Gemini’s documented SynthID checks are Google-only; the standalone detector names partners

Measurement technique

How to read this report

1. 01Synthesize the supplied official announcements, product documentation, verification help and model card without new collection, uploads or experiments. Recovery of saved evidence does not reset the evidence cutoff.
2. 02Bound the external-partner sample to OpenAI, NVIDIA, Kakao and Apple. Do not extend the findings to every SynthID adopter, Google product, downloadable model or export path.
3. 03Plan an evidence_matrix with separate columns for partner, named product or workflow, modality, watermark insertion, rollout status, standalone Detector compatibility, verification surface, source and uncertainty. Use documented, announced/forthcoming and unresolved labels; distinguish explicit product compatibility from inference based on insertion plus a partner announcement.
4. 04OpenAI matrix rows: ChatGPT, Codex and API images have documented SynthID insertion; ChatGPT and API audio have documented insertion. Google names OpenAI but does not enumerate these individual product paths. Codex audio remains unresolved. Preserve the API provenance exceptions for gpt-image-1, gpt-image-1.5 and sora-2. TextGrain for text and Content Credentials for video are not confirmed SynthID coverage.
5. 05NVIDIA matrix rows: the hosted Cosmos3 Nano endpoint documents mandatory SynthID insertion into returned images and videos, without a request-level disable option. Numeric action arrays are explicitly not visually watermarked. Standalone Detector compatibility is inferred from insertion and Google’s NVIDIA announcement, not file-level testing; local deployments remain unresolved.
6. 06Kakao matrix rows: Kanana-Kollage images, Kanana-Kinema videos and KakaoTalk Kanana Template videos are announced integrations. The sampled records do not establish completed product rollouts, despite Google naming Kakao as a Detector partner. The visible Kanana logo is not confirmation of invisible SynthID insertion.
7. 07Apple matrix rows: Image Playground generation and Apple Intelligence editing in Photos are forthcoming integrations in the cited records. Named editing tools include Spatial Reframing, Extend and Clean Up. September documentation limits edited-image inclusion to most images, depending on edits; Google also describes Apple Detector compatibility as forthcoming.
8. 08Keep result interpretation separate from coverage. OpenAI and Gemini document reasons for missing or inconclusive signals, but their language is not a verified result taxonomy for the standalone Detector. The public portal documents image, video and audio uploads, not text.