Deepfake Detectors Lose 15‑20% Accuracy on Real‑World Social Media Content
The authors assembled a large, realistic dataset of videos, images, and audio from Instagram, TikTok, YouTube, and Facebook, then benchmarked several popular deepfake detection models. Across these platforms, detection accuracy fell by 15‑20%, with fast models still missing many fakes, highlighting the need for more robust, adaptable solutions and occasional human review.