Do AI detectors actually work?

What the testing shows

Independent evaluations — academic studies and journalist-run tests alike — consistently find the same shape: top commercial detectors catch a large majority of unedited AI text (often 80–95%+ depending on model and genre), miss more as text gets edited or as new models ship, and false-flag some percentage of human writing, with formal genres and non-native English writers hit hardest. OpenAI shut down its own public AI-text classifier in 2023 over low accuracy — a vendor admission worth remembering.

Accuracy also swings with length (short text is noise), genre (creative prose scores human; technical prose scores robotic) and time (every model release resets the arms race for a while).

The asymmetry that matters

A detector that's "95% accurate" sounds great until it processes a thousand honest essays: dozens of false accusations, each landing on a student who can't prove a negative. That's why Turnitin tells institutions scores should start conversations rather than verdicts, and why Grammarly built process-tracking — assertion is weak evidence in both directions.

If you're on the wrong end of a false flag, our false-positive guide covers what to do.

So how should you treat a score?

• As a strong signal, never a verdict — in both directions.
• Trust longer samples more than short ones; trust agreement between detectors more than any single one.
• If you're being evaluated: verify your own text first, and keep process evidence (drafts, history).
• If you're evaluating others: published error rates mean a score alone can't carry an accusation.

Where Humanize Studio sits in this

We build on the honest version of this picture: detectors work well enough that you should check your text, and unreliably enough that you should check it yourself rather than trust anyone's promise — including ours. Humanize, verify, iterate; no stored text, no forever-guarantees.