People can tell. What they are picking up on is a short list of specific, repeatable patterns, easy to name once you have seen them.
Which is good news, because a thing you can name is a thing you can fix.
Why it reads that way
After a model learns to continue text, it gets a second round of training on human ratings: people judge its answers, and it is pushed towards whatever they marked well. What gets marked well is the safe middle of every choice. Make that choice a few hundred times in a row and you get prose that is relentlessly, structurally average.
Average is not the same as bad. The grammar is clean, the structure is sound, the transitions work. That is exactly the problem. There is nothing wrong with it and nothing in it, and readers register that as a texture long before they consciously identify why.
Every tell below is a symptom of the same cause.
The tells
The em dash. The most-cited tell of the lot, and the one people spot first. Models reach for it constantly, because it is the punctuation mark that works almost anywhere and commits to nothing: it can stand in for a comma, a colon, a full stop, or a pair of brackets, so the safest choice is always available. Human writers pick the one that fits and vary as they go.
Nobody has established why the habit is this strong. The explanations on offer are guesses, and so is that one.
Triples everywhere. "Clear, concise, and compelling." "Faster, cheaper, and more reliable." Three-item lists are pleasant and slightly over-represented in edited prose, so they come out constantly. Real writing varies: sometimes two, sometimes five, sometimes one blunt word.
The not-just-but construction. "It's not just a tool, it's a partner." "Not merely faster, but fundamentally different." Once you notice this one you will see it in every third paragraph.
Throat-clearing openings. "In today's rapidly evolving landscape." "As technology continues to advance." Paragraphs that spend a sentence warming up before saying anything.
Signposting that carries no information. "It's important to note that." "It's worth mentioning." "Let's dive in." Delete these and nothing is lost, which is the test.
The closing restatement. A final paragraph that summarizes what you just read, adding nothing. Human writers usually end on the strongest point, or a turn, or simply stop.
Suspicious symmetry. Five sections, each with three bullets, each bullet roughly the same length. Real emphasis is uneven: the important thing gets four paragraphs and the minor one gets a clause.
Inflated adjectives. Robust. Seamless. Powerful. Comprehensive. Leverage. Delve. Words that sound like they are doing work while asserting nothing checkable. "Delve" is the one that got famous, after researchers counted how sharply it climbed in academic abstracts through 2024.
No specifics. This is the big one, and it is the one that actually matters. No numbers, no names, no dates, nothing that happened to a particular person on a particular day. Generated prose stays at a level of abstraction where nothing can be wrong, and nothing can be interesting either.
The fix is not editing the tells out
You can strike every "delve" and still have prose that reads as machine-made, because the tells are symptoms. The cause is that you asked it to write, so it wrote the average thing.
The move is to stop asking it to write, and start using it where it is genuinely strong. In practice that is three jobs.
1. Editor, not ghostwriter
This is where it earns its place, and it is badly underused. Write the draft badly yourself, then:
Cut this by 30% without losing any point I'm making. Show me what you cut, as a list, so I can put anything back.
Which paragraph here is weakest, and why? Just that one paragraph.
I'm trying to say [X]. Where does this drift off that?
Find every sentence that could be cut without losing meaning.
Notice these all return judgements about your text. The words stay yours, and you keep the decision.
2. Option generator
The reason "give me three options" produces three near-identical rewordings is that all three are aiming at the same safe middle. Fix it by naming the axis you want them to differ on:
Give me three openings for this piece:
- Starting with a concrete example
- Starting with the counterintuitive claim
- Starting with the reader's problem
Same argument, genuinely different entry points. One paragraph each.
Now they have to differ, because you have made them differ structurally rather than asking nicely for variety. The same works for headlines by angle, for explanations by audience, for arguments by which objection they pre-empt.
3. Structure
Handing it a mess and asking for an outline is legitimately useful, and none of its prose survives into your draft:
Here are my notes. What's the strongest order to make this argument in, and what's missing?
Getting your voice into it
When you do want it generating text, the fix is the same as everywhere else on this site: give it the actual material.
Paste your own writing as the sample. Two or three paragraphs you are happy with beats any amount of describing your style. "Match the voice in the sample: sentence length, how formal, how much hedging."
Describe your reader, specifically. Not "a general audience", but "an operations manager who has been burned by a bad software rollout before."
Give it your specifics. The number, the name, the thing that went wrong on Tuesday. It cannot invent your details, and details are what separate writing from content.
What readers are actually detecting
The tells are real and worth knowing. But a piece that has had every "delve" removed and still says nothing checkable will read as machine-written anyway, because the thing readers are detecting is absence, not vocabulary.
Related: How to Write Better Prompts for the shape of the ask, and Context Is Everything for why pasting your own writing works better than describing it.