Google Lets Users Remove Visible AI Watermark While SynthID Persists

August 15, 2026news

Google announced on August 14 that it is decoupling visible watermarks from the underlying provenance signals embedded in AI-generated images, videos, and audio. Users can now disable the visible mark while SynthID's invisible watermark and C2PA-compliant metadata remain intact and detectable. For engineers building content authenticity pipelines, the policy clarifies something previously ambiguous: Google treats the visible badge as a UX affordance, not as the canonical provenance record.

Josh Woodward, VP for Gemini, confirmed on X that the toggle applies to the Nano Banana, Omni, and Lyria models. The setting surfaces under Settings > Media Watermark and is rolling out over the coming days across Gemini and Google's video editor Flow, with Search support listed as coming soon.

Two-Layer Architecture

Google's approach separates provenance into two orthogonal channels. The visible watermark is a pixel-level overlay—human-readable but trivially cropped, screenshotted past, or aesthetically disruptive for professional output. SynthID operates differently: it encodes a signal imperceptible to human vision directly into the pixel distribution of an image (or the audio waveform of a song), so the mark survives standard re-export workflows that strip metadata. C2PA metadata travels in the file's manifest and is readable by any C2PA-compliant validator without requiring Google's infrastructure.

Decoupling the visible layer from these two invisible channels is defensible because detection capability doesn't depend on the badge being present. A downstream detector querying SynthID or parsing a C2PA manifest gets the same signal regardless of whether the visible watermark was enabled at generation time.

Signal Type User-Removable Survives Re-export / Screenshot Requires Google Infrastructure to Detect Scope
Visible watermark badge Yes (Settings > Media Watermark) No — cropping or screenshotting removes it No Images, video, audio
SynthID invisible watermark No Yes — embedded in pixel/audio data Yes (or via Credentio locally) Images, video, audio
C2PA metadata No Partial — stripped by some re-encoders No — open standard, any C2PA validator Images, video, audio

The C2PA caveat is significant: C2PA manifests are not pixel-embedded and are routinely stripped by platforms that re-encode on upload. Teams relying on C2PA as their sole verification layer for user-generated content pipelines are accepting a gap that SynthID is specifically designed to close.

Credentio Open-Source Library

The more consequential developer-facing announcement is Credentio, a new open-source library Google is releasing to let developers embed local validation logic directly in their applications. The source material does not specify which runtimes or package registries Credentio targets, nor its license, but the stated purpose is local validation—meaning an app can check provenance without routing a query to a Google-hosted API. For teams already thinking about AI agents interacting directly with content corpora, a locally embeddable validator removes the latency and dependency of a round-trip verification call, and keeps provenance checking viable in offline or privacy-sensitive contexts.

Regulatory Context

The announcement follows Anthropic's move to embed watermarks in Claude-generated text and files to satisfy EU regulatory requirements—a decision that drew debate about whether text watermarking is technically robust enough to matter. Google's approach targets modalities (image, video, audio) where imperceptible steganographic embedding is more mature, sidestepping the text problem entirely. Credentio signals that Google wants third-party developers to build detection into products proactively rather than waiting for platform-level enforcement—a posture consistent with C2PA's ecosystem strategy of distributing validation rather than centralising it. Whether invisible watermarking survives adversarial removal at scale remains the open question no provenance standard has yet answered definitively.