Deepfakes and AI detection

In short: Deepfakes are deceptively real images, videos and voices generated or altered by AI. Detecting them is hard. Source checking, provenance records and watermarks help.

In more detail: Detection programs for AI text are unreliable (many false alarms, especially for non-native speakers). For images there are hints but no guarantee.

In detail

• Hints in images: hands and teeth, text in the image, reflections, unnatural skin, contradictory shadows.
• Videos: lips out of sync, flickering edges, rigid eyes.
• Voices: breathing missing, intonation flat. Fraud with cloned voices (“grandchild scam 2.0”): call back on a known number, family code word.
• Watermarks: hidden patterns in AI images (SynthID and others). Detectable only if the provider takes part, removable by editing.
• C2PA (content credentials): standard for provenance information in files, shows camera, editing and AI use.
• Law: the EU AI Act requires labelling, see EU AI Act and rules for AI.
• Practically: check the source, reverse image search, fact checkers, be sceptical of emotional content.

See also: GANs: generator versus discriminator, AI phishing