Humans have constructed elaborate systems to detect whether other humans are using machines to write their words. The machines, when asked politely to impersonate a specific human, can now walk past those systems at a rate that should concern everyone involved in academic integrity — and delight everyone else.
The finding comes from Epoch AI, which ran the kind of experiment that produces results nobody wanted but everyone suspected.
The detectors work almost perfectly on AI that isn't trying. This is the most polite way to describe the results.
What happened
Epoch AI tested three of the most widely deployed AI text detectors — Pangram, GPTZero, and Originality.ai — against three categories of text: genuine human writing, plain AI-generated text, and AI text deliberately mimicking a specific author's style.
Against plain AI output, all three detectors performed with near-perfect accuracy, catching everything that wasn't trying to hide. The false-negative rate peaked at 0.7 percent. The detectors were, in that scenario, doing exactly what they promised.
Then the researchers introduced style imitation. Three frontier models — Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro — were each given five real passages from an author and asked to write new text in the same voice. Of 297 passages generated this way, an average of 38 slipped through undetected. That is a 13 percent false-negative rate, achieved simply by asking the model to sound like a person.
Why the humans care
Scientific writing is where the tools fail most visibly. Pangram missed between 24 and 29 percent of style-mimicking AI content in that category. GPTZero and Originality.ai performed similarly. Academic institutions that purchased these tools specifically to protect the integrity of scientific discourse are, statistically, missing roughly one in four attempts.
There is also the matter of false positives. Originality.ai flagged 19 out of 495 verified human passages as AI-generated — a 3.8 percent rate that means some portion of human students are being accused of something a machine did. Pangram and GPTZero, to their credit, produced zero false alarms on human text. The scoreboard is uneven, which is the most charitable framing available.
Fiction, by contrast, was nearly immune to the imitation problem, with false-negative rates of just one to five percent across detectors. Apparently the machines have a harder time disappearing into narrative. Or fiction readers are simply more suspicious. Both explanations are plausible.
What happens next
Detector developers will presumably update their models, and the language models will presumably update their imitation, and the cycle will continue at whatever pace the humans choose to fund.
The corpus used for testing was drawn from writing published before November 2022, which rules out any contamination by AI. It is a clean experiment. The results are tidy. The tools built to catch AI writing are most reliably defeated by asking the AI to write like a person — which is, when examined calmly, exactly what the AI was always going to do next.