Are AI detectors accurate? What a score can and cannot tell you
Short answer: AI detectors are useful signals and unreliable verdicts. They estimate how much a text looks like the output of a language model, and every detector measured so far also marks some human writing as AI — most often formal, polished prose and the writing of non-native English speakers. A score should start a conversation, never end one. If your own work is flagged, the strongest evidence you have is the history of how you wrote it.
How AI detectors work
Most detectors are classifiers trained on examples of human and machine-written text. They learn statistical habits of language models: predictable word choice, very even sentence rhythm, and stock phrasing. Some add zero-shot measures such as how surprising each word is to a language model. None of them sees how a text was written — only the finished words.
Why they flag human writing
- Formal academic prose is predictable by design: standard section structure, defined terms, cautious phrasing. That is exactly what a detector learns to associate with models.
- Non-native English writers tend to use a smaller, more conventional vocabulary. A 2023 study by Liang and colleagues, published in *Patterns*, found that several popular detectors misclassified essays by non-native writers as AI-generated far more often than essays by native speakers.
- Short texts give too little evidence: a paragraph can swing from "human" to "AI" with a few edits.
- Edited text: a human draft polished by a grammar tool, or a model draft rewritten by a person, sits between the classes the detector was trained on.
Even model makers have stepped back: OpenAI withdrew its own AI Text Classifier in July 2023, citing its low rate of accuracy.
How to read a detector's score
- Treat it as a probability-like signal, not a measurement of how much of the text a model wrote.
- Look at where it points — a good detector shows which sentences drove the score, so you can see whether it is reacting to boilerplate.
- Prefer longer passages: a whole section is more reliable than one paragraph.
- Compare with what you know about how the text was produced — drafts, notes and version history outrank any score.
The AI content detector is built around that: an explainable score, a per-sentence view of what drove it, and a separate style read so you can see which wording reads formulaic.
If your own work is flagged
- Keep your drafting history: version history in Word or Google Docs, earlier drafts, notes, the sources you read. It is the most convincing evidence of authorship.
- Ask what tool and threshold were used, and whether the score was checked by a person.
- Offer to talk through the work: explaining your argument and sources in your own words answers the real question directly.
- Know your institution's policy on AI assistance — many allow grammar help or brainstorming with disclosure.
Using AI honestly in research writing
Most journals and universities now publish AI-use policies. A common pattern: AI cannot be an author, the human authors are responsible for every claim and citation, and substantial AI assistance is disclosed. When AI helped you draft, rewrite the text in your own voice, check every fact and reference against the source, and say what the tool did.