When AI Detectors Flag Your Own Writing as Not Human

When a growing chorus of readers and critics accused Jamir Nazir, winner of the prestigious Commonwealth Short Story Prize, of using AI to write his winning entry, I did what any writer with a healthy amount of professional dread would do — I read it.

Nazir’s work, “The Serpent in the Grove,” certainly bears many of the characteristics we’ve come to associate with AI writing. Bizarre metaphors and overly compressed language initially come across as bold, but a second read — and the strange language demands one — reveals the words for what they are: eye-catching tissue paper wrapped around an empty box.

The question of who wrote “The Serpent in the Grove” was one of several AI-generated writing controversies to emerge in May — a book reviewer for The New York Times and a Nobel Prize-winning novelist were at the centers of two others. Soon after the Commonwealth Foundation announced that Nazir’s story was a regional winner, a wave of response pieces followed, most arguing that AI writing was easy to spot and that anyone who relied on it would eventually face Nazir’s fate — even though the Commonwealth Foundation’s investigation concluded that he didn’t use AI, his name and any future work will always be attached to the accusations.

I mostly agreed that the story read like something AI would write. But I also noticed that the voices accusing Nazir tended to belong to writers, a population with an obvious stake in believing, and insisting, that an AI could never do what they did.

Those who want a more objective opinion point to AI detectors, tools that analyze text for patterns, structures, and other elements common in AI-generated writing. Pangram, the industry-leading detector, flagged Nazir’s text as 100% AI-generated. Curious to see how these tools worked — and how accurate they are — I entered an article I’d written months earlier into Pangram’s prompt box, pushing aside the irony of using AI to detect AI. Sure of my innocence, I clicked “check.” But as the tool did its thing, a thready nervousness crept into my chest. I knew I hadn’t used AI. What did I have to worry about?

A few milliseconds of unpleasant anticipation later, the screen loaded with a wash of green highlighter over my text. Phew. Then I read the score: 97% human-generated.

My anxiety curdled into indignation. I flung my cursor down the screen, furiously scrolling through the piece in search of that 3%, as if the software could sense my outrage. And there it was. My closing paragraph, highlighted in pale yellow, with a note attached: “Light AI Assistance.”

Rereading the paragraph in question, my steam began to cool off and settle into a pool of defeat and self-pity. I decidedly had not relied on AI assistance, but I could see why Pangram flagged the text: I’d written a list of three and couched it in em dashes — two common AI “tells.”

The thing is, I really liked that article, and I particularly liked how that ending came to be — born from the kind of human editorial process at the heart of most good writing. Originally, I’d submitted the piece with a different ending. My editor suggested I add a few sentences to tie back to open questions I posed in the introduction. It was a good idea, and a lightbulb flickered on as I began drafting options. She liked what I came up with, I liked what I came up with, and the main subject of the article, a scientist leading a new body of research, used it as a pull quote when he shared the piece. Readers commented and emailed to say they liked it, too. According to every standard I understood, I’d done good work.

Except now it read like something AI might write. Doubt crept into my mind. Was it really a good ending? Or was the point too tidy, too cliché, too, well, robotic?

A few months ago, I would have brushed aside the entire premise of using an AI detector as absurd, even comical — something I might try out of curiosity, but not read too much into. But by the time I was facing that yellow paragraph, I was mentally exhausted, stumbling off an edge I’d been riding all year, trying to reconcile AI’s existence with both my professional life and my identity as someone who writes. The rise of AI detectors was just another twist I couldn’t find footing on, another factor clouding my ability to read my situation, let alone anyone else’s.

My initial worry about AI centered on practical obsolescence: whether a client would literally replace me — a flesh-and-blood human being, one who grows tired and distracted — with a chatbot. Now that fear seemed almost quaint, a sign of a simpler time when AI only threatened my external world.

At some point over the past year, sitting down alone to work no longer meant sitting down alone to work. Even with the AI tools off, the tabs closed, I was, in some ways, considering the machine. That’s when I began to reach for the white flag. AI, I surrender: just get me out of here, out of this place where you’re all over my process and my voice — even when you’re not.

Will write (or chat with AI) for food

I didn’t always know I wanted to be a writer. My writing habit developed the way someone with a nutritional deficiency might start craving the foods that would correct it. It was a form of self-medication, a way to calm my restless mind, which constantly swerves between several trains of thought. Putting what I was thinking into words, usually bound in a journal, forced me to slow down and focus on one track at a time, only to discover they often intersected in surprising ways.

Professionally, I pursued ecology, drawn to the field because I was spending more and more time outside and wanted to understand the world around me. I followed that field to a Ph.D. because I found the subject to be intellectually interesting. Ecology is a messy science; we tease signals from noise, but we also accept a level of unexplained variance that experts in many other sciences would find anathema. Part of being an ecologist is testing for patterns while accepting that the irreducible complexity of the natural world may always leave unsolved questions. I turned out to be good at that kind of thinking.

My favorite parts of graduate school were the opportunities to reflect on my research — my process, my discoveries — and consolidate it into something I could cleanly communicate to others, usually through writing. When several postdocs didn’t pan out, I began editing for an academic journal and then writing science stories. Eventually, I stopped applying for academic positions altogether.

I’ve never had a byline in The New York Times, and every short story I’ve ever sent to a literary magazine has been rejected. Still, by the time ChatGPT, Claude, and the rest of the LLM chatbots arrived, I was making a decent living as a working writer. It’s true that for every piece that was meaningful — every sentence I was proud of — I had to accept several less glamorous gigs to keep the lights on. When one of those gigs ended last fall, I took it in stride. My portfolio was solid. Surely I’d find more work, maybe even a steady job.

I was familiar with AI when I started looking for work in late 2025. I’d already used ChatGPT for SEO research and had even edited AI-generated copy at a social media agency. I tolerated AI in these contexts, but I wasn’t exactly inspired to unleash the tool on my writing. At the same time, the handful of steady freelance gigs I had landed before LLMs went mainstream had insulated me from the extent to which the industry had begun to revolve around AI — not just as a tool more people were using, but as a force that would commodify writing, bringing opportunities and entire categories of work to a new and disorienting kind of low.