AI Video Detection in 2026: Watermarks Advance as Detectors Falter

The verification problem has changed

For most of the deepfake era, the main question was whether a face had been swapped onto someone else's body. In 2026 that question has been replaced by a harder one. Sora 2, Veo 3.1 and similar text-to-video systems have made fully synthetic footage common, so verifiers now have to ask whether the scene ever happened at all.1 There is less to compare against: a text-to-video clip has no original recording whose seams might show.3

The industry has answered with two different strategies. One is detection, which means analysing pixels after the fact. The other is provenance, which means marking content when it is created. The evidence from this year points one way. Detection is useful for sorting content but unreliable as proof, and the only real progress on scale has come from provenance, even though large gaps remain.

How the generators raised the bar

The current situation goes back to late 2025. OpenAI released Sora 2 in September 2025, and realistic synthetic video then spread across social platforms.22 Coverage at the time described Sora clips as high-resolution, with synchronized audio, and noted that the "cameo" feature let users place real people's likenesses into generated scenes.21 NewsGuard tested Sora 2 with 20 known false claims. It produced convincing news-style videos for 16 of them, 11 on the first attempt. The visible watermark could be removed with free online tools in under four minutes.27

The technical change matters as much as the realism. Older deepfakes were built with generative adversarial networks, which left frequency-domain artifacts that classifiers learned to spot. Diffusion-based systems such as Sora 2 and Veo 3 build video by removing noise step by step, which avoids many of those signatures.25 Deepfake researcher Siwei Lyu has said current models can produce stable faces without warping, and that voice cloning has crossed a threshold where it can no longer be told apart from real speech.22 Clip-length limits that once helped flag suspect videos are also disappearing as models produce longer, higher-resolution output.25

Detection: confident marketing, worrying benchmarks

The detection market is growing quickly. NVIDIA introduced a Synthetic Video Detector at SIGGRAPH 2026. It scores clips frame by frame for newsrooms, and NVIDIA reports accuracy of up to 92% on uncompressed video, 87% at 15% compression and 82% at 50% compression.4 It has been integrated with Wowza's video intelligence framework, which reaches livestreaming deployments in more than 170 countries.4 ZeroGPT added a video detector in August, extending its text and image checks to a third format.5 Gartner published its first deepfake-detection quadrant in June 2026 and named Reality Defender a "Market Shaper."1

The accuracy numbers vary widely depending on who is reporting them. One vendor comparison lists its own tool at 97% and competitors between 80% and 96%.9 An industry summary says leading tools reach 92–98% on standard deepfakes, but only 70–80% on live calls.24 Academic work also reports high figures under controlled conditions: the EA-Swin paper reports accuracy around 98% across Veo 3, Sora 2, Kling and other generators on its own dataset.28

Independent stress tests tell a different story. The DF26 benchmark contains 271 real videos and 2,420 synthetic clips from seven modern video models. It found that both human viewers and state-of-the-art detectors performed close to random chance.7 Its authors do not claim that all synthetic video is undetectable. Their argument is that high scores on familiar test sets do not show that a detector will work against generators it has never seen.7 Earlier benchmarking found the same pattern: detection accuracy dropped by about half when moving from lab datasets to real social media content.29

The most likely explanation for the gap is that detectors learn the fingerprints of specific generators, and those fingerprints change with each model release. Even vendor-friendly reviews acknowledge that new generators create a detection lag for every provider.8 Hands-on testing has shown failures in practice: one tool labelled a Google Veo 3 demo video as real.23 Some practitioners say openly that detection is in an arms race it is losing, partly because newer generative models are trained specifically to evade existing classifiers.2

Our view is that published detector accuracy figures in 2026 should be treated as upper limits measured under good conditions, not as performance anyone can expect in the field. NVIDIA presents its own product this way: it says the detector supports, and does not replace, source checking and provenance tracking.4 Testers also found that its confidence score changed across different parts of a clip, which could help flag videos that are only partly manipulated.4

Humans are not a fallback

If detectors struggle, people struggle more. One summary of recent research puts human accuracy against high-quality synthetic video and images at about 25%.6 Earlier iProov research found that only 0.1% of participants identified every deepfake correctly, and that video fakes were 36% harder to spot than manipulated images.29 Common advice still includes watching for odd blinking, teeth that look like a single block, waxy skin, flickering hair and garbled background text.23 These tips are still free and occasionally useful. But guides now admit that compression removes many of these signs and that current models pass a casual viewing.3 Asking someone on a live call to turn their head slowly still catches many face-swap models, but that only applies to real-time impersonation, not to clips already circulating online.6

Provenance has momentum

The most significant event of the year came on May 19, 2026. OpenAI joined the C2PA steering committee and said it would add Google DeepMind's SynthID watermark to images from ChatGPT, its API and Codex, alongside the C2PA Content Credentials it already attaches.11 On the same day at Google I/O, Google announced that C2PA verification and SynthID detection would come to Search and Chrome. Pixel phones would embed C2PA in video, and a partnership with Instagram would apply Content Credentials labels automatically.11 Sundar Pichai described the user experience: right-click an image in Chrome, ask whether it was AI-generated, and get an answer with context.20 Kakao, ElevenLabs and Nvidia joined as adopters,11 and Anthropic announced in August that it would add text watermarking and C2PA metadata for supported files.15

The two layers work in complementary ways. C2PA is signed metadata that records origin and edit history, but screenshots, re-encoding and platform pipelines can strip it. SynthID is embedded in the pixels or audio and survives those operations, but carries much less context.18 OpenAI has described the two as reinforcing each other, with SynthID keeping a signal when metadata is lost.20 Regulation adds pressure: California's SB 942 took effect January 1, 2026, and enforcement of the EU AI Act's Article 50 transparency rules began in August.12

Reports on scale differ in a way worth noting. One account says SynthID has marked more than 100 billion files since 2023,11 while another cites more than 20 billion images.14 The two figures may measure different things, all files versus images only, but the gap shows how loosely deployment numbers are reported. Sources also disagree on the date of OpenAI's layered provenance announcement. One gives April 2026,14 while most coverage dates it to May 19.1120

The gaps provenance cannot cover

Provenance only works when the generator takes part. Midjourney does not embed C2PA credentials. Nikon's C2PA implementation was suspended after a signing vulnerability forced full certificate revocation.14 SynthID detection only identifies content from systems that use it, so output from unwatermarked tools, including open-weight models, leaves no signal to find.1413 Both Google's and OpenAI's documentation say a missing watermark or manifest proves nothing: it means the file was not detected as output from a participating generator, not that a person made it.1816

Distribution is the other weak point. X strips metadata to reduce file sizes, which removes the C2PA manifest along with it.2 Platform AI labels are spreading, but enforcement is still inconsistent.10 C2PA also has a limitation in principle: it records what was declared when the file was signed, not whether that declaration is true.14

What it adds up to

The sources broadly agree on a layered workflow. Check provenance first, use statistical detectors to sort content, look at the source and its history, and leave decisions with real consequences to human reviewers.167 Where they differ is in how confident they are, and on that point the independent benchmarks are more convincing than vendor marketing.

In our reading, 2026 is when verification moved away from asking whether something looks fake and toward asking whether its origin can be proven. For the large share of video made with participating tools such as Sora, Veo and ChatGPT image models, verification is becoming easier and is being built into browsers and search.2026 For everything else, the defenders are behind. This lopsided result means that people spreading disinformation have a reason to use unwatermarked tools and to post through platforms that strip metadata. Until provenance becomes universal, and until detectors are tested against generators they have not seen before, an unlabelled clip in 2026 should be treated as unverified, not as authentic.