EU AI Law Requires Deepfake Labeling, Study Shows Human Limits in Detecting Fake Videos

A study by University of Florida psychologist Natalie Ebner and her team compared human and AI performance in recognizing synthetic media. While algorithms identified deep‑fake images with up to 97 % accuracy, humans performed only around 50 % on photos and 63 % on short videos, where AI accuracy fell to near‑random levels. The researchers are investigating the visual cues that help humans spot deepfake videos but evade AI models, and suggest that combined human‑AI approaches may offer stronger protection against synthetic‑media fraud.

At the same time, the European Union’s Artificial Intelligence Act now obliges AI developers and providers—including firms such as Anthropic and OpenAI—to clearly label deepfakes and other AI‑generated content as synthetic. The rule aims to curb misinformation, political manipulation (e.g., a forged video of President Volodymyr Zelensky urging surrender) and non‑consensual exploitation (e.g., AI‑generated pornographic images of Taylor Swift). Exemptions exist for clearly artistic or satirical works, such as a 2023 viral AI‑created image of Pope Francis in a white puffer jacket.