AI detector

Text detectors are usually classifiers trained on examples of human and AI writing, or methods that measure how predictable a text is to a language model, since generated text tends to be more predictable. Some image and audio tools look for generation artefacts, or for watermarks and provenance metadata embedded by the generator.

Why it matters: schools, publishers and employers use detectors to check submissions, so accuracy has real consequences. Research has found that detectors can be evaded by paraphrasing (Sadasivan et al., 2023) and can disproportionately flag writing by non-native English speakers as AI-generated (Liang et al., 2023). When choosing one, look for published, independent accuracy data, clear false-positive rates and guidance against using a score as sole evidence.

Example: a teacher runs an essay through a detector, which reports a high AI probability; a fair process treats that as a reason to talk with the student and review drafts, not as a finding of misconduct.

Limits: no detector is reliable enough to be the only basis for a decision, results differ between tools for the same text, and accuracy falls as AI models improve and as people edit generated text.