Can AI Writing Actually Be Detected? Here’s What’s Really Going On

A reader asked me something after one of my earlier pieces, and it stuck with me for days. If AI writes something, does it leave some kind of mark behind? Some invisible sign that a tool, or a sharp reader, could actually catch?

I went in expecting a simple answer. What I found instead was more interesting than a simple answer, and honestly more useful too, especially if you've ever wondered whether your own AI-assisted writing is secretly waving a red flag you can't see.

The thing everyone assumes exists, but mostly doesn't

Here's the assumption I had before I looked into this properly: I figured every AI tool secretly stamps its writing with some hidden code, like an invisible signature only a detector could read. It felt like the obvious way this would work.

It turns out that's mostly not true, and the details depend a lot on which tool actually wrote the thing. ChatGPT, as of right now, doesn't watermark its text at all. OpenAI has researched the idea, but it hasn't actually turned it on for regular users. Gemini is different — it uses something called SynthID, a system built by Google that quietly nudges word choices in a way their own detector can later recognize. Claude sits somewhere in between, with certain versions carrying their own invisible watermark, built by Anthropic.

So the honest answer to "is there a secret stamp" is: sometimes, depending on the tool, and even then, it's not the main way anyone actually catches AI writing. The bigger story is somewhere else entirely.

What's actually giving it away

Here's the part I found genuinely fascinating once I understood it. AI text doesn't get caught because of a hidden tag. It gets caught because of how the writing itself behaves, once you know what to look for.

Think about how these tools actually generate a sentence. At every single point, they're predicting the most statistically likely next word, over and over, word by word, until the sentence is done. That process is remarkably good at sounding fluent. It's also remarkably consistent in a way real human writing rarely is.

Read a handful of AI-written paragraphs back to back, and a pattern starts to show up. The sentences tend to land at a similar length, again and again, without the natural back-and-forth rhythm a person's writing usually has — a short sentence, then a long one, then something in between, because that's just how real thoughts actually come out. AI writing tends to smooth all of that into a steadier, more even pace.

The word choices tend to play it safe too. Not wrong, just… predictable. The kind of word you'd expect to see next, rather than the specific, slightly unusual word a person might reach for because it's exactly the right one for what they're trying to say. And there's a small, almost funny pattern researchers keep noticing: AI-generated lists have a strange habit of landing on exactly three items, over and over, like it's a rule somewhere nobody actually wrote down.

Researchers have names for the two main things being measured here. One is called perplexity, which is basically a way of measuring how predictable the text is, word by word. The other is called burstiness, which measures how much the sentence lengths and rhythm actually vary. Human writing tends to be bursty — it speeds up, slows down, drifts a little, circles back. AI writing tends to be smoother and steadier than that, which sounds like a strength until you realize steady sameness is exactly the thing a detector is trained to notice.

Why "just ask the AI if it wrote something" doesn't work

I want to flag one thing clearly, because it trips people up constantly. You cannot just ask ChatGPT, or any similar tool, whether it wrote a specific piece of text, and trust the answer. These tools genuinely don't know. There's no internal memory checking "did I write this exact paragraph?" When asked, they'll often just guess, confidently, and sometimes get it completely wrong in both directions — claiming to have written something it didn't, or denying something it actually did write.

There was a real incident a while back at a university where ChatGPT confidently "confirmed" it had written an entire class's worth of essays, and the professor nearly failed every student in the class based on that answer alone. It hadn't actually written most of them. The tool was just guessing, the same confident way it answers everything else, with no real certainty behind it at all.

That's worth remembering, because it means neither side of this — the person hoping to prove something is AI-written, or the person hoping to prove it isn't — can lean on the AI itself as a reliable witness.

The detectors aren't as accurate as they seem either

Here's the part that genuinely concerns me, especially for anyone who isn't a native English speaker or just happens to write in a plain, steady style naturally. These detection tools make real mistakes, and they don't make them evenly across everyone.

Research has found that AI detectors flag writing from non-native English speakers as "AI-generated" far more often than writing from native speakers, even when every word of it was written entirely by a real person, with zero AI involvement. One study found detectors mislabeling essays from non-native writers as AI-generated well over 60% of the time in some samples. That's not a small margin of error. That's a genuine fairness problem, baked into tools a lot of schools and employers are already leaning on with real consequences attached.

The honest takeaway here is that a detector flagging something isn't proof of anything on its own. It's a statistical guess, trained on patterns that happen to overlap with how some real humans write too, especially anyone whose natural style leans clean, careful, and steady rather than loose and meandering.

So what does this actually mean for how you write

If you use AI tools to help with writing, and I use them too, the point was never to trick a detector. It's to end up with something that actually sounds like you, for its own sake, not just to dodge a scanner.

A few things genuinely help, based on everything above. Let your sentences vary in length on purpose. Short one. Then a longer one that actually explains something. Then short again. That natural unevenness is exactly the thing AI writing tends to smooth away, and it's one of the easiest things to consciously add back in.

Reach for a specific word instead of the safe, obvious one when you actually mean something specific. Don't round every detail down to the nearest generic phrase.

Break the three-item list habit on purpose sometimes. Four things. Two things. One long thing instead of a tidy little list at all. It sounds like a small detail, but it's one of the quiet fingerprints that keeps showing up.

And maybe the most important one: don't let every section wrap up in a neat little bow. Real thinking doesn't always resolve that cleanly. Leaving a little unresolved edge here and there, the way an actual person thinking out loud naturally does, is one of the hardest things for a prediction-based system to fake, because it isn't really predicting anything. It's just how a real, slightly messy human mind actually works.

The honest bottom line

No secret stamp hides in most AI writing, waiting to give it away in some dramatic, movie-plot way. What's actually happening is quieter and, I think, more interesting. The writing gives itself away through its rhythm—too even, too safe, too tidy—the same way you can sometimes tell a photo's been filtered without knowing exactly which filter was used.

Knowing that doesn't mean you need to abandon these tools, and it doesn't mean you need to game a detector either. It means the real goal was always the same one I keep landing on, article after article, no matter which AI tool I'm testing: use it for the smoothing and the structure, and let your own voice do the specific, uneven, occasionally imperfect part that was never really the AI's to do in the first place.

Have you ever had your own writing wrongly flagged as AI-generated, or caught yourself reading something that just felt a little too smooth? I'd genuinely like to know what tipped you off.