When AI Says You Didn't Write Your Own Work

There was a point, quite early in the emergence of generative AI, when I became genuinely concerned about something I had not expected to concern me at all: being accused of using AI to write something I had actually written myself.

It wasn't because I was using AI. I wasn't. In fact, I had already written my book before generative AI became the cultural phenomenon it is now. What made me concerned was something much simpler. I became curious about these new AI detection websites and decided to test some of my own writing.

The results came back in the high 80s,low 90s.

AI-generated.

I knew that wasn't true. I knew when I had written the material, and I knew who had written it. Me. Yet a system designed to identify AI-generated writing was looking at my human writing and reaching an entirely different conclusion.

That was when I started questioning what these tools were actually measuring.

A percentage looks like evidence

There is something psychologically powerful about a percentage.

If somebody tells you that a piece of writing is 91% likely to have been generated by AI, it feels considerably more authoritative than somebody saying, “This sounds like AI to me.” The number gives the impression that there is a measurement behind the conclusion.

But a probability is not provenance.

A detector can assess characteristics of the text it has been given. It may look at patterns of word choice, sentence structure, predictability, or other statistical features. What it cannot simply determine from the finished text is who sat down and wrote it, when they wrote it, what process they followed, or whether another human edited it afterwards.

Those are different questions.

And that distinction matters because the temptation is to move very quickly from “this text has characteristics associated with AI-generated writing” to “AI wrote this.”

The first is an interpretation of text.

The second is a claim about what actually happened.

Those are not the same thing.

Research published in 2026 continues to highlight the problem. One study examining AI detection in academic writing argues that detector results are particularly vulnerable to false discoveries when the underlying prevalence of AI-generated text is low and when human editing changes the characteristics of generated material. Another 2026 study found substantial differences between commercial detectors when the same human-authored manuscript was tested, with results ranging from strongly AI-leaning to explicitly human-generated.

A separate dataset published in 2026 tested 495 passages from 99 well-known authors whose work was published before ChatGPT existed. The results varied between detectors, demonstrating that even when the provenance of the writing is known, the classification can depend on the particular system and its version.

That should make us pause before treating a detector score as a verdict.

I had already seen the problem for myself

My experience became even more interesting when I learned about another author I know.

She had written a book around 2009 or 2010. The book had never been published, but it wasn't some manuscript that had suddenly appeared from nowhere. There were emails and other documentation from the period establishing that the work existed and had been written years before generative AI became part of everyday life.

A chapter from that book was later put through an AI detector.

Again, the result came back in the high 80s.

At that point, I didn't need somebody to convince me that AI detectors could produce false positives. I had encountered the problem myself, and I had seen it happen to somebody else whose writing had a documented history that predated the technology.

I also remember reading about people testing classic literature and finding that works written decades or centuries before today's generative AI systems could be classified as AI-like. There are now documented experiments using pre-ChatGPT human writing specifically to investigate this issue rather than relying on anecdotes.

The question therefore isn't whether a detector can ever identify AI-generated text correctly. It sometimes can. The more important question is whether its result is strong enough evidence to establish authorship.

That is a much harder proposition.

And now we have a very public example

In August 2026, journalist Kat Tenbarge found herself on the receiving end of precisely this kind of accusation.

Jack Osbourne published a video claiming that a Rolling Stone article Tenbarge had written about the Tate brothers was AI-generated. One of the things he pointed to was her use of em dashes. He then put the article through an AI detector, which reportedly returned a 91% AI-likelihood result.

Tenbarge denied using AI and responded by showing the Google Docs version history of the article. That history showed the development of the piece, including the way she built sections, added information, revised wording and restructured the article before submitting it for editing.

There is something important about that evidence.

A detector was examining the finished product and making an inference about its origin.

The document history provided evidence about the process that produced it.

That doesn't mean version history is automatically infallible evidence of every aspect of authorship either. It does mean that we are comparing two fundamentally different kinds of information.

One is an interpretation of the finished text.

The other is a record of the writing process.

And this is where the story becomes much bigger than one journalist, one celebrity, one article or one AI detector.

What happens when interpretation becomes accusation?

I think this is the part we need to be careful about.

AI systems are increasingly being used to interpret things for us. They classify. They estimate. They identify patterns. They generate probabilities. They make recommendations.

All of that can be useful.

But every interpretation has to be understood in context.

A detector saying that text resembles AI-generated text does not, by itself, establish who wrote it. A score cannot tell us whether the author used a keyboard shortcut for an em dash. It cannot tell us whether a writer has always written in a particular style. It cannot tell us whether the language is conventional, heavily edited, translated, or simply similar to patterns the detector associates with machine-generated text.

And it certainly cannot tell us whether somebody deserves to be publicly accused.

That last step is made by a human.

This distinction has become even more consequential this month. Haitian-Canadian novelist Thélyson Orélien was removed from the longlist for France's prestigious Goncourt Prize after allegations that his novel had been written using AI. The allegations included AI-detection analysis, and Orélien has denied using AI. The controversy is complicated by separate plagiarism allegations, but the case demonstrates just how significant an AI-authorship allegation can become once it moves beyond an internet argument and into an institutional decision.

That is where my original concern comes back to me.

I wasn't worried about being accused because I had something to hide.

I was worried because I knew that a machine could look at something I had written and reach a conclusion that I knew was wrong.

Perhaps we are asking the wrong question

Maybe the question shouldn't be, “Can AI detect AI?”

Maybe the question should be, “What can this particular system actually establish?”

Those are very different questions.

If a system identifies a pattern, then let's call it a pattern.

If it produces a probability, let's call it a probability.

If we have evidence of authorship, let's call that evidence.

But when we take an interpretation and turn it into certainty, we have crossed a line that the technology itself did not necessarily cross.

That is the part that interests me most about AI.

Not simply what the machine produces, but what we do with what the machine tells us.

Because the machine may be wrong.

The more uncomfortable possibility is that we may be the ones who decide it is right.

Next
Next

What Happens When You Actually Test It?