The Provenance Paradox: How Video Generation, Content Credentials, and the Detection Arms Race Are Rewriting Computer Vision

When Louis Daguerre fixed the first photograph onto a silvered copper plate in 1839, the immediate reaction was not wonder but epistemic panic. Courts in France and England had to decide, within a decade, whether a photograph was evidence of reality or merely a mechanical illustration that could be staged. The legal apparatus eventually settled on a working presumption: photographs were truthful unless proven otherwise. That presumption held for roughly 170 years. It is now collapsing, not because photography has changed, but because the inverse operation — manufacturing a photograph indistinguishable from a capture — has become trivially cheap.

What the Record Actually Shows

Three documented developments define the current state of computer vision, and none of them is a research curiosity. First, the proliferation of video-capable diffusion and transformer models — OpenAI's Sora architecture, Runway's Gen-3, Kuaishou's Kling, and Google's Veo — has moved synthetic video from laboratory demonstrations to consumer-accessible products with multi-minute temporal coherence. Second, the Coalition for Content Provenance and Authenticity (C2PA), a cross-industry standards body including Adobe, Microsoft, the BBC and the New York Times, has shipped its 2.1 specification for cryptographic content credentials, now deployed in Adobe's Content Credentials ecosystem with over one billion assets signed. Third, the EU AI Act's Article 50, which entered into force in 2024, mandates that providers of AI systems generating synthetic audio, image, video or text must ensure their outputs are marked as artificially generated or manipulated, with technical solutions that are "effective, interoperable and reliable."

Read together, these three developments describe a field that has moved past the research phase and into an asymmetric deployment phase: generation is consumer-grade, provenance is standards-grade, and regulation is jurisdiction-grade — but none of them is solving the same problem. That mismatch is where the real commercial pressure is accumulating.

The Provenance Stack and Its Structural Limits

The first implication for [[Computer Vision Infrastructure and Trust]] is that the industry is attempting to solve a detection problem with a provenance solution, and these are not equivalent. C2PA works by cryptographically binding a chain of custody to an asset at the moment of capture or creation. It answers the question "where did this file come from?" It does not answer the question "is this file authentic?" — a distinction that matters enormously in adversarial settings. A synthetic video generated outside the C2PA ecosystem will simply lack credentials; it will not carry a "forged" label. The verification burden therefore shifts downstream to the consumer of the media, who must treat any uncredentialed asset as unverified. This is a meaningful improvement over the pre-2024 status quo, but it is not a solution to deepfakes; it is a reclassification of the problem from "detect the fake" to "trust only the credentialed."

The second implication is architectural, and it is being resolved in the training pipeline rather than the inference pipeline. The shift from supervised ImageNet-style training to self-supervised vision models — exemplified by Meta's DINOv2 and similar contrastive approaches — has changed what computer vision models actually learn. Self-supervised models trained on billions of internet images learn representations that are malleable in ways that supervised classifiers are not; they can be fine-tuned to generate, segment, or describe with the same underlying weights. This convergence of perception and generation into a single model family means that the same architecture that powers an autonomous vehicle's perception stack can, with minor adaptation, power a synthetic media generator. The security boundary between "vision model" and "vision synthesiser" has effectively dissolved, and enterprise procurement processes have not caught up.

Where the Sceptics Have Been Right

A one-sided reading of the synthetic media narrative would ignore how often detection claims have been retracted by generation improvement. The academic literature on deepfake detection is littered with models that achieved 95%+ accuracy on one benchmark and collapsed on the next. The FaceForensics++ benchmark, widely used in 2019–2021, reported detector accuracies that dropped by 20–40 percentage points when evaluated on out-of-distribution generation methods. According to a 2023 NIST study on deepfake detection, "no single detection algorithm generalises well across different generation methods, and performance degrades significantly when tested on unseen synthesis techniques." The honest conclusion is that detection is a cat-and-mouse game in which the generator currently holds the initiative, because each new generation method invalidates the previous generation of detectors. Any enterprise deploying a single detection model as a permanent control is buying a depreciating asset.

But the Pessimism Frame Is Also Wrong

The reflexive counter-position — that detection is hopeless and provenance is the only answer — understates what multimodal consistency checking actually achieves. Modern detection research has moved past pixel-level artefact hunting toward semantic consistency verification: does the reflection in the window match the lighting? Does the speaker's lip movement match the audio at the phoneme level? Does the background geometry remain physically coherent across frames? These checks are not defeated by better diffusion models; they are defeated only by models that explicitly simulate physics, which is a much harder problem than generating plausible pixels. Furthermore, the C2PA ecosystem is not a static standard; its 2.1 specification includes provisions for asserting training provenance and model identity, which means the provenance layer can evolve to capture information that detection cannot. The downside case is not that provenance fails; it is that adoption is uneven, creating a two-tier media environment where credentialed content is trusted and uncredentialed content is suspect — which is arguably a more honest epistemic state than the pre-2024 presumption of photographic truth.

The Photoshop Precedent

The instructive precedent is not the invention of photography. It is the mainstreaming of Adobe Photoshop in the early 1990s, which made image manipulation accessible to non-specialists for the first time. The immediate prediction was that photography would lose all evidentiary value. That prediction was wrong. What actually happened was a slower, more nuanced adaptation: photojournalism organisations developed internal verification protocols, wire services established chain-of-custody procedures, and the legal system learned to treat photographs as claims rather than proof. The medium did not collapse; the institutions around it evolved. The lesson for synthetic video is precise. The technology will not destroy visual trust; it will force the institutions that consume visual media — newsrooms, courts, regulators, platforms — to develop new verification workflows. The organisations that survive the transition will be those that built those workflows early, not those that waited for the technology to stabilise.

What to Do Before the Next Quarter

For operators, the sequence is specific. First, deploy C2PA credential verification at your ingestion layer; any media entering your systems — whether from employees, contractors, or the public — should be checked for provenance metadata, and uncredentialed assets should be flagged for manual review. Second, audit your computer vision models for dual-use exposure; if your perception models are based on architectures that can be trivially adapted for generation, you need to restrict their deployment environments and monitor for misuse. Third, do not rely on a single deepfake detector as a permanent control; implement ensemble detection with regular retraining on current generation methods, and treat detection scores as probabilistic inputs to a broader verification workflow rather than binary decisions. Fourth, mandate that all synthetic media generated internally carries C2PA credentials, and that your procurement contracts with third-party content providers include provenance requirements.

The Next Two Quarters

Expect the following, stated so it can be checked. The first major legal case involving synthetic video evidence will reach a court in a jurisdiction with AI Act-style disclosure requirements, and the ruling will establish precedent on whether C2PA credentials are sufficient to meet the "effective, interoperable and reliable" standard of Article 50. Detection model performance will continue to degrade on out-of-distribution generation methods, and at least one prominent commercial detector will be shown to fail on a new generation architecture. Provenance adoption will bifurcate: news organisations and regulated industries will adopt C2PA rapidly, while social media platforms will adopt it selectively, creating a fragmented trust landscape. And the first enterprise will publicly disclose a synthetic media incident — a deepfake used to bypass visual verification in a fraud or social engineering attack — that could have been prevented by provenance checking. The organisations that come out ahead will be those that treated this as a trust infrastructure problem, not a detection problem.