A news photo now sometimes carries a small icon in the corner — a Content Credentials mark, built on the C2PA standard backed by Adobe, Microsoft, the BBC, and other members of the coalition — that a reader can click to see the image's edit history back to the moment of capture. It is the first widely deployed answer to a problem that did not used to need one: proving a piece of media is what it claims to be.
Human perception was never built to detect synthesis. It was built to detect the kinds of manipulation that were expensive under old production methods — splices, obvious edits, poor lighting continuity. Generative systems like OpenAI's Sora do not need to fool a forensic expert, only the ordinary act of watching, and ordinary watching was already a low bar — which is exactly why the practical response, provenance systems that cryptographically sign content at the moment of capture, only solves the problem for footage that opts in, leaving the absence of a signature meaning less than people expect.
Institutions whose authority rests on being trusted eyewitnesses — courts, newsrooms, insurance adjusters — face the heaviest burden, because their processes were built around the assumption that a recording is reasonably strong evidence on its own, an assumption that no longer holds without a chain of custody attached.
A market for verification-as-a-service is emerging around exactly this gap — tools that assess likely synthetic origin, track provenance metadata, and flag inconsistency, including Google DeepMind's SynthID watermarking approach — though these tools are themselves in an arms race with the generation systems they are trying to catch, and will likely never close the gap completely.
The durable adaptation is procedural rather than technological — treating unverified recordings the way courts have always treated uncorroborated testimony, as a claim requiring context and corroboration rather than as self-evident proof. C2PA's growing adoption is a start. It is not, by itself, the fix.