Google said on July 24, 2026 that it is signing the European Union's Code of Practice on Transparency of AI-Generated Content, adding the largest supplier of generative advertising tools to a voluntary framework that becomes commercially relevant nine days later, when Article 50 of the AI Act starts to bite.

The announcement came from Karen Massin, Head Government Affairs and Public Policy - EU Institutions at Google, in a post on the company's News from Google site. It was short. It was also, in places, openly critical of the regime the company was joining.

According to Google, the company is signing the code "to help support the responsible use of AI in Europe," and the move builds on its 2025 signing of the General-Purpose AI Code of Practice as well as its work on the C2PA provenance standard and its own watermarking system. Two paragraphs later, the same post warns that layering further obligations onto an unsettled technical field carries costs. According to Google, adding more regulatory complexity while technical solutions are still evolving "could contradict Europe's goals for competitiveness and simplification." The company added that if online content is flooded with overlapping AI labels and legal disclosures, it becomes harder for people to get the clear context they need.

Both positions can be held at once. What matters commercially is the second half of the sentence: Google has now accepted a specific set of technical measures that its advertising, video and search products will be measured against.

What the timing means

The European Commission published the finalised code and its accompanying Article 50 guidelines on 20 July 2026, four days before Google's post. PPC Land reported at the time that organisations seeking a presumption of compliance through the code were expected to sign by 22 July 2026, and that non-compliance with Article 50 carries administrative fines of up to EUR 15,000,000 or, for an undertaking, up to 3 percent of total worldwide annual turnover for the preceding financial year, whichever is higher.

Google's announcement therefore lands two days after that date and nine days before the obligations themselves take effect on 2 August 2026. Whether the signature is processed as part of the initial signatory cohort or as a later addition is a procedural question the Commission has not publicly resolved. The substantive point is unaffected: the underlying legal duties apply from 2 August regardless of who signs, and the code exists to give companies a recognised way of showing they have met them.

According to the Commission, the code was drawn up by independent experts through a multi-stakeholder process facilitated by the AI Office, and the Commission and the AI Board have confirmed it is an adequate voluntary tool to demonstrate compliance with the AI Act transparency obligations. Providers and deployers that comply through other means have to demonstrate that those measures are adequate, assessed individually by different market surveillance authorities.

Two sections, two sets of duty-holders

The document runs to 38 pages and splits cleanly in two.

Section 1 covers providers of generative AI systems under Article 50(2) and (5), and was chaired by Kalina Bontcheva with Dino Pedreschi and Christian Riess as vice-chairs. It governs machine-readable marking and detection. Section 2 covers deployers under Article 50(4) and (5), chaired by Anja Bechmann with Giovanni De Gregorio and Madalina Botan as vice-chairs. It governs the visible labelling of deep fakes and of AI-generated text published to inform the public on matters of public interest.

Google sits on both sides of that line. It builds and markets generative systems. It also operates platforms and advertising products through which synthetic content reaches audiences. The code makes the drafting convention explicit: "will" marks mandatory measures whose compliance market surveillance authorities will monitor, "encouraged" marks optional measures that are recommended, and "may" marks optional measures or points of implementation flexibility.

The marking rules Google has accepted

Under Measure 1.1, signatories implement a marking solution built from at least one machine-readable technique. For audio, images, video and containerised text that can circulate online, one technique is not treated as enough. The code requires a multi-layered approach with at least two layers: digitally signed and time-stamped metadata under Sub-measure 1.1.1, plus an imperceptible watermark under Sub-measure 1.1.2.

The 200-token threshold is the sharpest number in the document. Free-form text longer than 200 tokens still has to be watermarked, even though the code concedes the resulting reliability may be lower than for longer passages. The glossary defines very short text as anything below that line and notes the expectation that the threshold will fall as methods improve. Because free-form text cannot carry metadata, a single marking layer is accepted for it. To compensate for weaker reliability, access to the corresponding detection mechanism may be restricted to verified expert users.

Two further commitments in Section 1 bear directly on platform operations. Signatories will not place on the market, promote or advertise tools whose purpose is to circumvent machine-readable markings. Separately, signatories who operate an online platform or a search engine, or who otherwise disseminate content publicly, are encouraged to ensure that the platform preserves metadata markings on AI-generated or manipulated content. That second provision is framed as encouragement rather than obligation, but it points at the exact failure mode the industry has documented for years: metadata stripped by upload pipelines.

Detection has to be free, and it cannot retain content

Commitment 2 is where the code imposes an ongoing operating cost. Signatories will make available a detection solution allowing deployers, end-users, third-party integrators and legitimate parties such as authorities, researchers, media organisations and civil society groups to verify whether content came from a given system.

That solution has to be free of charge. The single carve-out is scale-based: providers with fewer than 1,000,000 monthly users of their generative system, whose detection solution incurs substantial operational costs, may charge a reasonable and proportionate fee where one user's request volume exceeds a reasonable threshold. Google is far above that line, so the carve-out does not apply to it. All signatories, regardless of size, must always give free and volume-unrestricted access to market surveillance authorities, other regulators, law enforcement, media, fact-checkers, trusted flaggers, independent researchers, educational and research institutions, and civil society organisations.

Sub-measure 2.1.3 imposes a zero-retention rule. Content submitted for detection is stored only for the duration of the detection and permanently deleted immediately afterwards, and signatories will not retain a verbatim copy. Detection results must be downloadable on request in a digitally signed format containing at least a hash of the submitted content, an identifier of the detection solution and a timestamp.

The code is candid about the limits of the state of the art. At the time of publication, forensic detection mechanisms, which identify synthetic content without any prior marking, were not deemed mature enough to comply with the quality requirements in Article 50(2). Signatories that offer them anyway may restrict results to verified expert users.

The interoperability clock runs to February 2027

Measure 3.4 sets the longest deadline in the document. At publication, the code states that relevant interoperability standards and best practices are yet to be developed, except for digitally signed metadata. It therefore adopts a staged approach, requiring signatories to implement an interoperability solution for their detection mechanisms by 2 February 2027 through one or more of four routes: a publicly available industry-standard access method for routing detection queries, a publicly readable signpost embedded in the content indicating which detection solution applies, a shared detection solution operated by a consortium of signatories and open to others including smaller providers, or another solution achieving comparable interoperability.

That deadline is where Google's existing infrastructure and its stated policy position converge. According to Google, it has been partnering with third-party AI labs including Apple, Eleven Labs, Kakao, NVIDIA and OpenAI to drive industry-wide adoption of interoperable watermarking tools using SynthID, which the company describes as its "industry-leading digital watermarking technology." The consortium route in Measure 3.4 is drafted in a way that such an arrangement could satisfy, provided it stays agnostic to the participants and open to other signatories, including small and medium enterprises.

What Google already had in place

The signature formalises commitments Google has been building toward in public for two years. PPC Land reported that Google integrated SynthID verification into the Gemini app in December 2025, letting users upload a video and ask whether it was created or edited using the company's AI systems, with timestamp-specific feedback on which segments contain synthetic elements. On the distribution side, YouTube moved its generative AI disclosures to more prominent positions in May 2026 and treats both C2PA metadata and SynthID watermarks as automatic triggers for labelling, a change PPC Land covered when YouTube shifted its AI labels to locations viewers actually see.

On the advertising side the picture is less settled. Earlier this month Google introduced an AI label setting across five ad products, and PPC Land reported that the accompanying documentation placed the disclosure duty on advertisers rather than on AdSense publishers, while confirming that assets generated inside Google Ads tools carry SynthID and C2PA markup automatically. Coverage of the rollout across the five platforms also documented that the lighter disclosure regime for artistic and satirical work does not extend to persuasive commercial content.

That distinction carries weight, because it determines whether a synthetic spokesperson in a campaign is treated as a creative work or as a deep fake requiring a visible label at first exposure.

Where deployers stand

Section 2 is the half of the code that reaches agencies, brands and publishers directly. Signatories commit to disclose the artificial origin of deep fakes and of published text through the EU icon in Annex 1 of the code or an equivalent label meeting the same design and placement rules. The icon carries the capitalised acronym "AI" as its main visual element, with variants reading "AI GENERATED" for fully synthetic content and "AI MODIFIED" for partially manipulated content.

According to the annex, the icon designs were subject to empirical user-testing across several Member States, measuring perception of authenticity, noticeability, understandability and trust. The tests found that all designs achieved high noticeability, but variants including a clear textual label performed significantly better on noticeability and clarity, with respondents indicating that explicit text on AI generation or manipulation reduces ambiguity about the nature of the content.

Placement is prescribed rather than suggested. The label must be clearly perceivable at the latest at the time of first exposure, placed where no intervening overlay elements exist, and directly embedded into the content unless an equivalent alternative such as a user interface overlay is used. For video, it appears at the start and, at minimum, after interruptions including advertising breaks. For published text, it belongs above or at the top of the text, near the headline, or in the colophon. Where visual disclosure is impossible, an audible disclaimer in plain language opens the content.

Commitment 4 preserves the editorial exemption. Media service providers within the meaning of Article 2(2) of Regulation (EU) 2024/1083, already bound by editorial standards, may rely on the Article 50(4) exception by applying their existing review procedures. Everyone else claiming it must maintain policies identifying the natural or legal person holding editorial responsibility, by name, role and contact details, and publish those contact details where they are not already public.

Why this matters for the marketing community

For buyers and publishers, the practical value of a Google signature is evidentiary rather than symbolic. Under the Commission's framework, signatories can point to the code's measures to demonstrate compliance, which reduces administrative burden and gives predictability across Member States. Non-signatories carry the burden of proving their own approach is at least as effective, interoperable, robust and reliable.

An advertiser generating creative inside Google's tools is a deployer relying on a provider's marking layer. If that provider is a signatory implementing two-layer marking and a free detection solution, the advertiser inherits a machine-readable audit trail it did not have to build. The visible label, however, remains the deployer's problem. Google's July documentation was explicit that neither the automatic setting nor a custom label guarantees legal compliance.

The regulatory calendar around all of this has been unusually stable. PPC Land reported that Brussels negotiations over the Digital Omnibus package collapsed in early May 2026 without delaying the deadline, and that the provisional agreement reached on 7 May 2026 moved high-risk system deadlines to 2027 and 2028 without touching Article 50. A grandfathering rule gives providers of generative systems already on the market before 2 August 2026 until 2 December 2026 to bring Article 50(2) marking and detection into conformity.

There is also a demand-side reason the labelling question refuses to settle. Research the IAB published in January 2026, showing a 37-point gap between advertiser assumptions and Gen Z consumer sentiment toward AI-generated ads, suggests the compliance label and the trust problem are not the same problem, and that solving the first does not automatically solve the second. Google's stated worry about label proliferation confusing audiences is, read charitably, a version of the same concern from the opposite direction.

Timeline

Summary

Who: Google, with the announcement attributed to Karen Massin, Head Government Affairs and Public Policy - EU Institutions at Google. The code itself was drawn up by independent experts chaired by Kalina Bontcheva and Anja Bechmann through a process facilitated by the European Commission's AI Office.

What: Google is signing the Code of Practice on Transparency of AI-Generated Content, a voluntary framework supporting compliance with Article 50 of the AI Act. Section 1 obliges providers to apply at least two machine-readable marking layers, watermark free-form text longer than 200 tokens, and offer a free detection solution under a zero-retention rule. Section 2 obliges deployers to label deep fakes and AI-generated published text using the EU icon or an equivalent. Google simultaneously warned that additional regulatory complexity could work against Europe's competitiveness and simplification goals.

When: The announcement carries the date 24 July 2026. The Commission published the finalised code and its Article 50 guidelines on 20 July 2026, with a 22 July 2026 date for signing to obtain a presumption of compliance. The obligations apply from 2 August 2026, with grandfathered marking conformity due 2 December 2026 and detection interoperability due 2 February 2027.

Where: The obligations apply across the European Union and European Economic Area, and reach organisations established outside the bloc where the output of their AI system is used inside it.

Why: Article 50 addresses the risk that synthetic content deceives audiences about its origin, and non-compliance carries fines of up to EUR 15,000,000 or 3 percent of worldwide annual turnover, whichever is higher. Signing gives Google a recognised route to demonstrate compliance and concentrates supervisory attention on whether it has implemented the code's measures, rather than on whether an alternative approach is adequate.