Deepfake detectors are meant to protect against misinformation, but their failures carry societal weight. A new arXiv preprint, "Time to Fake It: Continual Machine Learning Intensifies Unfairness in Deepfake Detection," argues that the common strategy of continually updating detectors on new generator outputs can backfire.

The abstract notes that detectors trained on static datasets miss newly arising threats as generators evolve. Continual learning is often proposed as a fix, but the authors suggest it may intensify unfairness—likely by amplifying biases in the data or the update process. The exact mechanisms are not detailed in the available abstract, which cuts off mid-sentence.

The paper frames this as a dual problem: technical accuracy and social equity are intertwined. If continual learning makes detectors more accurate overall but less fair for certain groups, the trade-off demands scrutiny. The preprint is a single source, so its claims should be read as a preliminary warning rather than a settled conclusion.