Google’s AI Watermarking Bet Pays Off as EU Delays Rules

Google signed the European Union’s Code of Practice on the Transparency of AI-Generated Content just hours before the window closed for companies to appear on the initial list of signatories, submitted before the 27 July, 18:00 CEST cutoff Brussels had set. The signing extends a commitment Google first made in 2025, when it backed the AI Act’s General-Purpose AI Code of Practice.

It lands five days before the EU’s AI Act begins enforcing its own transparency rules on 2 August. And it arrives with a warning attached: Google told regulators that stacking watermarks, labels and legal notices on the same piece of content could confuse the very users the rule is meant to protect.

Google Signs On With Hours to Spare

The code covers how companies identify content that AI created or altered, whether that is a synthetic photo, a cloned voice or an edited video. Google’s compliance plan leans on two tools. One is the Coalition for Content Provenance and Authenticity standard, known as C2PA, which attaches a traceable edit history to a file. The other is SynthID, Google’s own watermarking system that embeds an invisible signal directly into pixels or audio waveforms at the moment content is generated.

Google is not new to Brussels’ AI rulebook. The company has spent years working through the bloc’s other digital statutes too, including a billion-dollar EU antitrust fine tied to Android, so this signature reads less like a first encounter and more like a company that has learned to move early on European rules rather than fight them after the fact.

Five Companies, One Watermark

Google says it is not building this alone. It has been working with five other AI developers to make watermark detection interoperable across platforms, so a file tagged by one company’s tool can still be read by another’s.

OpenAI has described its own participation as support for a trustworthy AI ecosystem in Europe, a phrase that mirrors Google’s own framing almost exactly. Five companies agreeing on one detection layer is a bigger deal than it sounds. It means a watermark planted by a Google model can, in theory, be read by an Apple device or an ElevenLabs tool without either company building its own separate detector from scratch.

Could More AI Labels Confuse Users?

Yes, according to Google itself. The company told regulators that piling a SynthID tag, a C2PA credential and a separate legal disclosure notice onto one photo or clip risks burying the signal it is meant to send, leaving users less certain, not more, about how a piece of content was made.

That is an unusual position for a company to take in public while still signing the underlying commitment. Google is not asking Brussels to drop the transparency rule. It is asking Brussels to slow down on adding new label formats until the technical standards, including C2PA and SynthID’s own detection tools, have time to settle into something consistent. Google has separately leaned on identity signals to fight synthetic fraud elsewhere too, rolling out selfie video sign-in checks as deepfake fraud has surged, which suggests the company sees label fatigue and verification fatigue as the same underlying problem.

Brussels Buys Itself Four More Months

Google’s warning about rushing regulation landed at an interesting moment. The EU’s own institutions had already agreed, weeks earlier, that part of this exact obligation needed more time.

1. July 2025: Google and OpenAI sign the AI Act’s General-Purpose AI Code of Practice; Meta declines.
2. November 2025: The European Commission proposes a Digital Omnibus package to simplify AI Act timelines, including a six-month delay to marking rules.
3. February 2026: The Commission publishes its first draft Code of Practice on Transparency of AI-Generated Content.
4. 26 March 2026: The European Parliament adopts a negotiating mandate seeking a shorter, three-month delay instead.
5. 7 May 2026: The Council and Parliament reach a provisional deal settling on a four-month extension.
6. June 2026: The AI Office publishes the final Code of Practice on Transparency of AI-Generated Content.
7. 27 July 2026: Google and its partners submit signatures ahead of the initial signatory-list cutoff.
8. 2 August 2026: Most AI Act transparency duties, including deployer disclosure of deepfakes, enter into application.
9. 2 December 2026: The machine-readable marking obligation under Article 50(2) now takes effect instead of 2 August.

The Council and Parliament agreement to simplify and streamline the rules did not touch every transparency duty. It targeted specifically the piece Google flagged as premature: the requirement that providers mark synthetic output in a machine-readable format detectable by outside tools. Regulators effectively conceded the same point Google made publicly, just in legislative form rather than in a policy statement. Companies that keep skipping the obligations entirely still face penalties under the AI Act’s enforcement regime, including fines that can reach €35 million for the most serious violations, so the delay buys engineering time, not an exemption.

Who Signed, Who Refused

Not every major AI developer has taken Google’s path. Meta walked away from the AI Act’s general-purpose code back in 2025, and it has kept its distance from the follow-on transparency code too, at least going by the partner list Google itself named.

Meta’s chief global affairs officer said at the time that the code “introduces a number of legal uncertainties for model developers, as well as measures which go far beyond the scope of the AI Act.”

• Signed or supporting: Google, OpenAI, Apple, ElevenLabs, Kakao and NVIDIA, through the watermarking coalition Google announced.
• Likely to sign, per its own president: Microsoft, which Brad Smith told Reuters would probably join.
• Declined: Meta, citing legal uncertainty over the 2025 general-purpose code.

The split matters because voluntary codes only work as industry standards if enough of the industry actually shows up. Six companies representing most consumer-facing generative AI tools is a meaningfully different starting point than a code two or three companies signed alone.

SynthID Has Already Tagged 100 Billion Files

While regulators were still negotiating the four-month delay, Google’s own tool had already scaled past anything the law was built to require. The company says SynthID has tagged more than 100 billion images and videos, plus roughly 60,000 years worth of synthesized audio, across Gemini, Imagen, Veo and Lyria output.

Verification checks run through Gemini have been used 50 million times globally, and Google is now rolling SynthID detection into Search, Chrome and Circle to Search on Android rather than keeping it confined to its own generative tools. That scale is the real substance behind Google’s bet. A voluntary watermark that already runs across a hundred billion files does not need a law to make it the practical standard. It just needs the law to stay out of its way long enough to finish rolling out, which is effectively what the December delay now gives it.

Frequently Asked Questions

What Is the EU’s Code of Practice on Transparency of AI-Generated Content?

It is a voluntary compliance tool published by the EU’s AI Office in June 2026 that translates the AI Act’s Article 50 marking and disclosure duties into concrete technical steps. Signing it is optional, but the Commission treats it as an accepted way for a company to demonstrate it is meeting the underlying legal requirement.

Is Labeling AI Content in the EU Legally Required, or Voluntary?

The code itself is voluntary, but the obligation behind it is not. Article 50 of the AI Act legally requires machine-readable marking from providers and clear disclosure from deployers once the law applies to them; the code just gives companies a Commission-endorsed template for proving they have done it.

What Is SynthID and How Does It Mark Content?

SynthID embeds a statistical pattern into an image, video frame or audio waveform at the moment the content is generated. The pattern is invisible to people but survives ordinary edits like cropping, resizing or compression, and a separate detector tool can scan a file afterward to report the odds it carries a SynthID signal.

Does the EU AI Act Apply to AI Content Made Outside Europe?

Yes, if that content reaches people in the EU. Like most of the bloc’s digital rulebook, the AI Act’s transparency duties attach to where content is placed on the market or shown to users, not where the company generating it is based, so a video made outside Europe still needs a compliant disclosure once it reaches EU audiences.

Did Meta Sign the AI Act’s Transparency Code Too?

Meta is not among the partners Google named in its coalition. The company declined to sign the AI Act’s earlier general-purpose code in 2025 over concerns about legal uncertainty, and there is no public sign yet that it has joined this follow-on transparency code either.

Can Someone See a SynthID Watermark Just by Looking at an Image?

No. The mark is invisible by design and only becomes readable through a matching detector, such as the SynthID Detector tool Google has connected to Search, Chrome and Gemini, which reports the statistical likelihood that a file was AI-generated or edited.