Stop Accusing Writers of Using AI

Apparently, we’ve reached the point where using an em dash is evidence that you used AI.

An em dash. A punctuation mark.

So I guess Agatha Christie has some explaining to do, which might be difficult considering she died in 1976.

I’m joking, obviously. But only kind of.

There is a growing obsession online with trying to figure out whether authors are using AI, and I understand where it came from. There are legitimate concerns about AI-generated books, undisclosed AI use, copyright, originality, and people pumping out low-effort books as quickly as possible.

I’m not here to defend any of that.

We’ve all seen examples floating around online of published books containing passages that look suspiciously like raw chatbot output. If Chapter 12 of your novel suddenly says something along the lines of, “Certainly! Here is the revised chapter in a more emotional tone,” then yes, we have a problem.

If you’re going to outsource your novel to a robot, apparently proofreading is still part of the job description.

But there is a huge difference between finding actual evidence of AI use and deciding someone must have used AI because their writing sounds like ChatGPT.

And that’s where I think we’ve gone too far.

Apparently AI Owns the Em Dash Now

There are entire conversations online devoted to identifying supposed “AI tells.”

• Em dashes.
• Certain metaphors.
• Words like “delve.”

Even sentence structures like: It wasn’t anger. It was fear.

Apparently, the machines own all of those now too.

People have even pointed to passages from Agatha Christie as a pretty funny counterexample because her writing contains some of the same stylistic choices people now confidently identify as signs of AI. Again, she died in the 1970s.

Here’s the obvious problem: AI learned from human language.

Of course AI writing contains patterns that human beings also use.

The machine did not invent em dashes. It did not invent metaphors, dramatic sentence structures, clichés, parallel phrasing, or any of the thousands of other things writers have been doing for centuries.

We wrote like this first.

So if your logic is, “ChatGPT does this, therefore any writer who does this must be using ChatGPT,” you have the whole thing backwards.

AI Detectors Aren’t Magic Truth Machines

Then we get to AI detectors, where suddenly everyone on the internet becomes a forensic scientist.

Someone copies a paragraph into a detector. The detector says it is “82% likely AI.” Case closed. The author is guilty.

Except these detectors are not magic truth machines. Even companies that make them acknowledge that they can produce false positives.

Turnitin, for example, has acknowledged that its AI writing model can misidentify human-written text and says the score should not be used as the sole basis for taking action against someone. It also stopped displaying exact AI scores below 20% because false positives were more likely in that range.

OpenAI itself released an AI text classifier in 2023 and later discontinued it because of its low accuracy.

Yes, the company behind ChatGPT tried to create a tool that could reliably determine whether something was written by AI and decided the tool wasn’t reliable enough.

There is something inherently funny about using AI to determine whether AI wrote something. We ask the robot if the robot did it.

But that isn’t even the biggest problem.

These systems are making judgments based on patterns. They are not uncovering some secret digital fingerprint that reveals exactly how a manuscript was created.

Language is messy. Human beings repeat themselves. We write predictably, use clichés, and use weird metaphors.

Sometimes a human being will naturally write something that sounds like what we have collectively decided AI sounds like. That does not automatically make it AI.

Then There’s the Jerry Falade Story

This is where the issue stops being a funny argument about punctuation and starts having real consequences.

Jerry Falade is a graduate student at Southern Methodist University and the author of the crime novel Call Me, I’ll Hide the Body, which, side note, is a fantastic title.

The manuscript generated enormous interest in publishing. It reportedly went through a 14-way auction in the United States and sold for around $2.4 million.

That is huge for any author.

Then concerns surfaced about possible AI use. According to reports about the controversy, an editor raised questions during the submission process about possible AI involvement in the manuscript.

Falade’s agent said Falade initially provided an explanation the agent considered clear and credible. After another conversation, however, the agent said aspects of the explanation had changed, and the agent ultimately ended their relationship.

Falade has denied using AI to write the novel and has said his AI use was limited to research.

There is another important detail here. According to reporting about the situation, the editor’s concerns were not based on AI detection software.

So I do not want to turn this into, “An AI detector cost an author $2 million,” because that isn’t what happened.

But the story does illustrate the larger problem perfectly.

I don’t know whether Jerry Falade used AI to write his book. You don’t know either.

What we publicly have are allegations, questions about how the manuscript was created, an agent who said his confidence in the author’s explanation changed, and an author who denies using AI to write the novel.

Which means the responsible answer right now is: We don’t know.

For some reason, that has become an incredibly difficult sentence to say online.

We Have Gotten Very Bad at Saying “I Don’t Know”

People who strongly oppose AI may want suspicious writing to prove somebody cheated. Meanwhile, people defending an accused author may be equally tempted to assume the accusation must be completely false.

Neither response is particularly useful. Sometimes somebody probably did use AI. Sometimes somebody didn’t. And sometimes we may never know.

If we want to understand how a manuscript was created, there are much better things to examine than an AI detector score.

• Draft history
• Version history
• Track changes
• Research notes
• Editorial correspondence
• Previous versions of the manuscript

Those things can tell us something about the actual writing process.

An AI detector looks at the finished product and attempts to reverse engineer how it was made.

That is a very different thing.

Are Writers Supposed to Change Their Writing So They Don’t “Sound Like AI”?

This is the part I find especially troubling.

• Are authors now supposed to change their natural voices because they are afraid of sounding like ChatGPT?
• Stop using em dashes?
• Avoid particular words?
• Rewrite certain sentence structures?
• Make their prose deliberately less polished so an AI detector believes a human wrote it?

AI is designed to imitate human writing. Humans should not have to stop writing like humans because AI learned how to write like us.

And none of this means readers have to be okay with AI-generated fiction.

You are absolutely allowed to say, “I don’t want to buy a novel generated by AI.” You can want disclosure while wishing publishers establish standards around AI use. It’s normal to be angry if someone represents machine-generated work as their own creative writing.

Those are completely reasonable positions. But we should be able to oppose undisclosed AI-generated fiction and oppose baseless accusations against authors.

In fact, false accusations make it harder to address legitimate AI abuse. If every em dash becomes an AI scandal, eventually nobody is going to take the scandals seriously.

We Need Better Standards for Accusing Writers of Using AI

I think we need a few basic rules.

• Stop treating detector percentages as verdicts.
• Separate suspicion from accusation.
• Look for actual corroborating evidence.
• And get comfortable with uncertainty.

There is a massive difference between saying:

“This passage makes me wonder whether AI was involved.”

and:

“This author used AI, and here is the proof.”

If you are going to publicly accuse someone of lying about how they created their work, the evidence should be much stronger than a percentage generated by a website.

AI-generated content has created a real trust problem. Readers don’t want to feel fooled. Authors don’t want to compete against mass-produced junk. Publishers certainly don’t want to spend millions of dollars on books and later discover serious questions about how those books were created.

Those concerns are real.

But the answer to a trust problem cannot be more irresponsible accusations. We can’t complain that AI makes things up and then turn around and make things up about authors.

Sometimes a Writer Just Wrote It

So yes, call out AI slop when there is actual evidence.

If somebody accidentally leaves “Here’s your rewritten chapter with improved emotional depth” in Chapter 12, by all means, tell me immediately because that is wild.

But stop treating writing quirks, punctuation choices, sentence structures, and AI detector scores like you just ran DNA evidence.

And in the Jerry Falade case, I don’t think the responsible answer right now is either “He definitely used AI” or “He definitely didn’t.”

There is not enough publicly established evidence to confidently make either declaration.

Not every suspicion needs to become an accusation. Not every controversy needs an instant verdict. And not every sentence that sounds vaguely like ChatGPT was written by ChatGPT.

Sometimes, and I know this is going to be a shocker, a writer wrote it.

So I’m curious: What level of evidence would you need before you felt comfortable publicly accusing an author of using AI to write a book?

For me, accidentally leaving the chatbot instructions in the published chapter would certainly do it.