AI Detection in 2026 – What Has Actually Changed
three years in. what’s actually different.
what improved: detection accuracy on raw, non-humanized AI text is genuinely better in 2026 than 2023. most major tools now catch 80-90% of unprocessed ChatGPT output. that’s real progress.
what didn’t improve: false positive rates on vulnerable populations – ESL students, French-language writers, neurodivergent writers – are still unacceptable. humanized AI text is still largely undetectable. the inter-tool variance (same text, 40-point score difference between tools) hasn’t narrowed significantly.
what got worse: the humanizer ecosystem. in 2023 humanizers were rough and detectable. now they’re sophisticated enough that major detection tools catch roughly 20-30% of humanized output. the arms race is not going well for detection.
net assessment: better, but not good enough for formal academic integrity use. probably not getting there given the humanizer dynamic. policy should reflect this.
5 Replies
Join the discussion.
Log In to Reply
the adaptation race is real. I've been tracking GPTZero and Turnitin updates against what students are doing and the lag is consistently 2-4 months. by the time a detection update ships, the evasion technique it targets has already evolved. the technical arms race is not winnable from the detection side.
Chris's point about confidence interval transparency is the technical thing that has to change for this field to mature. a score without a confidence interval is not a measurement - it's a black box output. the tools have improved at detection but none of them have improved at communicating their own uncertainty, which is actually the more important capability for classroom use.
what's actually changed is teacher practice more than the tools. the teachers who have moved to process-based assessment aren't waiting on detection to improve - they've redesigned around it. that's the real evolution and the tools are becoming irrelevant to that group regardless of how they improve.
the raw detection improvement is real and worth acknowledging. I've seen it in Turnitin. but your point about humanizers is the dynamic that limits how far that improvement can take us.