Provenance signals are taking center stage as researchers and technologists look for reliable ways to verify authentic media and blunt the spread of deepfakes. That shift follows mounting skepticism of visual evidence online and a growing view among experts that creators now bear the burden of proof to show their content is real.
The forensic approach to deepfake detection
An incident earlier this year illustrates the challenge. Reporters sought analysis of a video aired by an Iranian outlet that appeared to show a U.S.-made Tomahawk missile striking a school and killing more than 175 people, most of them children. U.S. officials denied responsibility, and given the prevalence of generative AI and fabricated media, analysts were wary of the footage. Initial doubts centered on the clip’s low quality, delayed release, and its sole distribution by an outlet with a poor record for accuracy.
To evaluate the video, a researcher aligned frames using a fixed background reference to check for a realistic flight path. In the stabilized sequence, the missile’s movement appeared nearly linear. The analyst then estimated scale, concluding the image was consistent with a roughly 6-meter Tomahawk at about 100 meters, which matched a 0.3-second gap between visible impact and the sound of the blast, given the speed of sound at 343 meters per second. After hours of checks, the assessment concluded the video was authentic.
Even so, the case underscored a broader shift. Prior studies at the University of California, Berkeley, found viewers increasingly doubt video evidence, struggling to accept authenticity even when artifacts of manipulation are absent. The experience reinforced a new posture: a presumption of fake that requires creators to demonstrate provenance.
Signaling a video’s provenance
Because traditional forensics cannot operate at internet speed and scale, attention is turning to provenance signals embedded at capture. These include content credentials and watermarks that can travel with media wherever it is shared.
Credentials aligned with the C2PA standard attach a cryptographically signed manifest at the moment of capture, logging device details, time and location, and subsequent edits. Any tampering with the content or its history can later be detected. Watermarking systems such as Google’s SynthID embed an imperceptible signal in pixels or audio that identifies origin. Apple has introduced sensor-level signing on some newer iPhones, cryptographically sealing raw pixel data as it is captured and saving a reference image alongside the editable photo so changes can be compared precisely.
These mechanisms are not designed to trap bad actors. Their value lies in enabling good actors to prove authenticity. In a media environment governed by a presumption of fakeness, widespread adoption could allow trusted content to crowd out manipulated material by shifting the focus from catching impostors to empowering verified creators.
Real-time verification with provenance signals
The same distrust now reaches routine video calls. In September 2024, Senator Ben Cardin of Maryland, then chair of the Senate Foreign Relations Committee, received what appeared to be a video request from Dmytro Kuleba, who had recently served as Ukraine’s foreign minister. During the call, the visitor pressed Cardin about allowing Ukraine to fire long-range missiles into Russia, an unusual line of questioning. Cardin ended the conversation. The State Department later confirmed Kuleba had not requested the meeting, and Senate security warned staff that foreign adversaries, possibly Russian, might attempt similar deceptions.
Real-time deepfakes rely on face-swap or avatar techniques and voice conversion, all feasible on standard laptops using commercial tools or open-source software. Live impostures must be generated at roughly 24 frames per second, which forces shortcuts such as simplified facial rendering, imperfect interaction with surroundings, or slight audio-video drift. These artifacts can be exposed with targeted forensic checks, but there is little time to analyze during a live call.
Provenance signals offer a path forward. Video-calling platforms could cryptographically sign each frame from a physical camera, similar to Apple’s approach for still images. Because a face swap occurs between the camera and the conferencing software, platforms could verify whether incoming frames are sensor-signed or altered in transit. The aim is not to hunt the impersonator but to let legitimate participants prove they are present.
For much of the digital era, forensic analysis was enough to resolve disputes about authenticity. The rise of AI has changed the speed and scale of manipulation. Trust now requires pairing forensics with provenance, allowing verified content to be recognized quickly while investigative techniques continue to expose fakery. Efforts such as AI microcredentials that help engineers deliver trustworthy systems reflect a broader push to build technical and professional frameworks for reliability in an era of synthetic media.