Epoch AI tested three leading AI text detectors — Pangram, GPTZero, and Originality.ai — against passages generated by language models instructed to imitate a specific author's writing style. The results were striking: up to 18 percent of AI-generated content went completely undetected.

The vulnerability is particularly pronounced in scientific and academic writing. When models produced text in the formal, structured prose common in research papers, detectors struggled most. Scientific writing follows predictable patterns — standardized section structures, cautious language, citation conventions — that apparently overlap enough with natural AI output patterns to confuse detection algorithms.

Three implications emerge from the study:

First, the tools are losing an asymmetric race. Making a language model imitate a writing style is trivially easy — it requires only a single prompt instructing the model to write like a specific author. Detection, by contrast, is a hard statistical problem that becomes harder as the models themselves improve. The gap between offense and defense is widening, not narrowing.

Second, the real-world consequences are significant. Universities use AI detectors to police academic integrity. Publishers screen submissions with them. Hiring managers evaluate written work using detector scores. A tool that misses nearly one in five AI passages when someone applies a simple style-imitation prompt is not a reliable basis for decisions that affect people's careers and academic standing.

Third, the detection gap will likely grow. As frontier models become more sophisticated at mimicking human writing patterns across diverse styles and domains, the statistical fingerprints that detectors rely on will become harder to isolate. Each new generation of language models narrows the gap between human and machine text, and the detectors must constantly play catch-up.

The study's authors recommend that institutions stop treating AI detectors as definitive evidence and start treating them as weak signals at best. The honest path forward, they argue, is redesigning assessment around process and verification — focusing on how work is produced and verified rather than trying to catch AI-generated outputs after the fact.