Preprint finds continual learning worsens deepfake detector bias
New research shows that AI detectors updated to catch new fake video techniques often suffer disproportionate performance drops for specific demographic groups.
ImportanceMaterialEvidenceE2 unreplicatedWrite-upQuick
An arXiv preprint submitted on October 15, 2026, reveals that continual learning (CL) methods used to update deepfake detection models tend to exacerbate demographic unfairness over time. The study, led by Kajaree Das and colleagues, addresses the challenge of static datasets failing to keep pace with evolving generative AI threats.
The researchers analyzed four dimensions of fairness while training detectors progressively on a chronological sequence of 11 real-world deepfake generation methods. They found that as accuracy degraded over time, it typically worsened unfairness, whereas improvements in fairness remained rare.
Crucially, the study notes that this inequitable forgetting is not explained solely by underrepresentation in training data but is likely entangled with representation learning mechanisms. The authors issue a call to action for CL methods to account for social outcomes when balancing the stability-plasticity tradeoff.