People say they can tell when something was written by AI, and they are usually right. Ask them how, and the answer is vague: "it just reads like it." That vagueness is worth taking seriously, because it means the signal is not in the words. If it were, they could point at one.
We score every draft our tool produces against a set of these tells before showing it to anyone. Here is what is actually in that list, and how to remove each one from writing you did yourself.
1. Every sentence is the same length
This is the strongest tell and the least discussed.
Human writing has uneven rhythm. A long sentence that develops an idea, then a short one. A fragment, sometimes. The unevenness is not decoration. It comes from thinking, where some thoughts take longer than others.
Model output tends toward a consistent middle length. Every sentence lands in the same twelve-to-eighteen word band, and the effect is a kind of hum. Readers do not consciously notice sentence length. They notice that nothing surprises them.
Fix: read your draft aloud and mark where you naturally pause. If the pauses are evenly spaced, break one sentence in half and merge two others. You are editing for rhythm, not for meaning.
2. The three-part list, everywhere
Three examples. Three reasons. Three adjectives. It is a genuinely good rhetorical structure, which is precisely the problem. A model reaches for it constantly because it is reliably fine.
Real writing has lists of two, and of five, and of one thing mentioned twice because it matters. When every list in a piece has three items, the pattern becomes visible and the writing starts to feel assembled.
Fix: count your lists. If they are all threes, cut one to two items. Usually the third was padding anyway.
3. The reversal
It's not about the tools. It's about the process.
Once you see this construction you cannot stop seeing it. Not X, but Y. It performs insight without carrying any: it sets up a position nobody held so it can knock it down.
Occasionally it earns its place, when X is genuinely what people believe. Most of the time X is a straw man invented one clause earlier.
Fix: for each reversal, ask whether a real person actually believes X. If not, delete the first half and keep the second. The sentence gets shorter and stops posturing.
4. Confidence with nothing underneath
Model output is fluent about things it has no information on. It will tell you a habit takes a few weeks to form, that most startups fail for a specific reason, that the best writers do a particular thing, all delivered in the same even register as a fact someone actually checked.
Human writing carries hedges in the right places, and specifics in the others. A person writes "I tried this for about two months and it stopped working around week six," because that is what happened to them. A model writes "consistency is key," because that is what the sentence shape wanted next.
Fix: find every general claim and ask where it came from. If the answer is "it sounds true," replace it with something that happened to you, including the numbers. Specificity is the single fastest way to stop sounding synthetic, and it is unfakeable, because only you have your details.
5. The closing question nobody asked
Ending on "What's your take?" or "Have you tried this?" is a habit models learn from a corpus full of engagement-bait. It reads as a request rather than an ending, and on a platform where everyone does it, it reads as a request everyone is ignoring.
Fix: end on your strongest sentence. If the piece genuinely invites a reply, the reply will happen without being solicited.
6. Transitions that announce themselves
Moreover. Furthermore. In conclusion. It's worth noting that. These are essay scaffolding. In a post of a few hundred words, they are half a sentence of pure overhead, and they signal that something has been composed rather than said.
Fix: delete them and read the result. The join almost always survives, because the logical relationship was already clear from the content.
7. Nothing at stake
The subtlest one. Model output rarely takes a position that could cost anything. It presents balanced views, acknowledges that reasonable people differ, and lands somewhere unobjectionable.
Writing people remember has a claim in it that someone could disagree with. The absence of disagreement-potential is itself the tell, even when every individual sentence is fine.
Fix: ask what a reasonable person could argue with in your draft. If the honest answer is nothing, you have written a summary, not a post.
Why we score rather than just prompt
You can put every rule above into a prompt. We did. It helps, and it is not enough. An instruction is a request, and a model under several competing requests satisfies some of them.
So each draft gets checked after generation rather than only guided before it: hook strength, readability, length against what you actually write, and density of the tells above. The score is the part that catches the drafts where the prompt did not take.
That distinction is worth stealing whether or not you use our tool. Prompting is a bet on the output. Checking is a measurement of it.
The short version
If you only remember one thing: the tells are structural, not lexical. Swapping "delve" for "explore" changes nothing. Breaking your rhythm, cutting a list from three to two, deleting a reversal, and replacing one general claim with a specific number will do more than any amount of word substitution.
Everything above applies to writing you did entirely by hand, which is the point. These are not AI problems. They are the habits that make writing forgettable, and a model just happens to have all of them at once.