Authenticity must become demonstrable

Digital-forensics researcher Hany Farid is calling for a shift in how people assess online photos, video and audio, arguing that creators should increasingly prove where legitimate material came from instead of expecting viewers to identify every fake.

Farid, a Dartmouth College computer science professor, described the change in an IEEE Spectrum essay published as synthetic media tools make convincing manipulation widely accessible. His proposed approach centers on provenance signals added when content is captured or created. Those signals would not eliminate deception, but would give news organizations, courts and ordinary users a way to distinguish material with a traceable history from files that arrive without one.

The argument grew out of Farid's own analysis of a disputed video that appeared to show a Tomahawk missile striking a school. He initially treated the recording with suspicion because of its low quality, delayed appearance and limited distribution. Frame alignment suggested the missile followed a physically plausible path, while its apparent dimensions and the interval between visible impact and sound were consistent with the scene. After several hours of analysis, he concluded the recording was authentic.

That experience exposed a new problem: the absence of obvious manipulation was no longer enough to create confidence, even for a specialist. Farid said conventional forensic examination is too slow to handle the volume and speed of media circulating online. Provenance systems, by contrast, can operate from the point of capture and remain attached through later editing.

One option is the C2PA standard, which uses a cryptographically signed manifest to record information about a file's origin and editing history. If either the content or its documented history is altered, verification tools can detect the inconsistency. Another approach is watermarking, including systems such as Google's SynthID, which places a machine-detectable signal inside generated pixels or audio. Farid also pointed to camera designs that sign raw sensor data and preserve a reference image for comparison with an edited version.

These mechanisms have limits. Metadata can be removed, not every device supports credentials, and a file without provenance is not necessarily false. Farid's case is that the systems should be judged by whether trustworthy creators can demonstrate authenticity, not by whether they catch every adversary. Broad adoption by cameras, publishers and platforms could create a recognizable pool of verifiable material.

The stakes extend beyond edited clips. Real-time face swaps and voice conversion can support convincing impersonation during video calls using ordinary modern computers. That makes source verification and trusted communication channels important alongside technical media analysis.

The emerging model therefore changes the default question. Instead of asking viewers to prove a suspicious recording was fabricated, publishers and creators would provide evidence that legitimate content has an intact, inspectable chain of origin.