An AI file can carry the SynthID watermark without falling inside Gemini’s documented recognition scope. Superpower Daily’s public-document review finds a consequential split: Gemini recognizes Google AI watermarks, while Google’s standalone Detector names outside partners. That partner list, however, does not certify every product workflow—and some named integrations remain forthcoming or unresolved.
Original analysis: We compared supplied official announcements, verification help, product documentation and a model card for four external partners: OpenAI, NVIDIA, Kakao and Apple. Google names these partners, with Apple forthcoming. Our question was which products and media types have documented watermark generation and detection support. We separated insertion, rollout status and checker compatibility rather than treating adoption as a single yes-or-no answer.
The review retains an October 11, 2026, 20:31:04.976 UTC evidence cutoff. Saved snapshots were retrieved afterward, so we cannot certify that mutable documentation had exactly this wording at the cutoff. We uploaded no media and measured no detection accuracy, robustness or false-negative rate. These are documentation findings, not experimental results. OpenAI itself cautions that coverage varies by product, model, export path, file type and creation time.
A hidden signal needs a compatible checker
SynthID embeds an imperceptible signal directly into generated media. Unlike metadata—information attached to a file—the watermark is part of the content itself. OpenAI describes it as an additional clue about origin that may survive some edits or transformations, even when attached metadata disappears. It is not a judgment based on whether an image looks artificial.
The two approaches carry different information. Content Credentials can record a file’s origin and history, but screenshots and conversions may remove them. Embedded watermarks carry less context, while potentially surviving transformations that strip metadata. OpenAI combines both for supported images. Neither technique guarantees survival through every edit; extensive cropping, compression or other changes can make watermarks undetectable.
Generation and recognition are separate promises. Gemini’s verification help explicitly says its SynthID checks currently recognize only content created by Google AI tools, despite adoption elsewhere. Gemini can also inspect compatible Content Credentials from other sources. The Google-only restriction therefore applies to its SynthID recognition, not to every piece of origin information Gemini can display.
Google’s October 7 announcement opened the standalone Detector globally in English. OpenAI’s own verifier and Content Provenance API take a narrower approach: they check supported images and audio for OpenAI-associated signals. They are not designed to detect other companies’ generated content. Sharing the SynthID name does not make these verification surfaces interchangeable.
The partner evidence, product by product
The records are strongest where a provider specifies what its generator inserts into returned files. They become less definite where a partner announcement must stand in for product-level detection documentation. The following evidence matrix preserves that distinction; unresolved means the sampled records do not settle the question, not that a capability is absent.
SynthID partner evidence matrix
- OpenAI — Insertion: documented for images generated through ChatGPT, Codex and the OpenAI API, alongside Content Credentials. ChatGPT and API audio generation also insert inaudible SynthID watermarks. Detection: Google names OpenAI at partner level but does not enumerate these individual paths or audio model IDs. OpenAI’s own tools check supported images and audio. Coverage caveat: its general API provenance statement excepts gpt-image-1, gpt-image-1.5 and sora-2; that is not an exhaustive SynthID-specific exclusion list. Codex audio remains unresolved in this sample.
- NVIDIA — Insertion: the hosted Cosmos3 Nano endpoint explicitly watermarks returned images and videos. Watermarking is mandatory, with no request-level switch to disable it. Numeric action arrays are explicitly not visually watermarked. Detection: Google names NVIDIA, making compatibility an inference from that announcement and documented insertion, rather than a file-level result from our desk. Deployment boundary: the hosted model card does not certify every downloadable Cosmos model or local deployment. Its instructions recommend preserving original JPEG or MP4 files because re-encoding, resizing and editing may affect detection.
- Kakao — Insertion and rollout: its May 20 announcement schedules SynthID integration beginning in the second half of 2026 for Kanana-Kollage images and Kanana-Kinema videos, starting with KakaoTalk’s Kanana Template video feature. Detection: Google’s October announcement names Kakao as supported. Unresolved: the sampled partner records do not establish which named product integrations completed rollout. Kakao’s existing visible Kanana logo is a different signal; its announcement describes invisible SynthID watermarking as an additional planned measure, so the logo cannot confirm that insertion has happened.
- Apple — Insertion: announced for Image Playground generation and Apple Intelligence editing in Photos. The named editing tools include Spatial Reframing, Extend and Clean Up. Rollout: September 14 documentation still schedules SynthID for a later-2026 software update, rather than treating it as part of the software already available. Edited-image coverage is qualified as most images, depending on the edits applied. Detection: Google likewise describes Apple support as forthcoming. These records do not establish an exact release date or eligibility for each possible edit.
Text does not follow the media portal
SynthID’s broader technology includes text watermarking. Google describes adjusting the probabilities of word choices to create a pattern that is not noticeable to readers. But the expanded public Detector documents uploads for images, video and audio—not text. The ability to create a watermark in one medium does not establish public detection access for that medium.
OpenAI’s text signal is a separate system, textGrain. Its documentation lists Content Credentials for video, not SynthID. Neither entry belongs in a confirmed SynthID text-or-video coverage row. OpenAI also limits text detector applications to qualifying organizations, with academic-integrity checks outside the stated program scope. That access policy is separate from its public image-and-audio verifier.
A missing watermark is not a human-origin certificate
A negative result cannot establish human origin. OpenAI says a missing signal can reflect older content, unsupported products or export paths, stripped metadata, degraded watermarks, or audio that is too short or substantially modified. Its advice is to use an original exported file when possible and consider the source and context even when nothing is detected.
Gemini’s help contains a tension worth retaining. Its simplified explanation says no detected watermark means an image or video was not created or edited by Google AI. Elsewhere, it allows inconclusive results for sparse detail or tiny edits and acknowledges possible nondetection after repeated alterations. Those qualifications prevent treating the simplified line as an absolute assurance. This guidance concerns Gemini and OpenAI verification; it is not a verified result taxonomy for the standalone Detector.
A positive signal also has limits. OpenAI says detection does not establish accuracy, ownership, creator identity or correct context. A watermark does not show how much a person contributed, or whether the model generated all the content rather than editing an upload. Verification supplies evidence about origin, not a universal authenticity verdict.
Reader comments
Newest comments first. Replies stay oldest first.