I Got Stuck on the Words "These Three Articles" While Reading a Draft My AI Wrote

It was half past ten at night. I was on the couch flipping through that day's draft, the one my AI drafting tool had pieced together, meaning to give it one read-through and send it off. Second paragraph in, my eyes snagged on a line: "combining the views of these three articles..."

I pulled the phone closer to my face. Three articles? Which three? I ran back through my memory and couldn't place a single one — I never told it to reference three specific pieces. The material was just stuff I'd bookmarked over the past few months, whatever caught my eye at the time. It picked a few, stitched them together, and in doing so couldn't help but announce: "here's where I copied this from."

This kind of tool is common now: you save things that sound right when you read them, too lazy to organize them one by one, so you let an AI mash a few into a rough draft and you edit from there. What it saves you is the step of "figuring out how to start from nothing," and that's a real convenience. But that night was the first time I noticed the convenience came with a tell I'd never spotted before.

When a real person writes something, they're pouring it out of memory. You already believe some idea, so when you write it down you just state the conclusion — you've long forgotten which article first convinced you, you only remember that this is what you think now. Only writing where the material hasn't been digested, just pasted together, bothers to announce that it "combined a few sources."

Over the past few months, "how to tell if a passage was written by AI" has become its own field of study. A viral guide from this past May basically lifted Wikipedia's list of AI writing tells: the reflexive "a historic moment," "a defining highlight," turning a pair of running shoes into "a commitment to a disciplined lifestyle," and that "from X to Y" construction that sounds connected until you actually think about it — "from a problem-solving tool to an artistic expression of scientific discovery."

The most-discussed signal is the em dash. A team that runs detection research crunched the numbers: GPT-4.1 uses 10.62 em dashes per 1,000 words, 3.3 times the human baseline of 3.23. The number alone sounds damning. But the team's own conclusion is that it doesn't hold up: Mark Twain used 10.13 per 1,000 words, more than GPT. Switch to a different model and Llama 3.1 8B barely uses any. Density is too easy to game deliberately for this to be a reliable test.

All these checklists are chasing surface symptoms of word choice: some word overused, some sentence pattern repeated, tone pitched too high. Nobody mentions the thing I ran into that night — whether the writer (or, in this case, the AI) can't resist telling you "I pulled this together from a few sources." That's not a style issue, it's a behavior giving itself away: someone who genuinely believes something doesn't stop to explain where the information came from. Only someone unsure of themselves, or someone who hasn't actually absorbed the material, does that.

Here's the trouble with a signal like the em dash — it can be imitated and dodged. Tell an AI to use fewer em dashes and the next sentence just uses a period instead, and it still reads like AI, because the real problem was never the punctuation. But "can't resist citing its sources" is hard to fake, because it's determined by how the content was put together: did you actually chew the information up into your own judgment, or did you just lay a few pieces side by side and summarize them? Whether it's been chewed up or not is impossible to hide.

And this isn't only an AI habit. I've met people who write the same way — always saying "some article said" or "some influencer mentioned" — and it's the same thing underneath: they haven't actually worked out what they think, so they lean on "someone else said so" to hold the sentence up. The only difference is an AI writes that way because it's literally assembling material, while a person does it because they're too lazy to take the one extra step of turning a borrowed idea into a genuinely own one.

After that night I changed a habit. Once the AI finishes a draft, the first thing I do isn't check for typos — it's hunt for words like "combining," "referencing," "these few articles/cases," and delete every one I find, then restate the conclusion in my own voice. Next time you read something and think "the logic checks out, but something feels off," try a different test: don't count the em dashes. Watch for whether it keeps trying to tell you "here's where I pieced this together from" — that check is more reliable than counting punctuation.