How AI detectors work, why they flag human writing, and how to sound like yourself
AI detectors estimate whether text was written by a person or a model by looking for statistical patterns, mainly how predictable the words are and how even the sentences are. They give a probability, and that probability is often wrong, especially for people who write in a second language. Watermark checkers are different: they look for a hidden signal that some AI models add while they write, and they can only find their own.
I first wrote this post in July 2025. Since then the evidence on accuracy has grown, and Google opened its watermark checker to other AI companies. This update covers how each kind of detector works, how much to trust it, and how to make writing you drafted with AI read like you.
How pattern-based detectors work
Most detectors you can paste text into are classifiers. They were trained on large sets of human and AI text and learned the features that separate them.
Predictability is the first feature. A language model picks likely next words, so its text has low perplexity (a measure of how surprised a model is by each word). People make odder choices: a strange word, a local phrase, a joke that half lands.
Rhythm is the second. People mix long sentences with short ones, which researchers call burstiness. Model text tends to keep every sentence at a similar length, so a page of it looks even and reads flat.
Vocabulary is the third. Models learned from a lot of formal and marketing writing, and they reach for the same words ("delve", "seamless", "robust") far more often than people do. A detector notices when a text is full of them.
How watermarks work
A watermark is a pattern a model adds while it writes, invisible to a reader. Google's SynthID Text nudges word choices in a way a checker can measure later, and Google published it as open source in 2024.
On 21 Oct 2026 Google made its SynthID Detector available worldwide, with partners including OpenAI, NVIDIA and Kakao, so one check covers content from all of them. Google itself says it is "not a general AI detector": it finds only watermarks from models that use SynthID, and heavy editing or translation weakens the signal.
How accurate AI detectors are
Less accurate than their scores suggest. OpenAI released its own AI text classifier in January 2023 and withdrew it on 20 Jul 2023 because of its low accuracy. In its own tests it caught 26% of AI-written text and wrongly flagged 9% of human text.
The people most often flagged by mistake are people who write in a second language. A Stanford study found that seven detectors labelled more than half of the essays written by non-native English speakers as AI, while they judged essays by native speakers almost perfectly. Plain vocabulary and regular sentences, which many of us learned in language class, look "predictable" to a detector.
Mixed text is the hardest case. Research from the University of Maryland shows that light AI polishing of human writing gives widely different scores from one detector to the next. Most writing today sits somewhere between human and AI, and a single percentage hides that.
If you review other people's writing
Treat a detector score as a reason to talk, and keep it out of the evidence. Ask the writer how they worked, look at their drafts and version history, and compare the piece with their earlier writing. Set a clear policy on AI use up front, so people know what is allowed before they write.
How to make AI-assisted writing sound like you
If your school, publisher or client bans AI, follow that rule. Where AI help is allowed, these habits make the final text yours, and they also make it better to read:
- Write the outline and the main argument yourself, and use AI for a first draft of the parts you know well.
- Add what only you know: a result from your own work, a customer's question, a mistake you made.
- Replace general claims with specific ones, such as a number with its source or a named example.
- Read it out loud and vary the rhythm. Split long sentences, join choppy ones and cut the filler.
- Delete the words you would never say to a friend, the "delve" and "seamless" kind.
- Use the active voice and plain words.
- Keep your notes and drafts, so you can show how the piece came together if anyone asks.
Questions people ask
Can AI detectors be trusted?
They can point a reviewer to text worth a closer look, and that is all. Their error rate on human writing, especially by non-native speakers, is too high to use a score as proof.
Do AI detectors work on text edited by a person?
They struggle with it. Editing changes the predictability and rhythm they measure, and it weakens watermarks too, so mixed text gets inconsistent scores.
What should I do if a detector wrongly flags my writing?
Show your process: your outline, drafts, notes and the version history of your document. Ask which tool gave the score, and point to the published research on false positives.
The AI Detector category on AI Agents Live lists detection tools with their reviews, if you need to compare them for your team.
