Roundup
The best AI detectors (and what they get wrong)
AI detectors are probabilistic classifiers with a documented false-positive problem. Here is what each one does and how much weight the number deserves.
Write It teamUpdated 3 min read
Every detector on this list is a statistical classifier making a guess. Some guess better than others. None of them know anything, and the confident percentage they return hides how much uncertainty is underneath it.
How they all work
Every detector measures roughly the same two things. Perplexity is how predictable each word is given what came before — machine-written text tends to pick likely words. Burstiness is how much sentence length and complexity vary — humans write unevenly, models write evenly.
Text that is plain, formulaic, or carefully edited toward simplicity looks statistically similar to machine-written text. That is the whole source of the false-positive problem, and it is a property of the method rather than a bug someone can fix.
1. Turnitin
Practically the most important one, because it is the one integrated into school submission systems. Its AI indicator reports an estimated percentage of the document that appears machine-written, alongside the similarity report it has always produced.
Students cannot usually run it themselves — it is licensed to institutions. What you see is whatever your teacher sees.
2. GPTZero
The most widely used detector students can access directly, with a free tier. It reports document-level and sentence-level assessments. Because it is free and popular, it is also the one most often cited in arguments about false positives.
3. Originality.ai
Built for publishers and content teams rather than schools, with per-sentence breakdowns and a plagiarism check alongside. Paid, and priced for professional use.
4. Write It AI Detector
Runs against your own draft inside the extension. The purpose is different from the others: it is not there to judge you, it is there to show you which paragraphs read as machine-written so you can rewrite those in your own words before anyone else sees them.
The false-positive problem
This is the part that deserves more attention than it gets. Detectors flag human writing, and not randomly — they flag some kinds of human writing more than others.
- Non-native English writers are flagged at measurably higher rates, because simpler vocabulary and more regular sentence structure look like low perplexity.
- Formal academic writing is flagged more than casual writing, because the register is more uniform by design.
- Heavily edited work can be flagged, because editing tends to smooth out exactly the variation detectors look for.
- Short passages produce the least reliable results, since there is not enough text to measure.
What a score actually means
It means a classifier assigned a probability. It is a reason for a conversation, not a verdict. Turnitin itself frames its indicator as something requiring human judgment rather than an accusation, and schools that treat the number as proof are misusing the tool.
See your draft the way a detector does
Install Write It and run the AI Detector on your own writing before anyone else does.