Google opens SynthID Detector. A watermark check is not a truth verdict

A detector can answer a technical question without answering the editorial one. Google's 7 October announcement expands access to SynthID Detector globally in English, giving more people a route to check supported media for a watermark. It does not make every clip a verified account of what happened. Dated announcement

Keep these distinctions

• A supported watermark check is narrower than universal detection.
• Finding generated material does not establish whether its surrounding claim is true.
• No detected watermark is not a certificate that a file came from a camera.

What changed this week

Google says the tool can check images, video and audio made with its systems or participating partners, naming OpenAI, NVIDIA and Kakao, with Apple described as forthcoming. That is Google's stated scope, not evidence that every output from every named company is covered. The original detector announcement was in May 2025; the October change is broader access, not the invention of watermarking.

The marker is not the message

DeepMind describes SynthID as embedding imperceptible signals during generation. Its product page says image and video watermarks are designed to withstand changes such as cropping and lossy compression. Those are supplier claims: PING has not run an attack, recompression or translation experiment. Do not turn a robustness description into a guarantee that all edits preserve a detectable signal. How Google describes the system

The cover photograph illustrates a different, visible kind of watermark. It is deliberately labelled as a historical contrast. Looking for lettering in a picture is not how SynthID detection works, and this image is neither a result from the detector nor a sample of generated news footage.

Use a check alongside the reporting

Our proposed newsroom workflow starts by preserving the file and its source: who supplied it, where it first appeared, when it was retrieved and what the accompanying claim actually says. If you run a supported detector, retain the result and the tool's stated scope. Keep that technical observation separate from what you can establish about the event.

Then ask the reporting questions the watermark cannot settle on its own. Can the location and timing be independently corroborated? Does the caption match the material? Is a synthetic illustration honestly labelled, or being passed off as an event recording? Contact an accountable source where necessary. These are editorial recommendations, not a claim that PING has authenticated a particular viral clip.

Be careful with the negative result

A tool that searches for one family's markers cannot logically certify that every unmarked file is genuine. Missing coverage, an unsupported workflow or a transformation may leave the editorial question unresolved. Describe the limited observation rather than converting uncertainty into the label 'real'.

For publishers working between London and African audiences, the practical benefit is a repeatable evidence trail—not a new reason to repost a sensational clip quickly. Include what the check did and did not establish. The same discipline applies to marketing images, translated clips and a source's own demonstration.

We have not evaluated detector accuracy or claimed that a particular watermark satisfies every legal provenance obligation. This is a documented guide to interpreting a newly accessible tool. Explore related reporting on PING's Networks desk.