A detectable watermark is not a provenance record

The shot

As retrieved on 16 September 2026, Google DeepMind's own SynthID page states the technology 'embeds digital watermarks directly into AI-generated images, audio, text or video,' extending an approach first built for images to video output, and DeepMind's Veo product page states directly that 'videos made with Veo will be marked with SynthID.' Both are living documents describing current products rather than a single dated announcement, though the Veo page distinguishes a current flagship version, Veo 3.1, from an earlier Veo 3.

What the documents show

The SynthID page describes, in its own words, how the video watermark works: it 'adds an invisible digital watermark to an AI-generated image (or video segment),' applied 'the moment content is created,' and 'designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression.' That is DeepMind's own account of what the watermark survives; the page calls the mark 'imperceptible to humans' but detectable 'by SynthID's technology' — a claim about detectability, not that the mark can never be stripped, and the material examined makes no stronger claim. Separately, the Veo page states outputs also 'undergo safety evaluations and checks for memorized content' for privacy, copyright and bias — a distinct process from watermarking, applied before generation reaches a user.

The workflow

For a production using Veo-generated shots, the sequence the pages describe is: a clip is generated, a SynthID watermark is embedded automatically without a separate opt-in step, and a viewer can later check it using DeepMind's own tools — a consumer-facing check through Gemini, where a user uploads a file and asks whether it was AI-generated, plus a separate SynthID Detector portal DeepMind describes as still gathering feedback from journalists and media professionals rather than as a finished product. A production relying on watermark detection to verify a clip's origin depends on a tool DeepMind itself calls still being refined with early testers.

What the tool does not change

This is an editorial boundary worth stating plainly: a watermark surviving common edits, as described here, is not the same as a chain-of-custody record of who generated a clip, when, and with what prompt — DeepMind's pages describe detecting AI generation, not the full manifest-based provenance the C2PA specification defines. A production still needs its own record of which shots were generated, by which tool and version, independent of whether a platform's detector later confirms it.

• Does the platform receiving a Veo-generated clip actually run SynthID detection, or is the watermark's presence being assumed rather than checked?
• Is a claim that SynthID content is 'undetectable as AI-generated' being confused with DeepMind's own, opposite claim that the mark is designed to be detected by its own tools?
• Does the production's own paperwork record provenance independently of the watermark, in case a pipeline step falls outside DeepMind's stated list of survivable modifications?

SynthID for video is DeepMind's own described mechanism for marking and later detecting its output, not an independently audited or unremovable seal. Treating it as full production provenance overstates what the company's own pages claim.