Detection tools lose half their accuracy in the wild. A practical deepfake verification workflow for newsrooms, and the EU labelling rules landing on 2 August 2026.
Detecting synthetic media is an adversarial problem that structurally favours the forger. The answerable question is not whether an image is real, but whether it can prove where it came from.
How data is hidden in images, why detection has overtaken concealment, and why the same techniques now matter more for proving provenance than for hiding messages.
Emotion lives in prosody, datasets are small, and most are acted. Where wavelet-based features and compact classifiers still outperform end-to-end networks, and where they do not.