A new preprint on arXiv examines how AI-text detectors hold up when attacks are applied to both human and machine text. The authors note that previous evaluations have usually focused on attacking machine-generated text, while ignoring the fact that human-written text is also frequently altered in real-world settings.
The paper argues that this one-sided approach gives an incomplete picture of detector robustness. By evaluating attacks on both sides, the researchers aim to better reflect realistic conditions where both human and AI-generated content may be edited or manipulated.
Because the abstract is truncated, specific results and methods are not yet available in the source. The key contribution, as stated, is the call for a two-sided evaluation framework that accounts for modifications to human text as well as machine text.