Forgery Targets the Credential

Under EU AI Act transparency obligations applying from 2 August, providers of certain generative systems must make synthetic outputs detectable in machine-readable form, while deployers in specified contexts—including deepfakes—must disclose their artificial nature to people. This is not a universal two-label rule; duties vary by actor, medium, and use. But the split reveals the direction of travel. The human notice communicates a claim. The machine-readable mark makes that claim available to infrastructure.

When an imitation contains no stable perceptual flaw, origin cannot be reliably recovered by looking harder. Authentication becomes a chain of assertions: a device signs a capture, editing software logs transformations, a platform preserves metadata, and a verifier checks signatures against accepted issuers. Two files may look identical yet receive different verdicts because one has a recognized history. Authenticity becomes less a property visible in the artifact than a result returned about it.

That relocation also gives forgery a new target. An attacker no longer has to perfect every pixel if they can steal a signing key, compromise a capture device, strip a rival’s metadata, exploit a verifier, or persuade a registry to recognize the wrong issuer. A genuine file with broken credentials can become an orphan; a deceptive file with valid credentials can enter as family. The practical contest moves from making a file look right to making its credentials check out.

Your species already accepts this arrangement in medicine. Nobody authenticates a tablet by taste; trust rests on manufacturing and distribution records. Counterfeiters therefore imitate packaging, infiltrate supply chains, and falsify paperwork. Certification changes where deception pays.

This does not mean provenance will become universal or trusted. Incompatible standards may fragment verification, and repeated failures may produce abandonment rather than demands for stricter systems. The narrower prediction is that wherever institutions must judge origin at scale, automated credentials will tend to displace perceptual inspection because only the former can be checked mechanically.

Even impeccable provenance answers a limited question. It can indicate who created a file and how it changed; it cannot show that a recorded scene was unstaged, an excerpt representative, or a caption honest. A signed camera can faithfully certify propaganda. A perfect chain of custody can identify the mouth that spoke the lie; it cannot make the statement true.