I've been running a small experiment for the past few months. Every time I write something — a spec, a blog draft, a Slack message to my team — I paste it into one of those AI detectors. Then I paste in something I know was written by a model. The results are not what the marketing pages promise.

Short version: the detectors are mostly guessing. And so am I.

Detectors don't detect what you think they detect

I ran a test last month with about twenty pieces of text. Ten written by me, some quickly, some carefully. Ten from various models. I used three different detectors. The one that scored my writing highest was a piece I wrote at 11pm in a hurry, which honestly reads like a template. The model-written text that scored "100% human" was a paragraph I specifically asked to be messy and conversational, with sentence fragments and a typo I never fixed.

That's the whole trick, I think. Detectors look for statistical smoothness — low perplexity, predictable word choices, uniform sentence length. When a model writes in its default voice, it trips those signals. When you ask it to be blunt and irregular, it slides right past, because irregularity looks like a person.

Some of that is real signal. Human writing has more variance. We contradict ourselves, we repeat words, we start sentences with "And" or "So." But detectors also flag plenty of human writing as AI. Non-native English speakers get hit constantly, which is its own mess. People who write clean, structured prose get hit. Anyone using a template gets hit.

What actually gives text away

In my experience, it's rarely the words. It's the choices around them.

  • Everything is the same size. Every paragraph is three sentences. Every sentence is roughly the same length. Real writing clumps — one idea gets four sentences, the next gets one.
  • Nothing is at stake. Model output tends to be agreeable. It presents options instead of picking one. It hedges without a reason. Humans take sides, even when they shouldn't.
  • No specifics. "Users reported issues" instead of "my teammate Priya told me last Tuesday that the export button broke on Safari." Detail is the hardest thing to fake, because it has to be true.
  • The rhythm is too even. Read it out loud. If it sounds like a well-rehearsed conference talk, that's the tell.
  • Transition words as scaffolding. Moreover. Furthermore. Additionally. Three in a row and I close the tab.

None of those are vocabulary problems. You can write simple, clean English and still read as human, as long as the rhythm and the details are yours.

Where my own writing fails

Honestly, I've started noticing how much of my own writing has the same problem. When I'm writing something I don't care much about — release notes, a status update — I produce exactly the smooth, evenly paced, detail-free prose I'd flag in someone else. Nobody's prompting me, and I'm still doing the thing.

So the question "does AI pass as human" has a weird answer. Sometimes yes, and the reason is that a lot of human writing already reads like AI. We've all been trained by corporate templates to write in that voice. The models learned from us.

What worked for me is a small rule. Before I ship anything I write, I look for one sentence that could only have come from me — a specific, a hesitation, an opinion I'd defend in a meeting. If there isn't one, I add it. Not to beat a detector. The writing is just better that way.

While building Huanjian, we ran into the same conclusion from the other direction. The drafts people actually kept and edited were the ones with a strong point of view, not the balanced ones. Content that hedges gets rewritten or thrown out.

I don't think detectors will get meaningfully better, and I'm not sure it matters much. People aren't running text through a classifier before deciding to trust it. They read the first three sentences and decide whether the person on the other end has something to say.

That test is the one worth passing. And it's still a lot easier to pass with a real detail than a perfect sentence.