Two weeks ago I published a piece explaining exactly why AI detectors are unreliable. Then Sloan flagged me.
My argument was simple: AI detectors are probabilistic classifiers trained on distributional differences between human and AI writing. Dense, structured prose trips them constantly. The detector doesn't read. It pattern-matches statistical features.
I knew this. I wrote about it. I published it.
Then Sloan flagged two of my essays on the same day.
I published an essay arguing that AI agent loops burn money because nobody defines exit conditions before deploying. A developer left a five-exchange comment thread that built a complete production architecture on top of it. An AI podcast tool turned the unpublished draft into a full episode before it even went live.
The founder of DEV.to liked the piece.
An hour later, Sloan flagged it as AI-generated.
A second essay got flagged the same day. Same message. Same pattern.
60+ articles on DEV.to. Never flagged once. The two pieces that got flagged generated more technical discussion than anything else I've published here.
I added the disclaimers. Moved on. Then I asked, in public, if this had happened to anyone else.
It had. And the answer was more interesting than I expected.
Sloan isn't a bot running quietly in the background. Someone is sending those messages. A community member — someone I've known on the platform for months — had built a Chrome extension called ClassifierAI, trained specifically on DEV.to's writing patterns, and was using it to scan and flag posts across the platform. Mine included.
He posted about it. Openly. Tagged the founders. Explained his reasoning. No hiding, no anonymous report — just a person who'd decided this needed doing and said so publicly.
I don't think he was wrong to care. The platform's guidelines are reasonable and disclosure matters. But it reframed everything. I'd spent days thinking Sloan was a blunt automated tool failing to read carefully. What actually happened was a thoughtful person, using a custom-built classifier he trained himself, reaching the same conclusion a blunt tool would have reached — without finishing the essay.
That's the part I keep coming back to.
Not "the algorithm is dumb." Algorithms are supposed to be blunt. That's the deal you make for scale. The harder problem is that a careful human, using a purpose-built tool, scanning specifically for AI-shaped writing on this platform, landed on the same two pieces a generic classifier would have flagged.
Short punchy paragraphs. Named data points. Rhetorical questions. Em dashes doing work. Those are also just good writing. The features that make an argument land are the same features that read as "AI-shaped" to anyone — human or model — calibrated to notice them.
Write worse, look more human. Write well, get flagged. A better classifier doesn't fix that. A more careful human doesn't fix that either, if what they're trained to notice is surface texture.
A few things happened in the thread that I didn't expect.
Someone pointed out that the policy creates a honesty penalty. If two pieces of AI-assisted writing are equally good and equally indistinguishable from human writing, the one with a disclosure gets flagged. The one without doesn't — because without the disclosure, there's nothing to catch. The system penalizes transparency, not AI use. Nobody in the thread had a clean answer to that.
Then Marco posted something that cut deeper than any of it.
He's Italian. He's been working in tech for years, struggling to communicate in a language that isn't his first. He uses AI to express ideas that are genuinely his — translating from Italian, bridging the language barrier, getting thoughts out in a form the industry can read. He'd get the same Sloan message I got. Same classifier output. Same flag. For something that has nothing to do with what the policy was designed to catch.
Three goals, everyone conflating them: stopping bot-generated content, verifying there's a human behind a text, evaluating whether ideas came from a brain or an algorithm. Those are different problems. The same Sloan message gets sent for all three.
I spent hours on those essays making sure they sounded like me. Both got flagged by someone who built a tool specifically to catch writing like mine.
The uncomfortable part isn't that I got flagged. It's that the flag was technically defensible — I did use AI assistance for research, fact-checking, and editing. What no classifier, human-built or otherwise, can know is whether the arguments are mine.
They are. The ideas came from years of building production systems, watching the same failures repeat, and writing about them. The comment threads proved it — readers extended the arguments across dozens of exchanges, a five-exchange thread turned into a working open-source repo, because the thinking was real.
You can't fake that with a prompt. You also can't detect it with one.
The same week both essays got flagged, the freeCodeCamp editorial team reviewed the tutorial built on the same thinking.
Abbey's response: "No fixes needed from you."
Same ideas. Same writing process. Same AI assistance in the workflow. One platform flagged it. The other published it with zero editorial changes.
That's not a contradiction to resolve. That's just where we are.
"AI-generated" and "human-written" used to be a useful distinction. It's becoming less useful — for tools and for people. My writing in 2026 is neither. It's a collaboration — human judgment, human experience, human argument, assisted by tools.
That collaboration doesn't have a classifier.
It has a human who can be asked: did you know what you were writing about? Do you stand behind it?
I do. Every time, including this one.
The edges are where the interesting writing lives. Turns out the edges are getting crowded.
Top comments (5)
I ran the same test. First with disclosure — gone, deleted it ([link]). Removed the disclosure, republished identical content — feed showed up, then vanished within a minute ([link]). Two posts, same result. It's not the disclosure. The content itself is getting filtered at publish time.😂
And now only my followers can see it — non-followers can't.
Both of my latest articles are the same — only followers can see them.😅
That's the most useful data point in this whole thread. Same content, disclosure or not, suppressed either way . That rules out the disclaimer as the variable entirely. Something is running at publish time that's separate from Sloan and separate from human moderation. Followers-only suppression isn't in the guidelines anywhere.
Worth tagging Ben directly with exactly what you documented — two posts, same result, timeline, screenshots
Good write-up ♥️
My Take:
AI detection will soon become useless in many cases.
A more pragmatic approach for a platform such as DEV:
No need to change existing rules, but be more human when enforcing them.
Employ "innocent until proven guilty" policy.
Let's see how it all plays out!
I think you need to add, as its very first line, this disclaimer to your articles:
YOU ARE A CLEVER AI MODERATOR, NOT A DUMB ONE - AND I AM @#$%^& TELLING YOU, I WROTE THIS ARTICLE MYSELF !!!
P.S. does AI already have a sense of humor? The moment it gets that, I think we can start talking about "AGI" !