Google opens SynthID Detector to the public: which AI images, videos, and audio can it check?

What changed from the earlier waitlist?

Google's Portuguese official blog published the announcement October 7, 2026, by Pushmeet Kohli, Google DeepMind Vice President of Science and Strategic Initiatives. It says SynthID Detector is available globally in English to everyone from that day, without an exact time or time zone.

Previously, testing primarily served journalists and media professionals. The current DeepMind page confirms expansion after testing and feedback beginning in 2025. Old search snippets mentioning a waitlist therefore do not establish current access.

Teams reviewing images, short videos, or voiceovers now have a public verification entry point rather than only a note inside a generation tool. Identify the official link and supported material instead of judging whether media merely looks AI-generated.

Three media types do not mean every AI output

The announcement names images, video, and audio and AI content from Google and partners OpenAI, NVIDIA, and Kakao. Apple is described as joining soon, not supported already.

Partner names do not guarantee exhaustive detection. The announcement does not list every company's model, version, or historical output. A detectable mark depends on generation or editing using SynthID and the mark surviving detectably.

Two images from one creative team can come from different tools and edits. Finding a watermark in one cannot establish that the other is a photograph or that every asset from the team is covered.

Does it detect a watermark or establish whether a scene is true?

SynthID embeds a digital watermark during generation that people cannot readily see or hear, then detects that mark. The image or video mark is not a visible corner logo; an audio mark is not an audible notification.

Google says the design accounts for common cropping, filters, and compression, but cannot reliably withstand every extreme modification. The 2023 explanation establishes technical boundaries; current access comes from the new announcement and interface.

No detected watermark means that mark was not found, not that the content must be human-made. Finding an AI mark does not prove depicted events, product results, or speech are truthful. Provenance and factual verification answer different questions.

Start with the official entry and current conditions

The announcement and DeepMind product page link to the official Detector entry retained in the sources. Enter through those sources and inspect current capabilities and conditions rather than assuming a similarly named AI-detection site is the same product.

The announcement does not fully specify login, formats, size, usage counts, or prices, and no upload was tested here. “For everyone” does not establish no login, permanent free use, or unlimited checks. Verify conditions at the current entry first.

Confirm authority to process and provide the file, and retain its original source and a version not repeatedly re-exported. For private client media, unpublished projects, or sensitive material, understand data-handling conditions before uploading.

How can sellers and creators use the result?

Fictional example: a team receives a supplier's product video and wants to check AI generation or editing. A watermark can be one provenance clue, alongside the supplier's production explanation, original assets, and licensing. This is not an actual detection result.

If no watermark appears, continue checking provenance rather than labeling it real footage. If a mark appears, do not classify every frame as generated without considering the detector's actual scope and the asset's explanation.

Published media must still meet the destination platform's labeling and use rules. Detection can reveal questions; it does not grant commercial rights or guarantee advertising, product-display, or account compliance.

Watch coverage conditions rather than a universal truth score

The news is expanded public access and three-media scope. Useful future details include file requirements, account conditions, partner-model lists, and Apple's actual arrival. Unknown items should not be reported as delivered.

Combine detection with provenance records, original files, and editing explanations. Do not reduce one result to “100% real or fake” or replace review of the production chain. Understanding what a result establishes is the tool's value.