Test-time adaptation improves AI text detection, MIT and Cornell report
MIT and Cornell researchers applied test-time adaptation (TTA) to AI text detectors, allowing deployed models to adjust using newly uploaded text without labels or full retraining. TechNews 科技新报 reported that the method detected more than 93% of AI text disguised as human writing, compared with about 27% for a commercial detector; in a simulated shift in human language, it reached 85% accuracy versus 73% for the traditional method. The research was scheduled for presentation at NeurIPS.
TechNews 科技新報 reports that TTA could help AI text detectors adapt as writing styles and model outputs change, a relevant capability for detecting disguised AI text without restarting training from scratch.
AI 文字偵測器學會邊用邊調整,識破偽裝文字準確率逾九成 | TechNews 科技新報


麻省理工學院(MIT)與康乃爾大學用「測試時調適」(test-time adaptation,TTA)技術,讓 AI 文字偵測器上線後仍能持續學習。實驗顯示,面對刻意偽裝成人類寫作的 AI 文字,商用偵測器只能辨識出約 27%,新方法可抓出逾 93%。這項研究將在神經資訊處理系統大會(NeurIPS)發表。
AI文字偵測器本身也是AI模型。開發者通常蒐集已知出自人類或AI的網路文字來訓練它,但模型上架後,如果遇到和訓練資料差異太大的文字,就可能判斷失準,這種情況稱為「分布偏移」。人類寫作習慣會隨時間改變,大型語言模型也不斷推出,還有人用「人性化」(humanizer)改寫工具把AI文字修飾得更像人。論文第一作者、康乃爾理工學院(Cornell Tech)研究生Kevin Ren說,偵測器和改寫工具的攻防就像重複的循環,可能永遠不會結束。