AI and content · Update and practical guide

Google expands SynthID Detector: how content teams can verify AI-generated media

A convincing image or rapidly spreading video arrives in your team inbox. Is it a camera capture, generated media or a partially edited file? Expanded SynthID Detector access adds a useful check, but publication decisions require more than one tool result.

A magnifying glass reveals a fingerprint inside a media frame, illustrating watermark detection
Key takeaway

A detected watermark can provide evidence of generation or editing by a supported tool. An absent watermark does not establish authenticity, and production provenance does not establish that a claimed event is true.

What Google announced

On 7 October 2026, Google expanded SynthID Detector access globally in English. The announcement covers images, video and audio from Google and named partners, including OpenAI, NVIDIA and Kakao, with Apple described as coming later. Official announcement.

This scope does not imply detection of every file from every generator, or that every historical output from a partner carries a detectable mark. Interpret each file result within the tool's supported coverage.

A watermark is not a visible logo

Google describes SynthID as embedding an imperceptible digital signal during generation that its detection technology can identify. It says image and video marks are designed to withstand common changes such as cropping and compression. Technology overview.

That is not a guarantee against every modification. A missing mark may reflect an unsupported generator or an undetected signal. The defensible editorial conclusion is that this check did not detect evidence, not that the file has been proven authentic.

A proposed content-review workflow

  1. Preserve the incoming file: request the original where possible and record the link, sender and receipt time.
  2. Separate the claim from the image: identify the asserted location, date and event.
  3. Check the file: use SynthID Detector for currently supported media and record its exact result.
  4. Investigate context: look for earlier publication, independent accounts and an explanation from the contributor.
  5. Make a human decision: publish with an accurate description, request evidence or hold the item.

This is CROMBA's proposed operating process, not a claim that the detector performs every step. Check service terms and team policy before uploading confidential material or files you are not authorized to share.

Turning results into editorial decisions

SituationSuggested response
A watermark is detectedReview the result scope, seek an explanation of generation or editing and disclose appropriately
No watermark is detectedContinue source and context checks; do not certify authenticity
The check fails or is inconclusiveRequest another copy or independent evidence and record the unresolved status

A genuine photograph can carry a false date. A generated image can be a suitable illustration when clearly labeled. Production method and the truth of the accompanying claim are separate questions.

Keep a compact review record

We suggest recording the item name, accompanying claim, earliest available source, tool and check date, result wording, other evidence, reviewer and publication decision. Associate the record with the actual file examined; do not assume a later version is identical.

Start with a sample of incoming materials and measure review time and escalation volume. The goal is a better decision with retrievable evidence, rather than a larger collection of verification badges.

Frequently asked questions

Does a missing SynthID watermark mean an image is real?

No. It means the check did not detect the mark within its capabilities and coverage.

Does the tool establish where or when a video was recorded?

No. Source, location and timing need other evidence.

Sources and editorial review

Original sources used in this article. Last editorial review: 8 October 2026.

  1. Google — Expanding SynthID Detector7 October 2026
  2. Google DeepMind — SynthIDReviewed 8 October 2026
Back to the top