The Detector Trap

“Yes, let’s give this a shot and see.”

James and I sat in silence as we pasted his essay into one of the popular AI detectors, the kind that promises to tell you whether writing is human or machine. We hit submit and waited as a spinning wheel rendered judgment on his essay.

After hours working together, this was the first time we’d reached a draft that we both agreed could be a final one. James was a dedicated, hardworking international student with a very impressive track record. He’d come to the United States from Taiwan as a 7th grader, hardly speaking any English. Through months of late-night review sessions, he’d developed an admirable grasp over the language, and by the time I was working with him, was a volunteer tutor, led the school’s robotics team, and was building a robotic dog as a personal project.

As I opened my mouth to break the long silence, the result loaded.

Likely AI-generated.

“That’s impossible,” James said with a laugh tinged with frustration. We both knew for a fact that he hadn’t used AI, because I’d literally sat with him and watched him type every single word himself.

I didn’t know what to tell him then. I do now.

First, the machine was wrong

Before anything else, I should reiterate that James did not use AI. The tool was wrong.

As we quickly found out through a couple of searches, these false positives were a lot more common than we originally imagined.

OpenAI, the company that builds ChatGPT, made a detector of its own and quietly shut it down in 2023. It caught only about 25% of AI text, and falsely accused human writing of being AI-generated roughly 10% of the time.

Vanderbilt turned off the detector built into Turnitin that same year, after its math showed even a tiny error rate would falsely accuse many honest students.

One particularly interesting finding, discovered by Stanford researchers, is that these false positives were especially common for students whose first language wasn’t English. In the study, researchers ran TOEFL essays written by non-native speakers through AI writing detectors. The results were astonishing. For English native speakers, the detectors were accurate. However, 61% of the papers written by English language learners were falsely flagged as being AI-generated.

For non-native English speakers, like James, this causes a real issue. At best, AI is a remarkably powerful tool to help develop yourself as a writer. At worst, it can be a crutch that generates an entire essay in seconds. Both ends of the spectrum tempt non-native speakers, and everyone knows it.

So someone like James is simultaneously under more scrutiny because people know how valuable a tool AI is, but at the same time, he’s far more likely to be wrongly accused.

What the detector actually saw

The detector that read James’s essay found something that was technically proficient, but unimaginative and predictable. James was a disciplined student who preferred technical precision in his writing over flowery language or daring displays of a unique voice.

When the tool read his essay, it read an essay that could have been written by dozens of other students, or, more relevant to this topic, an AI agent.

Every sentence was correct. He had clean grammar, careful structure, and a measured rhetoric that he’d taught himself over years of learning the language. And that was the problem. In his effort to sound “right”, he’d lost something crucial, something that would differentiate him.

He’d lost his voice.

That is what these tools actually react to. Not AI, exactly. Flatness. A machine can’t tell careful second-language formality from AI polish, because on the page they look the same. Both are voiceless.

We knew James hadn’t used AI to write his essay, but the false positive still gave us something incredibly valuable.

The false positive taught us that somewhere along the line, we’d lost James’s voice, and that the draft wouldn’t work until we rediscovered it.

Why this traps bilingual students

This is the trap for almost every student writing in a second language, and it’s a cruel one. They are taught, correctly, to be correct. Fix the grammar, choose the formal word, don’t make mistakes. Over years, correctness quietly masks the student’s unique perspective.

So when the stakes are highest, such as during a college application, they reach for the safest, most formal English they can produce, or they let AI smooth it for them. And in doing that, they erase the very thing that will make them stand out.

What actually fixed it

In our redrafting process, we leaned heavily on James’s own experience of what happened in his story. We went back through the essay, and wherever I found vague generalities like “we overcame the obstacle”, I asked the same question. What did you actually see here? What did that night feel like? What, specifically, happened, and what was your inner voice telling you at the time?

In short, we put James back in. The specific detail, the honest thought, the robotic dog he built alone at his desk, his excitement when it took its first step, coupled with his disappointment when it tumbled over and flailed on the floor. We did everything we could to bring his perspective into the story.

We put the new draft in, and after a similarly loaded pause, James saw the words:

Unlikely AI-generated.

He sighed with relief and pride.

What to do with this

If your child is second-guessing their writing because a detector, a teacher, or their own nerves called it into question, don’t reach for another tool. Reach for one question.

Ask them: what’s in this essay that only you could have written? A specific memory, a real opinion, a detail from your own kitchen table. If it’s there, the piece is alive, and no score should make them change it. If it isn’t there yet, that’s the actual work, and it has nothing to do with detection.

The goal was never to pass as human, or to sound correct. It was to write something so unmistakably yours that the question never comes up.