends where it began, with us around a campfire. In 1974 the anthropologist Polly Wiessner recorded 174 conversations among the Ju/’hoansi in northwest Botswana and published them in PNAS in 2014. By day people talked about money and complained about each other. At night by the fire, 81% of the long conversations were stories:

By day

Stories

6%

Economics and complaints

65%

Everything else

29%

By firelight

Stories

81%

Economics and complaints

11%

Everything else

8%

Storytellers matter too. Among the Agta in the Philippines, skilled storytellers had roughly double the odds of being picked as someone people wanted to live with, more than good foragers. And they had more living children.

Each of us is full of new experiences from our own lifetime. So we tell stories around the fire, we sing songs together, and maybe we watch something unhinged (or hinged) together. We chase novelty together, we chase food together. My favorite part of conferences has always been the conversation and that’s literally where this whole post came from: a guy with a good question after a talk.

The Bible has a whole book about this. The Teacher in Ecclesiastes tried everything under the sun and found that “the eye is not satisfied with seeing, nor the ear content with hearing”, which is habituation, written down long before anyone scanned a midbrain. “There is nothing new under the sun,” he says. So what’s left? Good food, a drink, good work and each other: “there is nothing better for them than to rejoice and do good while they live” and “every man should eat and drink and find satisfaction in all his labor”. And “Two are better than one”.

So I hope this ends with us closer together as a species, while all the boring stuff that isn’t new gets handled by AI.