You can't actually tell AI writing from human writing. But you can tell if someone was present when writing it.
You can catch a spam email with your partner’s name in the from box in a heartbeat. And you can catch an AI written email from your partner just as fast. You can probably even catch an AI email from your accountant or your boss’ boss just as fast. And if any of these people sent you Gemini’s output, you’d be livid.
Everyone hates AI slop.
Except, not really. In a study published in Scientific Reports in November 2024 (imagine those archaic models!), Porter and Machery found that people were more likely to call AI poems human, than they were to call human poems human. And, even as of 2023, Jakesch, Hancock, and Naaman found that across 6 experiments, people couldn’t tell AI slop from human content.
So what is it that people hate? If not the slop itself (since it’s indistinguishable), they definitely hate the “made with AI” sticker. In a July 2025 study by Zhu et al., people couldn’t tell the difference between AI vs human generated content, except when informed they definitely preferred the “Human Generated” content over the AI generated stuff.
Despite all this evidence that people can’t actually tell the difference, I’ve been called out for sending an email or document that had LLM generated content. Once, by someone I work closely with, who said “Sidharth, was this written by AI?” which is default in 2026 for “this doesn’t sound like you.”
It’s because people were never AI detectors (how could we be - the LLMs are trained on our words!). We’re person-detectors. Read enough of someone and you build a model of how they sound, without ever deciding to. It’s made out of samples - every email, every text, every argument you’ve ever had with them. And when something arrives that doesn’t fit the model, all sorts of bells go off in our heads.
It feels like forgery. Except I didn’t forge anyone in my email and the LLM that forged me was directed by me. The forgery isn’t of me as a person, but of my presence. Call it presence forgery. And that’s what makes it feel icky.
The studies don’t find this because they are all designed around a stranger’s prose. But you don’t know the stranger, so you couldn’t detect a forgery. It’s either human stranger or LLM-stranger.
This also explains why nobody minds ghostwriting. A politician’s speech or a CEO’s memo is a forgery in the strict sense, but you’ve never met the senator. You have no prior for how they should sound, and the forgery passes, because there was never any presence to forge. Now imagine finding out that your mother’s letter of life advice to you was ghostwritten.
There’s one exception, one place where people really can spot LLM writing: an expert reading in their own field. Chakrabarty et al. found in 2025 that while lay readers couldn’t tell AI from human prose, MFA-trained writers could, and rated the AI far lower in quality.
I can understand why. Often, when chatting with ChatGPT or Claude, it will say something that at first glance seems sensible, even brilliant, and then two seconds later reveals itself as completely hollow. It’s a fake Rolex. Sitting still on the counter, it’s perfect. It’s only when you watch the second hand that you see it tick instead of sweep. It’s not a matter of style. It’s the absence of thought.
So when we try to detect LLM written content from strangers, we see nothing. But when it’s from someone who knows us, it sticks out like a sore thumb.
Unless, the LLM is really good. In the same Chakrabarty study, when the model was fine tuned on a particular author, prose generated in that author’s style got past the MFA’s forgery detectors far more often - flagged as AI only 3% of the time, versus 97% for ordinary prompting. To get past forgery detectors, we need the same thing the forgery detectors need - samples. And the samples are the relationship itself.
And a good forgery isn’t forgiven. It’s only undetected. When Han van Meegeren forged Vermeers in the 30s, they were hailed as the finest Vermeers in existence, until he confessed and they all became worthless - no paint moved, but they went from brilliant to disgusting. Nothing on the canvas told anyone anything. The paintings only became forgeries when van Meegeren said so. Which is why the sticker matters so much: it is what people fall back on when the object itself has stopped confessing.1
The reason my colleague could tell I had sent him LLM-generated words was sample size - he had a decade of meetings, emails, slack threads, and documents to know me. It is also exactly what you would train a model on to make it sound like me.
Intimacy is a large sample size.
Detection and forgery drink from the same well. The people close to me are far better at detecting me. But they hold a rough impression of how I sound. The model can hold all of it, exactly, as it gets better. So, as LLMs get more intimate and start genuinely sounding like us, we too can get past our friends’ forgery detectors. But should we?
1 Notably, van Meegeren was never actually caught. He confessed only to save his life from a different charge - Nazi collaboration - and the disgust of forgeries plus the disgust of Nazis in postwar Holland actually made him a hero. Two hates make a like.