AI Wrote My Paper and All I Got was This False Negative:* Measuring the Efficacy of Commercial AI Text Detectors

Academic institutions and publishers are increasingly relying on commercial AI-generated-text (AIGT) detectors to combat plagiarism and verify authorship in the era of large language models (lLMs). As the number of submissions to academic security conferences increases at an exponential rate, the temptation to deploy detectors to protect academic integrity commensurately increases. Unfortunately, even when benchmarks are disclosed, these detectors lack appropriate performance characterizations for use in evaluating academic security writing. In this paper, we conduct a comprehensive empirical evaluation of leading AIGT detector performance on academic security writing. We collect a dataset of all papers ($\mathbf{N} \boldsymbol{=} \mathbf{6, 2 9 5}$) from Tier-1 conferences (IEEE S&P, CCS, NDSS, and USENIX) prior to the public release of ChatGPT. We then create an AIGT version of each of these papers and use this combined dataset to evaluate the top five most popular AIGT detectors, based on Tranco-list rankings. Our evaluation not only finds that performance varies wildly across AIGT detectors (e.g., FPRs between 0.05 % and 68.6 % and FNRs ranging between 0.3 % and 99.6 %). Even more critically, these detectors are trivially circumvented by a simple adaptive adversary (e.g., a TPR reduction from 94.2 % to 2.5 %). Ultimately, the limitations of current detector-based approaches create an adversarial environment in which achieving authorship verification remains out of reach.