Researchers publish method for detecting synthetic training data
The technique could help labs audit whether their training sets contain unlabelled AI generated content.
A research team has published a method for detecting synthetic, AI generated content within large training datasets, a tool the authors say could help labs audit their own data pipelines for unintended contamination.
The method relies on subtle statistical signatures left by generation models, and the team said accuracy remains imperfect but well above chance across the content types tested.
