It goes without saying GPT lags at writing. Even a pro subscription of Astra produces prose that is predictably AI, and easily identified by readers and software like Pangram as such. It has usual “failure modes” such as wordiness and lack of affect. Why can’t it do writing as well as it does math or coding? That is the million dollar question.
I think this needs to be reframed; it writes well…for an audience of other agents. Agents have no concept of wordiness or overuse of hedging language. If anything, it’s a positive, because it minimizes the risk of being misunderstood, which is good, but when overdone, leads to bad writing. A lot of what goes into writing is what is unsaid.
With coding or math, the output is what matters, so the code being bloated or bad math notation is not fatal. End users cannot see it or will just work around it, albeit frustratingly . AI math papers are sometimes described as confusing but the results are correct most of the time. Writers tend to have certain tics or patterns , whereas Ai tends to be genetics or bland.
Writing does not play this way: the cards are always face up. It’s not so much if the author is right or wrong, but also how it’s presented, unlike coding or math, where the output or correctness matters more. This introduces much more difficulty, as style is inherently subjective. Bad writing can kill the piece even if the author’s argument is sound. So AI has to take into account all of this.
It also doesn’t help that readers’ tastes can be capricious or contradictory. Readers hate overly pretentious or stylistic prose, but they also complain about bland, AI-generated prose. Striking the right balance is possibly even harder than advanced math. This is why writing as a profession still exists despite AI. “Writers” are those who consistently get it right, and are paid accordingly.
Anyone whose has used AI for writing, such as pasting their essay into Claude or GPT, can observe that it gives a lot of suggestions. At some point, you’re faced with a choice: either spend hours or even days writing what is effectively a second essay, or just have the AI rewrite the whole thing, fixing all the errors and possible factual inaccuracies, but at the cost inauthenticity. For many people, this is a worthwhile compromise, and can explain the overuse of AI in writing.