On October 7, 2026, Google announced global English access to SynthID Detector for image, video and audio checks. Newsrooms, educators and communications teams can now use the public portal to look for supported AI watermarks, expanding beyond the early-access tool introduced for media professionals and researchers in May 2025.
Use it to investigate how media was made, not to certify that a file is real. The portal’s FAQ says a missing watermark does not rule out AI generation or editing. A positive result is evidence to investigate alongside the file’s history and source.
Choose the right place to run a check
Google names OpenAI, NVIDIA and Kakao alongside its own models as supported partners; Apple detection is still forthcoming. The launch announcement does not establish coverage for every model, product version or historical output from those companies.
The distinction between Google’s verification products matters: they do not all document the same coverage or access conditions.
Service | Documented scope | Practical choice |
|---|---|---|
Supported SynthID watermarks in images, video and audio from Google and participating partners. | Occasional manual media review; not an automated bulk service. | |
Its help page currently limits SynthID recognition to Google AI content. Content Credentials verification is a separate capability. | Do not treat a negative Gemini watermark check as a check across all partners. | |
Separate private-preview service for JPEG, PNG and WebP images; access requires an application. | Assess separately for integration. It is not documented as an equivalent API for the public three-media portal. |
The public portal’s supported-content list does not include text. Readers evaluating essays or chatbot output face a different access and detection question, covered in our guide to text-watermarking controls and detector limits.
Engadget reports that the portal is free and requires a Google, OpenAI or Apple sign-in. Apple sign-in does not mean Apple-generated media is already detectable. Exact portal quotas, reset rules and file-size or duration ceilings were not established in the primary documentation reviewed; do not substitute Gemini’s limits.
Keep three review outcomes
SynthID looks for a signal intentionally embedded by a participating model. According to the current FAQ, detection suggests generation or editing by a SynthID-enabled model. It also acknowledges missed detections and rare false triggers, without establishing a numerical error rate for this public deployment.
For a review log, we recommend three outcomes. These are workflow categories, not claimed portal response fields:
Recognized watermark: Preserve the exact result and any displayed scope. Ask the supplier what was generated or edited before deciding whether a disclosure is needed. The signal alone does not establish that the whole file is synthetic or that someone misrepresented it.
No recognized watermark: Leave origin unresolved. The FAQ identifies unsupported models, output predating watermark adoption and heavy modification as possible reasons AI media can go undetected. Continue source and history checks.
Not checked: Record why a check could not be completed—such as an unsupported file, exhausted quota, service failure or lack of permission to upload. Do not collapse this into a negative result.
Our analysis: coverage has three separate limits. A generator must have added a supported watermark, enough of it must survive subsequent handling, and the detector must recognize it. Strong detection on watermarked examples cannot establish how much of all AI media the service will catch. This reporting provides no independently measured portal accuracy rate.
Add Content Credentials and source context
C2PA Content Credentials can supply signed, asset-bound statements about creation and editing history. Their validation concerns the integrity and association of those statements—not whether a depicted event or accompanying claim is factually true. That history can help explain what a watermark result alone cannot.
The two approaches are complementary. The C2PA explainer notes that metadata can be removed, and that durable credentials can use watermarks or fingerprints to rediscover it. This does not mean every SynthID result retrieves a C2PA history. Preserve the original file and inspect available credentials separately.
A visible AI label is a disclosure to the audience, not the invisible signal itself. Our recommendation is to decide labels using the combined evidence and the relevant publication policy. For example, evidence of a permitted AI edit calls for understanding that edit, not automatically treating the entire submission as fabricated.
Check upload permission; keep automation separate
The portal’s privacy notice says uploaded originals are deleted when results are returned. It retains a media signature linked to the user identifier for 24 hours for anti-abuse purposes. It also retains basic third-party sign-in identifiers, without specifying the same lifetime for those account details.
That is not a zero-retention promise. Before submitting confidential or unpublished material, check whether your organization permits this upload and whether the notice answers its retention and handling requirements.
The portal terms prohibit using the frontend for programmatic bulk verification and circumventing quotas. For product-scale use, seek separately authorized API access and check its actual media coverage. Google Cloud’s private-preview image API documentation describes probabilistic image analysis, not a universal watermark endpoint, and requires results to supplement rather than determine critical decisions.
A minimal review procedure
Confirm upload permission and obtain the highest-quality available copy, as Google recommends. Keep the original and note if the checked file is a transformed copy.
Record the exact service, check time, result text and any indicated region or segment. Associate the record with the file—using a file hash where practical—so a later reviewer knows which version was checked.
Compare the result with available Content Credentials, supplier records and the claimed source. Keep disagreements visible rather than forcing a binary real-or-fake verdict.
Record the publication or disclosure decision separately from the detector outcome. If provenance remains unresolved, say so.
SynthID Detector is useful when a supported watermark would add evidence to a manual review. It is not sufficient when the task demands proof of human authorship, factual truth or unrestricted automated verification.
Methodology: AI-assisted reporting and analysis of Google’s announcement, public portal documentation, Google support and Cloud documentation, the C2PA explainer and corroborating reporting, checked October 7, 2026. No media was uploaded; detector performance, account eligibility and API behavior were not tested.