Short answer, because it’s probably why you’re here: you usually can’t know for certain whether a piece of text was written by AI, and the tools that claim they can are wrong often enough to ruin someone’s day. A detector that says “92% AI” is not proof. It’s a guess from a machine that flags real human writing as fake all the time — and flags it hardest when the writer learned English as a second language.
That’s the uncomfortable truth underneath one of the most-Googled questions of 2026. People type “is this written by AI” millions of times a month — students worried they’ll be accused, teachers staring at a submission that feels off, editors and hiring managers and ordinary readers who just want to know if a human was on the other end. Here’s the honest version of how to check, what the detectors actually do, and what to do if one of them is wrong about you.
(This is the text companion to our guide on how to tell if an image is AI-generated — same “is this real?” question, different medium. Text is harder.)
The 30-second check
You won’t get certainty. You can get a reasonable read. Here’s the fast version:
- Run it through one detector — and treat the number as a rumor, not a verdict. GPTZero or a similar tool gives you a percentage. That percentage tells you something only if it’s extreme and consistent. “97% AI” on a whole document is a flag worth investigating. “40% AI” means nothing.
- Read it out loud for the texture. AI writing tends toward smooth, evenly-weighted, slightly hollow prose — every paragraph the same length, every point politely balanced, no rough edges, no real opinion, no specific lived detail. Human writing has bumps. It runs long, then short. It has a favorite weird phrase.
- Look for the absence of specifics. This is the most reliable tell. AI is great at “restaurants offer a variety of delicious options” and bad at “the guy at the counter wouldn’t sell me the last bagel because he was saving it.” Vagueness where you’d expect a detail is the strongest signal there is.
- Stop there, and hold it loosely. You now have a hunch, not a fact. If the stakes are real — a grade, a job, an accusation — a hunch is where you start a conversation, not where you end one.
That’s it. Notice what’s not on the list: “trust the detector’s score.” Here’s why.
Why AI detectors get it wrong
The detectors aren’t useless. They’re just sold as something they’re not. They’re probabilistic classifiers — they estimate a likelihood based on patterns in text. They are not lie detectors, and their output is not calibrated to any standard of proof. Three things break them in the real world:
They collapse the moment text is edited. This is the big one. In controlled tests on raw, untouched AI output, the top tools look great. The second a human edits the AI text — or runs it through a “humanizer” — the accuracy falls off a cliff. One audit of 2,000 samples found Turnitin caught 98% of raw AI but only 12% of lightly humanized AI. Since almost nobody pastes raw output into a real document, the number that matters is the small one.
They falsely accuse real humans. Independent benchmarks put false-positive rates anywhere from a few percent on clean test essays to dramatically higher in actual classrooms. The University of Waterloo discontinued Turnitin’s AI detection after it flagged genuine human writing as 100% AI. Real students have lost sleep, faced dishonesty investigations, and been told to “redo” work they wrote themselves — on the say-so of a tool with no appeal process. As one student put it: the detector flagged “something I wrote using my own functioning brain.”
They’re biased against non-native English speakers. This is the part that should stop anyone from using a score as proof. Studies have found detectors flag a majority of genuine essays by non-native English writers as AI — one tool misclassified nearly all of a set of TOEFL essays. The “tell” these tools latch onto — clean, simple, structured sentences — is exactly what a careful second-language writer produces. Neurodivergent writers who write in plain, structured prose get caught the same way. The tool isn’t detecting AI. It’s detecting “doesn’t write like a casual native speaker,” and calling it cheating.
What this means for you
If you’re a student who got falsely flagged: the detector is not evidence, and you can say so. Keep your process. Version history in Google Docs or Word, your drafts, your notes, your search history while researching — that’s the real proof a human wrote it, and it’s far more credible than any percentage. Ask, calmly, what the accusation is actually based on. If it’s only a detector score, point to Waterloo, point to the non-native-speaker research, and ask for a human review of your draft history. (We have a full breakdown of why AI detectors falsely flag honest writers if you need to share the evidence.)
If you’re a teacher or editor: use detectors as a prompt to look closer, never as a verdict to act on. A flag means “read this carefully,” not “penalize this.” The defensible move is a conversation about the work — ask the writer to talk through their argument, or look at their draft history — not a number-driven accusation you can’t substantiate.
If you’re a hiring manager or reader: lower your confidence. You can often sense AI text by its hollowness, but you can’t prove it, and acting on a guess about a real person’s work is a fast way to be unfair. Judge the work on whether it’s good and true, which matters more than how it was produced.
If you’re a writer worried about being flagged: keep your drafts and your edit history as a matter of habit. It’s the one thing that actually settles the question. And don’t sand the personality out of your writing to “sound less like AI” — your specifics and your voice are the most human thing you have.
What this can’t fix
- No tool gives certainty. If anyone — a vendor, a professor, a manager — tells you a detector is definitive, they’re wrong, and it’s worth knowing the research that says so.
- The manual tells are weak too. Em-dashes, the word “delve,” the rule-of-three, “it’s not just X, it’s Y” — these were AI tells last year. Plenty of humans write that way, and AI is being tuned to stop. Don’t accuse someone over a punctuation mark.
- It’s an arms race. Every improvement in detection is met by better generation and humanizer tools. There is no stable, future-proof “AI test,” and there probably won’t be.
- Process is the only real answer. The durable way to know who wrote something is to look at how it was made — drafts, history, a conversation — not to scan the finished text. Build that habit and the detector question mostly goes away.
The bottom line
“Is this written by AI?” feels like it should have a clean yes/no answer, and it doesn’t. You can get a reasonable hunch in 30 seconds — run one detector and distrust it, read for hollowness, look for missing specifics — but a hunch is all it is. The detectors that promise more are wrong often enough to hurt real people, especially the most careful and the non-native writers among them. When the stakes are real, the answer isn’t a better scan. It’s the draft history, and a human conversation about the work.
Want to actually understand how these tools think — and how to use AI for writing without producing the flat, flag-bait prose everyone’s learning to spot? Start with AI Fundamentals for how language models really work, then Advanced Prompts and Content Creation to write with AI in a way that keeps your own voice front and center.
Sources
- Are AI Detectors Accurate in 2026? Reliability, False Positives, and Real Tests — Walter Writes
- How Do Professors Detect AI in 2026? Tools, Accuracy, and False Positives — Thesify
- The False Positive Epidemic: The Evidence Against AI Writing Detectors — Arab World Books
- AI detecting AI in academic writing: Why most AI detector findings are false — ScienceDirect
- False Positives in AI Detection: Complete Guide 2026 — Proofademic
