AI Policy

Google's Watermark Opt-Out Tests the EU AI Act

Google let Gemini users hide the visible AI watermark on images and video from August 14, twelve days after the EU AI Act's Article 50 transparency mandate became enforceable with fines up to EUR 15 million.

AI PolicyBy Servola Tech Desk2026-08-143 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

Google's Watermark Opt-Out Tests the EU AI Act

Key takeaways

On August 14, 2026, Gemini VP Josh Woodward said in a post on X that Google will let users switch off the visible watermark on Nano Banana images, Omni video and Lyria music, rolling out to Gemini and Flow in the coming days, with Search support to follow; removing the corner mark was one of the top ten requests after Woodward asked users for their biggest Gemini frustrations.
The toggle arrives twelve days after the European Commission's AI Office gained the power, from August 2, 2026, to fine general-purpose AI providers up to EUR 15 million or 3 percent of global annual turnover for breaching the AI Act's Article 50 transparency duties, which require synthetic image, video, audio and text content to carry a machine-readable AI mark.
Google says the invisible SynthID watermark and C2PA content-credential metadata stay embedded regardless of the visible-watermark setting, so the toggle only removes the mark a human eye can see, not the machine-readable trail the law names.
The opt-out is not universal: Gemini work and school accounts get no toggle at all, and in India, South Korea and Vietnam it sits behind the paid AI Ultra subscription tier, meaning three different compliance postures ship from the same company in the same week.
No general-purpose AI provider has been fined under the Article 50 powers since they went live, so Google's wager that invisible-only marking satisfies a 'machine-readable' requirement is, for now, untested by any published enforcement decision.

A Feature Request Granted Two Weeks After a Deadline

Google's Gemini team turned a user-experience complaint into a settled feature almost overnight. Gemini VP Josh Woodward had asked his followers on X for their biggest frustrations with the product; removing the visible corner watermark on Nano Banana images, Omni-generated video and Lyria music tracks came back as one of the top ten answers. On August 14, 2026, Woodward announced that Gemini and Google's Flow video editor will ship a Settings > Media Watermark toggle within days, letting a user turn the visible mark on or off for every future generation. Search support is described as coming later.

The timing is what turns a minor UX fix into a policy story. Twelve days earlier, the enforcement machinery behind the EU AI Act's transparency rules had switched on for the first time, with the European Commission's AI Office newly empowered to investigate and fine the exact category of company Google is.

What Article 50 Actually Asks For

The EU AI Act's Article 50 transparency obligations became enforceable on August 2, 2026, according to the European Commission's AI Office, which can now request technical documentation from general-purpose AI providers, evaluate their models, demand risk mitigation, and issue fines of up to EUR 15 million or 3 percent of global annual turnover, whichever is higher. The rule's core demand is that AI-generated or manipulated image, video, audio and text content be marked in a way that is machine-readable and detectable as artificial, and that deepfakes in particular carry a clear label.

Nothing in the AI Office's own published materials specifies that the mark has to be visible to a human. That silence is the gap Google's product decision now sits inside.

The Bet Behind the Toggle

Google's position, stated plainly by Woodward, is that SynthID's invisible digital watermark and the C2PA content-credential metadata travel with a file regardless of what the visible-watermark setting is doing. Both are machine-readable in the literal sense: a detector tool or a C2PA-aware application can read them even though a person looking at the picture cannot see anything. That is a defensible reading of Article 50's text. It is also a different product than what most people who followed the AI Act's public rollout were told to expect, which was closer to a visible label they could rely on with their own eyes.

The practical effect is to move the burden of detecting AI content from the viewer to whichever software the viewer happens to be using - and most consumer apps, social feeds and messaging clients do not run a SynthID detector or read C2PA fields today.

What an EU Business Publishing AI Media Should Check This Week

Any EU organization that uses Gemini, Nano Banana or Flow output in marketing, product photography or social content now has a decision to make once the toggle reaches its account, and a second, less obvious one to verify: whether its own publishing pipeline preserves C2PA metadata at all. Many content management systems, social schedulers and image compressors strip embedded metadata on upload as a routine optimization step - which means that once the visible watermark is switched off, an image can leave a company's own systems carrying no detectable AI-origin signal whatsoever, invisible watermark included, even though Google's file did carry one when it left Gemini.

Frequently asked questions

What exactly changed in Gemini's watermarking on August 14, 2026?

Google said a new Settings > Media Watermark toggle is rolling out to Gemini and its Flow video editor within days, letting users switch the visible watermark on Nano Banana images, Omni video and Lyria music on or off for future generations. Search support comes later.

Does turning off the visible watermark break the EU AI Act?

Google's position is no, because the invisible SynthID watermark and C2PA metadata remain embedded either way, and it reads Article 50's machine-readable requirement as satisfied by those alone. No enforcement decision has tested that reading yet.

Why can't every Gemini user turn the watermark off?

Work and school Gemini accounts get no toggle at all. In India, South Korea and Vietnam, the option is restricted to the paid AI Ultra subscription tier; free users there keep the mandatory visible watermark.

The AI Act's transparency rules were sold publicly as a way for an ordinary person to tell, by looking, that a picture or a clip was made by a machine. What Google shipped on August 14 keeps the letter of that promise and quietly drops the part people actually understood: the mark is still there, but only software can see it now, and most of the software people actually use to view a shared image was never built to look. Enforcement has been live for twelve days. Nobody has tested in public yet whether a regulator agrees that invisible counts the same as visible.

GoogleGeminiEU AI ActArticle 50SynthIDEU AI Office

Label Your AI Text or Give It an Editor

The Commission published the final Article 50 transparency guidelines on 20 July 2026. From 2 August, a deployer who publishes AI-generated text on a matter of public interest without human editorial review has to disclose it.

4 min read

Decisions

800,000 Reservations Did Not Buy A Launch

Google cancelled its AI Studio mobile app on 31 July with about 800,000 preorders logged. The reason it gave was not demand and not quality. It was that it did not want you to install a second app.

5 min read