I’ve been testing every new AI model the same way for a while now. I give it a real coding task, and more often than not it comes back with something solid, impressive results. Then I give it a pile of good writing samples and ask it to write an article or a design doc in that voice, and it fails. Doesn’t matter how strong, expensive the model is or how many examples I feed it. It always comes back sounding like AI.
For a while I assumed it was on me. Bad prompting, not enough samples or wrong instructions. So I kept tweaking it for a while, and it never got better in any way that mattered.
Software or code has a right, finite answer. It compiles or it doesn’t, the tests pass or they fail. And the model has something concrete to aim at and correct itself against. The best practices and patterns are actually what make good, maintainable software. However, repeatable patterns in writing are what make it slop and annoying.
At its current stage, AI is good at anything with a checkable output. And at things where repeatable patterns are a sign of quality. Software and coding happen to be full of those. Most other work isn’t, especially writing. So if you want a decent result from AI for writing, the best approach right now is not to use it at all.
Thanks for reading,
Adlet Balzhanov
Connect with me on LinkedIn, just use the button below. I read every message. Cheers!
As a software engineer, I tried this one as well.
The harsh truth is that we often run to ai to fix our problems these days.
And doing art which writing in our case, it sucks. Sounds artificial no values in sentences.
And in my experience, I just give it to polish, since my English is not my native language, to fix grammar.
The other day, I was trying to resolve an issue in a react project of a friend of mine.
Got stuck at a long answer and a short advice which didn’t change anything.
Copilot said change file to version b so I did but then it asserted I hadn’t done what I did.
I am ND so it got me either touch grass first or delete the whole fork 😁