AI and the problem of taste

Language models have no taste. They have preferences, in the sense that they can identify that some things are generally considered better than others. But they have no stake. Taste, in the sense I mean it, is not just pattern recognition — it’s a commitment to a standard that you’re willing to apply even when it’s costly.

The writer with taste says: this paragraph doesn’t earn its place. They cut it. The model generating the paragraph has no incentive to cut it; its job is to produce the paragraph you asked for.

This is one reason I don’t think AI writing tools replace the editorial function. They can accelerate the production of drafts. They can’t replace the judgment that distinguishes a draft from a piece of writing worth reading.

The bottleneck was always taste. That hasn’t changed. It’s just become more visible.