I often hear people say AI models lack taste. Much AI writing is bland, indeed. There are too many words, too few surprises. The writing is of course flawless, but it’s almost too polished¹. Similarly, AI jokes rarely work – they’re quite repetitive, and many have a dad-joke quality. It’s also unclear whether they have research taste. For example, a paper from November 2025 found that LLMs largely failed in matching humans judgements of interestingness.

Naturally, Silicon Valley is trying to teach models good taste. In a META paper this week, the authors trained models to optimise for the metric ‘interestingness’, defined as proof length over description length in Lean². Back in February this year, the YC founder Paul Graham tweeted he thought taste would become more important. Greg Brockman, president of OpenAI, called it ‘a new core skill’.

Taste is usually defined as the ability to discern quality, although, interestingly, until the 1700s, taste was just a physical sensation. During the Enlightenment, the word ’taste’ acquired a new meaning, with Voltaire defining taste as ‘capacity for discriminating’ in the Encyclopédie.

The obsession with taste is hardly new. With the abundance of the industrial revolution came the need to discern. In 1835, a parliamentary committee was formed in Britain to investigate how to improve British taste (i.e., make it as good as French taste). Some fifty years later, the Arts and Crafts movement took off, largely as a reaction to industrial mass production. Normal cotton textile was the AI slop of the 1800s, and it needed to be resisted.

I agree with Paul, Greg and the committee that taste is essential for businesses. Now that everyone can produce shiny b2b software, perhaps you’re pitching your taste rather than your product to investors. So, can taste be learned by AI?

There is one obvious argument for ‘no’: taste requires human perception. But how much perception is needed for taste? Can a hearing-impaired have taste in music? And what do you even mean by perception? Reductio ad absurdum by ‘you need to do philosophy of mind’. Likewise, any argument saying taste requires an appreciation of human culture would fall (‘you need to do philosophy of mind’).

If it’s all just about discernment, I don’t see why AIs couldn’t learn taste. Cultivating taste within, say, painting largely comes down to studying many paintings, and AIs can ‘study’ more paintings within a second than a human in a lifetime³. At the very least, AIs can make intelligent-sounding comments about an enormous number of paintings.

There’s actually some evidence suggesting AIs could learn taste. The paper from November 2025 also noted that AI could generate engaging problems after a screening process. Remember that November 2025 was a long time ago – it was before AI began solving Erdős problems – so I’d expect AIs doing much better now. Moreover, the META paper also had moderate success: optimising for interestingness led to more novel results (affably described as ‘out-of-distribution math’).

However, there’s one possible way out for the believers in human-only taste. As highlighted in this IoT episode on taste, taste is closely related to the notion of exclusivity. If taste could be learned by AI, then we’d be flooded with… slop with style? Perhaps there’s no such thing – I don’t know. But if the cheap must be tasteless, then normal LLMs have no hope of learning taste.

Regardless of whether taste can be learned by AI, you can’t automate the production of the exclusive. Whatever requires human attention, then, might become a luxury good.