AI writing is everywhere, and so are the tools that promise to detect it. Teachers run essays through them, editors screen submissions, and hiring managers scan cover letters. The pitch is reassuring: paste text in, get a percentage, know the truth. But how accurate are AI detectors really? The honest answer is uncomfortable — not accurate enough to trust for any decision that matters. Here’s why, and what to do instead.
How AI detectors claim to work
Most detectors look for statistical fingerprints of machine-generated text. Two ideas come up constantly: perplexity (how predictable the word choices are) and burstiness (how much sentence length and rhythm vary). AI text tends to be smooth and predictable; human writing tends to be messier and more varied. Detectors are trained to flag the smooth, predictable pattern.
It’s a reasonable idea. The problem is that it produces two kinds of errors, and both are damaging.
The false positives: real writing flagged as AI
This is the serious one. Clear, well-structured human writing often looks “too clean” to a detector and gets flagged as machine-made. Non-native English speakers are hit hardest, because they tend to use simpler, more predictable sentence structures — exactly the pattern detectors associate with AI. Studies have repeatedly shown detectors disproportionately flagging non-native writers, and there are well-documented cases of detectors confidently labelling human-written classics and student essays as AI-generated.
When a false positive can cost someone a grade, a job, or their reputation, a tool that’s wrong even 5–10% of the time isn’t a minor inconvenience — it’s a fairness problem.
The false negatives: AI text that sails through
The other failure is just as real. Lightly editing AI output — changing a few words, rephrasing a sentence, running it through a paraphrasing tool — is often enough to drop a detector’s confidence to nothing. So the people deliberately trying to evade detection are precisely the ones most likely to succeed, while honest writers get caught in the crossfire. That’s the worst possible combination.
OpenAI made this quietly official when it retired its own AI-text classifier, citing a low rate of accuracy. When the company that builds the leading model can’t reliably detect that model’s output, that tells you something.
What the percentage score really means
A detector saying “87% AI” sounds precise and scientific. It isn’t. That number is a confidence estimate from an imperfect statistical model, not a measurement of a fact. It cannot prove how a piece of text was written — it can only say the text resembles patterns it associates with AI. Treating that estimate as evidence is where people get into trouble.
A saner approach
- Never use a detector score as sole proof. Treat it, at most, as a weak signal that prompts a conversation — never as a verdict.
- Look at the process, not just the product. Draft history, version control, and being able to talk through the work reveal far more than any detector.
- Design work that AI can’t shortcut. In-class writing, personal reflection, oral defence, and assignments tied to specific recent context are more robust than any detection arms race.
- If you’re the writer: keep your drafts and notes. A visible trail of your process is the best protection against a false accusation.
AI detectors aren’t useless, but they’re nowhere near reliable enough to carry the weight people put on them. The technology to definitively separate human from machine text at a glance doesn’t exist yet — and given how fast the models improve, it may never catch up. Until it does, the smart move is to lean on judgment, process, and honest conversation, and to treat that confident-looking percentage with a healthy dose of skepticism.
What to do if you’re falsely accused
Because AI detectors are unreliable, false accusations are a real and rising problem — especially for students and non-native English writers, whose clear, structured prose is exactly what detectors tend to misflag. If it happens to you, staying calm and methodical matters more than panicking.
First, show your process, not just your product. This is your strongest defence, and it’s why keeping a trail matters before any accusation. Draft history in your word processor, saved notes and outlines, version history in Google Docs, even browser research history — all of it demonstrates the work happening over time, which no detector score can override. Second, calmly explain how detectors work and their documented error rates. A detector output is a statistical guess, not proof, and reputable institutions increasingly recognise that a percentage score cannot establish how something was written.
Third, offer to demonstrate your knowledge. Being able to talk through your argument, explain your choices, or reproduce similar work in person reveals far more than any tool. An honest writer can almost always do this; it’s a fair and revealing test.
The broader lesson cuts both ways. If you’re a writer, protect yourself proactively by keeping your drafts and notes — a visible trail of your process is the best insurance against a false flag. If you’re an educator or editor, treat detector scores as, at most, a weak prompt for a conversation, never as evidence for a verdict. The technology to reliably separate human from machine text at a glance simply doesn’t exist yet, and acting as if it does causes real harm to real people.