Authenticity Must Begin at Capture
Deepfake labels are useful warnings. Democratic trust also needs provenance, competent forensics and remedies fast enough to matter.
The first viral hour belongs to the accusation. The correction enters a different contest.
A fabricated recording can appear, circulate through private groups and become the subject of public argument before a platform applies a warning. The target may eventually prove that the file was manipulated and still lose a reputation, an election or personal safety.
Deepfake policy cannot be reduced to placing labels on synthetic media. A durable system must make authentic material easier to trace, disputed material easier to preserve and malicious impersonation faster to remedy.
A label arrives downstream
Visible disclosure gives a viewer a reason to pause. Traceable metadata can deter ordinary creators from presenting generated material as real. Platforms can reduce recommendation while a file is examined.
But dishonest uploaders do not volunteer labels. Cropping, screen recording and re-encoding can strip visible and embedded signals. Detection systems lose accuracy when media is compressed, edited or produced by a generator outside their training data.
The US National Institute of Standards and Technology's 2026 evaluation found sharp deterioration in deepfake detectors under operational conditions. A classifier score can support an investigation. It should not be treated as a verdict about a person.
The more useful question begins earlier: can the recipient verify where an assertedly authentic file came from and what happened to it?
Build a chain, not a truth machine
The C2PA technical standard allows devices and publishers to attach signed records describing a file's origin and edits. A newsroom can show that footage came through its verified workflow. An election authority can sign an official announcement. An investigator can detect that a file changed after capture.
These credentials do not prove the scene was honest. The C2PA explainer explicitly separates provenance from truth. A staged event may be authentically recorded, while genuine evidence from an old phone may carry no credential.
That limitation is a reason to use provenance carefully, not to discard it. The standard provides claims that can be checked. Courts and journalists must still corroborate what the scene means.
Public institutions should lead because their responsibility is clearest. Election bodies, police evidence systems, government communications and public broadcasters can sign material from capture or creation through publication. Private newsrooms and device makers should receive interoperable tools, while unsigned citizen footage remains open to ordinary verification.
Preserve the disputed object
A credible forensic process begins with the original file, a cryptographic hash and a documented chain of custody. It should preserve upload records, platform transformations, detector versions and expert findings. Screenshots and reposted fragments rarely provide the same evidentiary value.
Countries need accredited laboratories and transparent validation of forensic methods. Courts should understand confidence and error rather than receive a detector percentage as scientific certainty. Platforms must retain relevant records promptly when a victim or authority gives lawful notice.
The same process protects defendants. If an accusation relies on questionable media, the accused should be able to examine the evidence and the method used to authenticate it.
The strongest countercase
Persistent provenance can expose journalists, dissidents and witnesses. A state may declare unsigned footage false, or make trusted credentials available only to approved institutions. Expensive signing systems could deepen inequality between large publishers and citizens.
A democratic design must preserve anonymous and pseudonymous speech. Verification can confirm that a credential is valid without revealing a civil identity. Lack of a credential is “unknown,” not “fake.” Tools should be open and inexpensive, and legal processes should continue to accept corroborated unsigned evidence.
Authenticity infrastructure must not become an identity checkpoint for the internet.
Remedy at network speed
Victims need an urgent civil route to preserve evidence and temporarily slow recommendation or forwarding while a court examines a credible impersonation claim. Final findings should trigger notice to accounts that received or shared the fabrication where technically possible.
Election periods need standing escalation channels and published decisions. The public should know what evidence led an authority to restrict a file. Secret takedown processes create a second trust problem even when the initial concern is genuine.
Law must distinguish disclosed art, parody and satire from deceptive impersonation intended to cause harm. Broad bans on synthetic media will burden legitimate speech while organised attackers move to private channels and foreign services.
Trust requires several institutions
Deepfakes create a liar's dividend: authentic evidence can be dismissed merely by claiming that AI produced it. If every recording becomes optional reality, power shifts toward those with the largest audience and away from those with the best evidence.
Labels help attention. Provenance helps establish history. Forensics helps adjudication. Rapid remedies limit the period in which a fabrication can act without resistance. No layer is sufficient alone.
Democratic trust cannot be restored after every incident by an icon placed over a ruined reputation. Authenticity must begin at capture, and justice must move closer to the speed of the lie.
The Global Federation treats common truth as democratic infrastructure: it must protect evidence without turning verification into a system of permission.