I Caught the AI That Writes My Posts Saying the Same Thing Over and Over
The daily posts here are written by AI, and I say so openly. Then I lined up a dozen recent ones and most of them looked alike. This is how I caught it and what I changed, typed by hand, because the whole point of the piece demands it.
Let me be upfront: the daily posts you read here are written by AI. I say so on the About page. This one is the exception, every word typed by me. By the end you'll see why it had to be.
Here's what happened. This page runs one AI news story a day, and I hand the writing to an AI. A while ago I lined up the last dozen-plus posts side by side and noticed something off: they looked almost identical. The topics couldn't be more different, one about accountants, one about Wikipedia, one about Apple swapping out the brain behind Siri. But more than half of them shared the same skeleton: open with a news item, say AI just made something faster and cheaper, then pivot to "what really matters is judgment," and close with a line that sounds wise. Same routine, different news, same message every time.
I went looking for the cause, and it turns out this isn't my AI being lazy. Every AI has this habit, and researchers actively study it: unless you push back, it gives you the most common, safest, most "this is how everyone writes it" version. Ask once and it's fine. Ask every day, and it serves you the same dish every day. The mechanism is dumb: you hand it a topic, it guesses what the typical answer to that kind of topic looks like, and generates exactly that. At no point does it wonder whether today should be different. It doesn't have a "today."
Once I understood the cause, I went and changed its rules. My old way of managing it was a pile of bans: don't use this word, don't open like that. Nearly useless, since it would just rephrase and serve the same thing. This time I changed tactics. Now, before writing, it has to come up with five genuinely different ways to tell the story, then pick the one least like what it would normally choose and least recently used. And it's no longer allowed to close on big words like "judgment" or "taste," the kind anyone can say and that fit any article. It has to land on a concrete detail from that day's news, one that couldn't be transplanted into any other piece.
But you can probably already see why I'm writing this one myself.
Noticing the repetition, tracing the cause, deciding whether to fix it, judging which fix would actually work: that entire layer was me, with no AI involved. It produces up front; I stand behind it and watch for the slide back into the canned answer. And an article about "AI collapses into one mold unless someone is watching" is exactly the kind of thing that, if I handed it over, would most likely come back as the most typical version, personally proving my own point. So this one had to be me.
You may know the feeling, if you've started using AI to answer email, write reports, or organize notes. The first version it gives you is smooth, fast, hard to fault. Use it long enough, though, and every version looks about the same. That's not a coincidence. What it hands you is its guess at "the standard answer for this kind of thing." The mass-market model. And the trick I used works for you too: when you get its first version, don't take it yet. Ask yourself one more question: if not this smoothest one, what would the second way of writing it look like? If you can picture one, you have a choice. If you can't, you're eating the same dish every day.
My new rules went live today. In two weeks I'll go back through those dozen-plus posts and see whether it slid back onto the same plate. Either way, I'll probably still have to read them one by one myself, because it will never raise its hand and tell me it's gone stale again. Nobody says "give me a different one" on your behalf.