Getting an email that says your writing was flagged as AI is a special kind of gut punch. You wrote every word yourself, yet some detector decided a machine made it up. It’s happening to professional writers, students, and job applicants, and the worst part is how little recourse you have. Here’s what’s actually going on inside those tools, and what to do if one of them burns you.
Why writing gets flagged as AI in the first place
Start with this: AI detectors don’t read minds. They read patterns. A detector asks, in effect, how predictable a piece of text is for a language model. Predictable equals suspicious, because machines statistically favor the most likely next word. Human writing wanders more, breaks rules, and surprises the model, which is exactly why we tend to read as human.
Except when we don’t. Chris Callison-Burch, a computer science professor at the University of Pennsylvania, explained the trap to NPR: language models also find text from their own training data unusually predictable. So if your writing resembles the kind of prose these models were trained on, you can score as machine-made without a single AI keystroke involved.
The false positive hall of fame is instructive. Janelle Shane, a science blogger, ran excerpts from her own 2019 book about how AI works through a detector. It returned 98.7% confidence the passage was machine-generated. Her dry verdict: “a clear, glaring, blinking-light false positive.” Getting flagged as AI for a book published before the current tools existed tells you everything about baseline reliability.
There’s a structural problem too, and editors admit it. Clarkesworld, the science fiction magazine, got hit with such a flood of AI submissions that its portal temporarily shut down in 2023. Submissions roughly doubled from around 1,100 a month. Publisher Neil Clarke estimates his workload rose 30%. Under that pressure, screening tools look tempting, even though everyone knows they’re imperfect.
What detectors actually look for
Detectors weigh clues across the whole piece, not one word at a time. Still, certain habits raise your score. Ghostwriter Sally McGraw found out the hard way: after 20 years of professional writing, a client’s screening tools flagged her original work as machine-generated. She’d done nothing wrong. Her style just happened to sit where the detectors hunt.
Here’s what trips them, according to the NPR reporting:
| Detector trigger | Why it misfires |
|---|---|
| The word “delve” | Models overuse it; some humans genuinely like it |
| Frequent em dashes | Taught as “good style” in journalism school for decades |
| Ideas grouped in lists of three | Clean rhetoric reads as machine-tidy |
| Openers like “notably” or “crucially” | Adverb-heavy starts mirror model output |
| Question immediately answered | A classic AI response shape |
| Very short samples | Little text means noisy guesses |
Notice the irony. Every item on that list is also standard advice from writing coaches. McGraw’s quote about it stings: “I miss my em dash. It was something that I learned in journalism school and I’ve used it for decades.” Meanwhile, models learned to write by reading us, so of course our habits look machine-like to a machine.
Why detectors disagree with each other
If one flag is scary, try running the same text through three tools. Researchers have found that different AI detection tools can reach dramatically different conclusions about the same passage. In one test NPR described, 40 papers were checked with multiple detectors. Pangram, one of the better-known options, performed best: it correctly identified AI use in 37 of 40 papers, and it also caught AI text that a human had rewritten. Impressive. That still leaves a 1-in-13 error rate on clean text, which is cold comfort if you’re the one error.
Even Pangram’s CEO, Max Spero, admits his tool can make mistakes, particularly when there isn’t much writing to evaluate. Short samples, heavily edited text, and non-native English patterns all raise the error bars. The Authors Guild put it bluntly in May: AI detectors create “an environment reminiscent of witch hunts,” where a false accusation can cost an author a contract and a reputation.
If you get flagged as AI: 5 steps that actually help
Okay, worst case. You submitted something real and got told it’s machine-made. Don’t panic-fire back an angry reply. Work the sequence.
- 1. Ask which tool flagged you and the score. Different tools fail differently, and a percentage gives you something concrete to address. If the client won’t say, that tells you something too.
- 2. Assemble your evidence trail. Drafts, version history, Google Docs edit history, notes with timestamps. Docs’ version history is especially strong because it shows writing evolving over hours, which no detector accounts for.
- 3. Offer to discuss the work live. Anyone who wrote a piece can talk about it: why this structure, where the idea came from, what got cut. A ten-minute conversation proves authorship better than any percentage score.
- 4. Show the messy middle. Send an early draft next to the final. Machines produce clean output on the first pass. Humans leave scars, and scars are proof.
- 5. Keep it professional either way. Clarkesworld’s Clarke had the right framing: “Submissions should end in rejection, not an accusation.” If a client treats a detector score as a conviction, you learned something important about them. Take the loss, keep the receipts, walk.
One preventive note: McGraw says the accusation changed how she writes. She now avoids certain words, skips commas she’d normally include, deliberately breaks her own polish. That’s a rational response to a broken incentive, and honestly, a little healthy. We’ve already made the case that polished-to-death prose reads as fake, because it does. Writing with a bit more rough edge protects you and reads better anyway.
How to write without living in fear
You can’t game every detector, and trying turns you into a worse writer. What you can do is understand the game. Our list of phrases that make writing sound like a bot covers the vocabulary side, and if you want the technical counter-move, we also looked at an AI model trained specifically to unlearn machine-slop patterns, which shows how blurry the line between human and machine style has become.
Three habits that reduce risk without wrecking your voice:
- Vary your rhythm. Long sentence, short one. Fragments are fine. Machines smooth things out; you shouldn’t.
- Keep your process artifacts. Version history is your alibi, and it costs nothing to preserve.
- Save receipts for anything that matters. The freelance gig you really want deserves a drafts folder, not just a final file.
Also worth remembering: this cuts both ways. If you legitimately use AI to assist your writing and pass it off as fully human, you’re part of the reason innocent people get scrutinized. The human premium is real, buyers pay more for verified human work, and the market only stays honest if we don’t launder machine text through it.
Where this is heading
Detectors will get better. So will the writing models they’re trying to catch. That arms race has no finish line, which means false positives stay possible for the foreseeable future. Editors know it, writers know it, and the tools’ own CEOs say it out loud. The reasonable stance is to treat detector scores as one noisy signal, never a verdict. Writers flagged as AI without cause will keep paying the price until that stance becomes the industry default.
Until the tools mature, your protection is evidence, not style. Keep drafts, keep history, and offer to talk about your work like a person who actually wrote it.
Takeaway
If your writing was flagged as AI, ask for the tool and the score, pull your version history, and offer a live walkthrough of your drafts. Going forward, write with a little natural roughness and preserve your process, because the receipts will matter more than the em dashes you gave up.