Why Isn't AI Better at Writing?

One of the amazing things about LLMs is how bad they are at writing, given how good they are at computer code. You would think that wouldn't be the case. After all, as Jill Lepore once said, writing history ain't rocket science. Bad?? You might ask. Yes, bad. AI is good at writing emails, and blog posts-- but not this one. But AI text is a kind of black parody of the best human writing.

The reason, I think, is simple: LLMs have been designed by left-brained engineers who themselves are not great writers. The only jobs they have for writers seem to be as editors, prettying up corporate content: blogs, system cards, annual reports, and the like. Techie to Fuzzy: make me talk good.

Sure, they will use gig economy companies who hire writers to grovel for basic tasks like grading AI output, like virtual undocumented laborers lining up in front of a digital Home Depot and rushing each imperious pickup truck that pulls up, begging to serve their technological superiors. But AI companies don't seem to employ a single writer to teach AI how to write in the first place.

Until last week I thought this was because AI companies had ethically and legally abstained from training their models on books. That's how naive I was. But then I discovered that they had pirated one of my books, like thousands of others. So that's not the issue.

The bigger issue is that, having scraped, harvested, and stolen billions of words of human-generated prose, they have no idea what to do with it. They mix it all together in their incredible neural networks and produce the same bland mush every time. It's like Steve Martin's depiction of a McDonald's that uses the same slop for hamburgers and cardboard cartons. That process makes for good emails, and lousy writing.

How would I fix that? God, I have so many ideas. If you pay me $250,000+ I'll teach you. It ain't rocket science.