Google’s Gemini Watermark Toggle Doesn’t Touch Your Article 50 Duty
- Google added a “Media Watermark” setting to Gemini and its video editor Flow on 2026-08-14, letting users turn off the visible “sparkle” badge on images and video from the Nano Banana and Omni models, and on Lyria-generated music.
- Turning the visible badge off changes nothing about the invisible SynthID watermark or C2PA metadata Google embeds regardless — on the record from Google’s own VP of Gemini, Josh Woodward: “while the visible watermarks are now optional, invisible SynthID watermarks and C2PA metadata are still being used for transparency.”
- The EU AI Act’s Article 50(4) deployer duty — disclosing AI-generated imagery that “would falsely appear… to be authentic” — was never discharged by Google’s visible badge in the first place. It binds the store publishing the image, not a UI toggle.
- If a WooCommerce or Shopify store’s AI product-photography compliance checklist depended on Google’s visible watermark being on, that checklist was already wrong before this update shipped.
Google gave Gemini and Flow users a switch to turn off the visible AI watermark on 2026-08-14. The mistake this invites: reading “I turned off the sparkle badge” as a compliance decision, in either direction. It isn’t one. The badge Google just made optional was never the mechanism that satisfies the EU AI Act’s transparency duties — the invisible layer underneath it is, and that layer doesn’t move when the toggle does. This piece specifies what actually changed, what didn’t, and where a WooCommerce or Shopify store’s own Article 50(4) disclosure duty sits regardless of which way a shopper’s Gemini account has that setting flipped.
What Google actually shipped on August 14
Google added a new “Media Watermark” setting, reachable from Gemini’s settings menu, that toggles the visible “sparkle” badge Google stamps in the bottom-right corner of AI-generated images, video, and music. Toggling it off applies to everything generated afterward across Gemini and Google’s AI video editor Flow, covering content from the Nano Banana and Omni models — Google’s current image and video generation systems — plus music from the Lyria model, confirmed independently by TechCrunch and The Verge the same day. Google says support for Search is “coming soon” but wasn’t part of the initial rollout.
The change is a genuine UI addition, not a rebrand of something that already existed — before August 14, the visible badge was not user-configurable. But “you can now hide the badge” is the whole scope of what shipped. Nothing in either independent report describes any change to what Google embeds in the file itself, which is the part that matters for a store’s own compliance posture, covered next.
Two marks, only one of them has a switch
Google’s own description of SynthID, its DeepMind-built watermarking system, states plainly that the technology “adds an invisible digital watermark to an AI-generated image (or video segment)” — imperceptible by design, verifiable later through the Gemini app or Google’s SynthID Detector portal. A second, separate mechanism — C2PA metadata, the open provenance standard Adobe, Google, Meta, Microsoft, OpenAI, and TikTok all back — attaches signed origin data to the file separately from SynthID. Neither mechanism is the sparkle badge. The badge was always a visible convenience layer sitting on top of two invisible mechanisms doing the actual marking work.
That ordering is confirmed on the record. Google’s Josh Woodward, VP of Google Labs, Gemini, and AI Studio, told TechCrunch: “while the visible watermarks are now optional, invisible SynthID watermarks and C2PA metadata are still being used for transparency.” Read literally, that sentence says the compliance-relevant layer was already the invisible one before this update — the new toggle only ever controlled a UI element sitting above it.
Where this touches a WooCommerce or Shopify product photo
Stores generating product photography or marketing imagery with Nano Banana through Gemini — one of the model families compared for cost and capability in the site’s Claude vs. GPT vs. Gemini comparison — now have an option their designers will actually notice: a clean product shot with no corner badge, indistinguishable at a glance from a studio photograph. That’s the scenario worth naming precisely, because it’s exactly the kind of synthetic image the EU AI Act’s deepfake clause was written for — not because the toggle creates new risk, but because it removes the one visual signal a shopper might have relied on to guess an image was AI-generated at all. The store’s own disclosure duty, detailed next, was never contingent on that signal existing in the first place.
Article 50(4): the deployer duty the toggle was never about
The EU AI Act defines a “deep fake” in Article 3(60) as “AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful.” A generated product photo that reads as an ordinary studio shot of the actual product fits that definition on its face — it resembles a real object and would falsely appear authentic to a shopper with no reason to suspect otherwise.
Article 50(4) puts the resulting duty on the deployer — the business using the AI system — not the provider: “Deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake, shall disclose that the content has been artificially generated or manipulated.” That’s the store calling Gemini’s API or using the consumer app to build a product image, not Google. Google’s badge, on or off, was never capable of discharging a duty the regulation assigns to someone else. This is the same provider/deployer split covered from Anthropic’s side in the site’s Claude watermark compliance check, applied here to a different vendor and a different content type.
On, off, or never checked — your duty doesn’t move
The table below separates what a shopper can see from what’s actually embedded, and from what the store still owes under Article 50(4), across every state Google’s toggle can be in.
| Toggle state | What a shopper sees | What’s embedded in the file | Your Art. 50(4) duty |
|---|---|---|---|
| Visible watermark ON (default) | Sparkle badge, corner of image | SynthID + C2PA (invisible) | Still yours, if the image is presented as authentic |
| Visible watermark OFF (new toggle) | No badge — looks unmarked | SynthID + C2PA (invisible, unchanged) | Still yours — nothing about the duty changed |
| No AI involved (real product photography) | No badge, none expected | None | No Art. 50(4) duty — not synthetic content |
Verdict: the only row where a store’s disclosure duty actually changes is the one where AI wasn’t used at all. Both AI rows carry the same duty, because both rows carry the same invisible marking — the visible badge was never the variable that mattered.
This isn’t a Google-specific problem
The same shape showed up when Anthropic began watermarking Claude-generated text and images in August 2026: a provider’s own marking mechanism discharges the provider’s Article 50(2) obligation, never a deployer’s separate duty under 50(1) or 50(4). Google’s toggle is the same pattern one layer down — it doesn’t even touch the provider mechanism, only a cosmetic setting sitting on top of it. Any AI vendor could ship an identical visible on/off switch tomorrow and it would change nothing about a store’s own disclosure obligations, because the switch was never the compliance mechanism to begin with. This is the recurring shape of every AI-transparency story in the site’s store-operations compliance guide — provider marks and deployer duties are separate obligations, and no vendor UI setting closes that gap.
The compliance check
The following applies regardless of how any individual staff member’s Gemini account has the Media Watermark setting configured — it’s a store-level policy question, not a per-user setting:
- Confirm whether any AI-generated product imagery on your storefront is presented as authentic photography — if so, Article 50(4) requires your own visible disclosure.
- Add that disclosure at the point of publishing, not conditionally on whether the generating tool’s own visible watermark happens to be on.
- Don’t treat a visible AI badge from any vendor’s tool as satisfying your store’s deployer duty — it discharges the vendor’s provider duty, not yours.
- Don’t treat the absence of a visible badge as evidence an image wasn’t AI-generated — Google’s own invisible marking (and most competitors’) persists either way.
- Check whether staff generating product images know the toggle exists — a designer switching it off to get a “cleaner” shot doesn’t remove your disclosure obligation, and shouldn’t be treated as if it does.
- Review the fuller two-item deployer checklist in the site’s small-business AI policy guide if AI-generated imagery is a regular part of your catalog workflow.
- Confirm the specifics with counsel for your jurisdiction — this is a technical and regulatory explainer, not legal advice.
What Google hasn’t shipped yet
The toggle is rolling out gradually — Google’s own statement to TechCrunch describes it as arriving “in the coming days,” not universally live at announcement. Search support is explicitly “coming soon,” meaning image generation surfaced through Google Search doesn’t have the setting yet even where Gemini and Flow do. A store auditing its own AI-image pipeline should check whether the specific surface it uses has the toggle at all before assuming either watermark state.
Also unchanged: there is still no self-serve way for a store to check, in bulk, which of its own previously-generated images carry the SynthID mark versus which predate Google applying it consistently. Verification runs one file at a time through the Gemini app or the SynthID Detector portal — useful for a spot check, not yet a pipeline-scale audit tool. None of this is legal advice; confirm the application to your specific content and jurisdiction with counsel before treating any item above as a compliance conclusion.
Auditing what AI touches in your product catalog?
A services conversation can map exactly where AI-generated imagery, copy, and chat sit in your WooCommerce, Shopify, or NetSuite-connected stack — and what, if anything, needs a disclosure line.
Get the working checklists
The runbooks and decision checklists from these guides, as printable PDFs — free in the SoftXone guide library.
References
- TechCrunch: Google will now allow users to remove visible watermark from its AI generationsFirst report, including Josh Woodward’s on-record statement and the model/rollout details.
- The Verge: You can now turn off Google Gemini’s visible watermarksIndependent same-day corroboration, including Woodward’s title and the Nano Banana/Omni model detail.
- Google DeepMind: SynthIDPrimary source for how the invisible watermark mechanism works and how it’s verified.
- EU AI Act — Article 50: Transparency obligations for providers and deployersFull article text, including paragraph 4 covered in this post.
- EU AI Act — Article 3: DefinitionsSource for the Article 3(60) “deep fake” definition applied here to product imagery.
- The Verge: What is the C2PA standard?Background on the second, separate provenance-metadata mechanism named in this post.
Frequently asked questions
Does turning off Google’s visible AI watermark remove the invisible SynthID mark too?
No. Google’s Josh Woodward, VP of Gemini, confirmed on the record that the invisible SynthID watermark and C2PA metadata stay embedded regardless of the visible “Media Watermark” toggle — the setting only controls the corner badge a shopper can see, not the mechanism Google actually relies on for transparency.
Do I need to disclose AI-generated product photos under the EU AI Act if I sell into the EU?
Yes, if the image is presented as authentic photography. Article 50(4)’s deep fake clause covers AI-generated or manipulated imagery that “would falsely appear… to be authentic” under the Act’s own Article 3(60) definition — that is a deployer duty on the store publishing the image, not something any vendor’s watermark setting satisfies on your behalf.
Which Google products does the new watermark toggle apply to?
Gemini and Google’s video editor Flow, covering the Nano Banana and Omni models for images and video, and Lyria for music. Google says support for Search is coming but was not part of the August 14, 2026 initial rollout.
Does Anthropic’s Claude watermark work the same way as Google’s?
The mechanism differs, but the compliance shape is identical: both are provider-side marking under EU AI Act Article 50(2), and neither discharges a store’s separate deployer duty under Article 50(4). The site’s Claude watermark compliance check covers the text and image specifics for that vendor.
Leave a Reply