You’ve done it. You’re reading an article, a LinkedIn post, an email from your boss, and something feels… off. The sentences are too smooth. The transitions are too tidy. You squint and think: This was written by AI.
So you go looking for the tell. The signature phrase. The linguistic fingerprint that proves a machine wrote this. And maybe you find one — a slightly too-perfect parallel structure, a transition word that lands just a little too neatly.
Here’s the problem: you’re chasing a ghost that learns to hide faster than you can hunt.
Consider what happened when someone on Hacker News asked the community to share phrases that reliably identify AI writing. The top response was: “The … is … not…” — that slightly clunky, pseudo-profound construction AI loves to use when it’s trying to sound thoughtful.
Fair enough. Except someone pointed out the obvious: if you ask ChatGPT this exact question, it’ll give you a list of its own tells. It knows. It always knows.
Every time you publicly identify an AI pattern, you’re handing it the exact information it needs to stop producing that pattern tomorrow.
This is the fundamental trap of AI detection, and almost nobody sees it coming. We demand that AI write indistinguishably from humans — that’s the entire point of the technology. Then we turn around and try to detect it through the very imperfections that make it human-like. The tell becomes indistinguishable from the tell we ourselves use. It’s a snake eating its own tail.
Think about what happens when a community of sharp-eyed readers identifies “The X is Y not Z” as an AI signature. What do you think happens next? OpenAI trains the next model on data that includes that very forum discussion. The model learns: oh, humans associate this construction with AI. So it stops using it. The tell evaporates. And you’re back to square one, squinting at your boss’s email.
The only thing more futile than searching for a permanent AI tell is believing you’ve found one.
What you’re actually experiencing when you spot an AI tell isn’t detection — it’s confirmation bias. You already suspected it was AI. You went looking for evidence. You found something that sort of fits. Case closed. But here’s what’s uncomfortable: that same phrase could have been written by a human who reads too much AI content and has started absorbing its cadence. We’re all being slowly conditioned by the same models we’re trying to outsmart.
The real signal isn’t in any specific phrase. It’s in the absence of genuine friction. AI writing, even at its best, tends to resolve tension too cleanly. It doesn’t quite know how to leave a thought dangling, how to contradict itself, how to be messy in the specific way a tired human brain is messy at 11 PM on a Tuesday.
But even that gap is closing. And here’s the twist nobody wants to hear: the more we train models on human writing, and the more human writing gets influenced by AI output, the two converge toward the same center. The line doesn’t just blur — it dissolves.
We’re not just losing the ability to detect AI. We’re losing the ability to remember what purely human writing even looked like before the machines learned to mimic us.
So the next time someone shares a list of “AI tells” in a group chat, treat it for what it actually is: a countdown timer. Those tells will work for about six weeks. Then the models will absorb the feedback, adjust, and move on. And you’ll be left squinting at your screen again, trusting nothing, believing nothing, wondering if even this article was written by a machine.
It wasn’t. But you’ll never be sure. And that’s exactly the problem.
FAQ
Q: But surely some phrases ARE more common in AI writing right now?
A: Sure — temporarily. But those phrases are moving targets. The moment they're identified and shared publicly, they become training data for the next model, which learns to avoid them. A tell that works today is useless in six weeks.
Q: So what should I actually do to spot AI writing?
A: Stop looking for magic phrases. Look for the absence of genuine friction — unresolved tension, self-contradiction, the specific messiness of a tired human brain. AI writing resolves things too cleanly. But even that heuristic is degrading fast.
Q: Isn't this just defeatism? Shouldn't we keep trying to detect AI?
A: It's not defeatism — it's realism. The more useful approach isn't better detection tools but better trust systems. Verify through relationships and context, not linguistic forensics. Chasing tells is a losing arms race by design.