GoogleがEU AI法「AIが生成したコンテンツの透明性」実践規範に署名Google is signing the EU AI Act Code of Practice on Transparency of AI-Generated Content.

Google has signed the European Union's Code of Practice on transparency for AI-generated content, a voluntary framework developed under the EU AI Act. The move is significant because it commits one of the largest producers and distributors of generative AI tools to a shared set of standards for helping users recognize when they are looking at, listening to, or reading material that a machine produced. As synthetic media becomes harder to distinguish from human-made work, the question of disclosure has moved from an academic concern to a regulatory requirement, and Google's participation signals that major platforms are prepared to align with the EU's approach.

The transparency provisions sit within Article 50 of the AI Act, which requires that providers and deployers of certain AI systems make clear when content has been artificially generated or manipulated. In practice, this covers a range of outputs, including text, images, audio, and video, and it applies with particular force to so-called deepfakes and to synthetic media that could be mistaken for authentic recordings of real people or events. The Code of Practice is intended to translate these legal obligations into concrete technical and operational commitments, offering signatories a way to demonstrate compliance before enforcement fully takes effect. Codes of Practice under the AI Act are not themselves law, but adherence is designed to create a presumption of conformity, giving companies a clearer path to meeting their statutory duties.

For Google, the commitment appears to build on work the company has already been doing around content provenance. Its SynthID technology, developed by DeepMind, embeds watermarks into AI-generated images, audio, text, and video that are intended to be detectable by software while remaining imperceptible to people. Google has also participated in the Coalition for Content Provenance and Authenticity, known as C2PA, an industry group whose standard attaches cryptographically signed metadata to media files to record how they were created and edited. These provenance approaches are complementary: watermarking survives some forms of copying and re-encoding, while metadata standards provide a richer record of a file's history but can be stripped more easily. Signing the EU Code is likely to reinforce Google's incentive to deploy such mechanisms consistently across products including its Gemini models and image-generation tools.

The broader context is a shift in how governments are treating generative AI. The EU AI Act, which entered into force in 2024, takes a risk-based approach, imposing the strictest obligations on systems deemed high risk while applying lighter transparency rules to many general-purpose applications. Its provisions are being phased in over several years, and the transparency requirements around synthetic content are among the elements the European Commission has prioritized through accompanying guidance and voluntary codes. The Act is widely watched because of its potential extraterritorial reach: companies offering AI systems in the EU market must comply regardless of where they are based, which tends to push global firms toward building compliant features into products worldwide rather than maintaining separate versions.

Other technology companies have taken varying positions on the EU's voluntary instruments. Earlier codes tied to the AI Act, including one addressing general-purpose AI models, drew a mix of signatures and public reservations, with some firms warning about compliance costs and legal uncertainty. Whether the transparency code attracts broad industry participation will help indicate how durable the EU's collaborative regulatory model proves to be. Adjacent efforts, such as media-literacy campaigns and platform-level labeling policies from social networks, also form part of the ecosystem that determines whether disclosure reaches ordinary users in a meaningful way.

Practical challenges remain. Watermarks and provenance metadata can be removed, degraded, or spoofed, and no single technique is currently robust against a determined adversary. Detection tools can also produce false positives, and disclosure labels only work if users notice and understand them. Google's signature does not resolve these limitations, but it does formalize a commitment to industry-wide standards and to the ongoing development of detection and labeling systems. As enforcement of the AI Act advances, the credibility of such commitments will likely be measured by how thoroughly the underlying technologies are implemented and how consistently synthetic content is disclosed across the services people use every day.