How accurate are AI text detectors?

Problem one: human writers in formal registers — academic English, legal writing, business prose — share many surface patterns with LLM output. Sentence-rhythm regularity, generic transitions, hedged claims. These patterns flag legitimate writing as AI.

Problem two: LLM text is easy to humanize. Light editing, paraphrasing tools, or simply running output through a second pass with different instructions can defeat detectors trained on the original outputs.

Vendor benchmarks frequently report 90%+ accuracy on long, unedited LLM samples. Independent evaluations on classroom-realistic essays — short, edited, mixed-author — typically see accuracy below 80% with non-trivial false-positive rates on real student work.

The OpenAI text classifier was retired in 2023 partly because of this gap. The problem has not become easier since.

Use it as a triage signal that prompts a conversation, not as evidence. A 'likely AI' score should lead to a discussion with the writer about their drafting process, not directly to a grade or HR decision. A 'likely human' score does not prove anything either.

If you're an educator, see the for-educators page for the workflow we recommend.

Long samples, unedited output, casual register, and tasks where LLM patterns are most visible (formulaic structures, list-heavy writing, generic openings). Short samples and creative writing are the worst cases for detection accuracy.