Can readers tell AI wrote it, and what one accusation costs
The question arrives as one line in a reply
You send the newsletter on Tuesday. On Wednesday a subscriber replies with four words.
Did AI write this?
The reply sits in your inbox all afternoon. You wrote the piece from a case you handled in March. A model helped you shape it. Now a stranger is asking a question that feels like an accusation. Every answer you draft sounds defensive.
Answer it well and the reader learns something about how you work. Answer it badly and they file you with the rest of their inbox. Nothing you send after that gets read closely.
Two audiences sit behind this worry. What search engines do with AI-assisted pages has an answer of its own. What readers see is the question here.
Readers identify AI writing at about chance
In August 2026, two researchers at Villanova published the cleanest test of this question so far. Sydney Sears and Deena Skolnick Weisberg ran three experiments with adults aged 18 to 81. Their work is published in the journal Judgment and Decision Making. They paired short stories from literary journals with ChatGPT versions of the same material.
Readers picked the AI story correctly 40% of the time in one experiment and 52% in another. One result is below chance. The other is chance.
Then the finding that should reframe the whole worry.
Readers rated the AI stories higher, and rated them higher still when told a human had written them. The label moved the scores. The writing did not.
Two things follow for you. Your reader probably cannot tell. And the moment they believe a machine was involved, they mark you down for the belief.
That is worth sitting with, because it means the fight is not about prose quality.
What readers notice is the absence of a person
Something does make a piece feel machine-made. It is rarely the sentences.
They notice that nobody is named. They notice that nothing in the piece cost the writer anything. They notice that the same article could have run under any firm's name, in any city, in any year.
A reader cannot articulate that. They feel it, reach for the nearest explanation, and the nearest explanation in 2026 is AI.
The Villanova team found one useful detail on top. Readers who use these tools often did spot the machine versions more reliably, while literary expertise predicted nothing. The tells they cited are stylistic habits. One is the em dash used as a tic. Another is the "it's not just X, it's Y" construction that models reach for.
A house rule against both is cheap insurance. Style is the small half.
The large half is material. Four things do most of the work.
- A date.
- A figure you watched happen.
- A decision you regret.
- A sentence your competitor would argue with.
A page carrying those does not read as machine-made. A model cannot supply any of the four. The problem is rarely that a model helped you write. The problem is that nothing of yours went in.
Detectors are not evidence
The tools that claim to settle this do not settle it.
In 2025 the Federal Trade Commission ordered a detection company to stop advertising its accuracy. The company sold its detector as 98% accurate. Independent testing put it at 53% on general writing. The FTC's consumer protection director called that no better than a coin toss.
The problem is structural rather than one bad vendor. Detectors return a probability with no reasoning attached, so an accused writer has nothing to rebut.
Publishing has already lived through this. In 2026, a winning story in the Commonwealth Short Story Prize scored 100% AI on a widely used detector. Granta had published it. Both organizations stood by the story and noted that detectors are imperfect. The Authors Guild now offers legal help to members falsely accused, which tells you how the risk runs.
So when someone waves a detector score at you, the score is not the argument. Say so once, without heat, and move to what you can show.
What one accusation costs, and what to say
Here is the accounting, because the number in your head is probably too high.
One reader asking is worth one reply. It costs you 10 minutes. It usually ends with that reader trusting you more, since few people answer this question straight. The expensive version is public. A stranger posts a screenshot and a score, and you are proving a negative in front of an audience.
The reply that works names the method and the judgment in the same breath.
Say what your process is. You record your thinking, a model drafts from your material, and you check every fact. Then say what is yours: the case behind the piece, the numbers, the call you made. Then offer the evidence, because you have some. The recording exists. The matter exists. Nobody with a keyword tool and a prompt can produce either.
What not to do is a shorter list. Do not apologize for a process you are not ashamed of. Do not run your own work through a detector and quote the score back. Quoting it agrees that the coin toss decides. And do not promise you will stop using the tools, unless you mean it.
Announcing your method before anyone asks is a separate decision. There is a right way to put it on the page.
Where this answer runs out
Three limits, stated plainly.
If a piece came out of a model with nothing of yours in it, the reader is right. No reply repairs that. The fix sits upstream, in what you put in.
Some contexts make method a rule rather than a quality question. Academic journals, contests, and certain client contracts forbid AI-assisted work outright. Their rules apply whether the writing is good or not. You agree to them or you do not enter.
And you cannot fully control a public accusation. Granta stood by its writer, and the story still spent a week being argued about. What you can control is having material to show when it happens.
The reader is asking whether a person stands behind the page
Your reader wants to know whether a person with something at stake wrote this. The alternative is the internet's average opinion with your logo on top. That question has an answer, and the answer is in the material rather than the method.
Put your cases, your numbers, and your judgment on the page, and the question stops coming. Leave them out, and the page reads as machine-made whether or not a machine touched it.
