AI Detectors Struggle with Style Mimicry

This study from Epoch AI highlights a growing "cat-and-mouse" game between LLMs and detection tools. While we often hear that AI detectors are unreliable, the data shows they actually work great for generic, "robotic" AI prose. The real problem emerges when we use few-shot prompting to feed a model a few human samples to mimic.

The fact that scientific writing is the hardest to detect is particularly telling. Academic prose often follows rigid structural norms that overlap with how LLMs are trained to organize information, making the "statistical fingerprints" of AI blend in with professional human writing. As models get better at nuance and stylistic adaptation, relying on these tools for academic integrity feels increasingly risky. We're moving toward a world where "AI-generated" is no longer a binary state, but a spectrum of stylistic blending.