Google has signed the European Union’s Code of Practice on Transparency of AI-Generated Content, committing to a common framework for marking and labeling content created or manipulated with artificial intelligence.
The announcement was made on July 24, 2026, shortly before the related transparency requirements in Article 50 of the EU AI Act begin applying on August 2.
The code is voluntary, but the underlying legal transparency requirements are not. Companies that do not sign may use other methods to comply, but they will need to demonstrate that their own systems meet the European Union’s requirements.
The rules cover AI-generated text, images, audio and video, including deepfakes and some publications about matters of public interest.
What Did Google Sign?
Google signed a Code of Practice created to help AI providers and businesses comply with the transparency requirements of the EU AI Act.
The code was developed through a multi-stakeholder process facilitated by the European AI Office. Participants included AI companies, researchers, civil society organizations, academic experts and specialists in content authenticity.
It provides practical guidance for two main groups:
- Providers: companies that create systems capable of generating or manipulating content.
- Deployers: organizations and individuals that use those systems to create and publish content.
Signing the code gives companies a recognized framework for demonstrating compliance throughout the European Union.
Companies can choose not to sign, but they must still comply with the relevant provisions of the AI Act when those provisions apply to their products or activities.
When Do the New Requirements Begin?
The Article 50 transparency obligations are scheduled to apply from August 2, 2026.
The rules are intended to reduce the risk that people will be deceived or manipulated by synthetic content that appears authentic.
They complement other parts of the EU AI Act dealing with general-purpose AI models, high-risk systems and specific prohibited uses.
Some older AI systems placed on the European market before August 2 may receive a transitional period, depending on the exact system and requirement.
Businesses operating in Europe should therefore examine whether their tools create content covered by the rules and whether they act as a provider, deployer or both.
What AI Content Must Be Marked?
Providers of generative AI systems are expected to make generated or manipulated outputs detectable in a machine-readable format.
This may apply to:
- AI-generated images.
- AI-generated or modified videos.
- Synthetic voices and audio.
- Music created by an AI model.
- Text produced by generative AI systems.
The marking system should be effective, interoperable, robust and reliable as far as technically feasible.
In practice, this could involve invisible watermarks, metadata, content credentials or other technical signals that automated systems can detect.
The code does not require every company to use one specific technology. The chosen method must be appropriate for the content and capable of functioning across the relevant distribution environment.
What Must Be Clearly Labeled for People?
The rules also contain disclosure requirements intended to be visible and understandable to ordinary users.
Deployers may need to clearly label content that constitutes a deepfake.
A deepfake may include an AI-generated or manipulated image, audio recording or video that resembles a real person, place, object or event and could falsely appear authentic.
AI-generated or manipulated text published to inform the public about matters of public interest may also require disclosure.
However, the EU framework recognizes an exception when the material has undergone human review and an individual or organization assumes editorial responsibility for the publication.
This distinction is especially important for publishers, news organizations and blogs that use AI as an assistant but still review and edit every article before publication.
What Is SynthID?
Google says its participation in the code aligns with the continued development of SynthID, its technology for watermarking and identifying AI-generated content.
SynthID embeds signals directly into content created by Google’s generative AI systems.
The watermark is intended to be invisible or inaudible to people while remaining detectable by compatible verification technology.
Google uses versions of SynthID for:
- Images.
- Video.
- Audio.
- AI-generated text.
For images and video, the watermark is designed to survive common changes such as cropping, filters, compression and adjustments to frame rates.
For audio, Google says the signal can remain detectable after operations such as compression, speed changes or the addition of background noise.
Text watermarking works differently. SynthID adjusts the probability of certain token choices during generation, creating a statistical pattern that detection tools can later identify.
Users can also upload certain media to Gemini and ask whether Google AI created or modified it. Gemini checks for the presence of a SynthID watermark.
Google Is Working With Other AI Companies
Google says it is working with several third-party AI companies to encourage interoperable watermarking tools based on SynthID.
The announced partners include:
- Apple.
- ElevenLabs.
- Kakao.
- Nvidia.
- OpenAI.
Interoperability is important because AI content rarely remains inside the platform that created it.
An image may be generated in one application, edited in another, published on a social network and downloaded by thousands of users.
A transparency system becomes more useful when different platforms can preserve and recognize the same origin information.
Without interoperability, each company could create its own incompatible watermarking system, making verification more confusing for publishers, users and automated moderation services.
Why Google Still Has Concerns
Although Google signed the code, the company also warned that excessive regulatory complexity could create new problems.
Google argues that technical solutions for detecting AI content are still evolving. It is concerned that requiring several overlapping labels, notices and legal disclosures could confuse users instead of helping them.
For example, the same piece of content might contain:
- A visible AI-generated label.
- An invisible watermark.
- Content provenance metadata.
- A platform-specific disclosure.
- A legal notice required by a particular country.
When these systems use different terminology or provide conflicting information, users may find it harder to understand what actually happened.
Google says it plans to work with regulators and industry partners to create practical systems that provide useful context rather than simply adding more warnings.
Why Machine-Readable Marking Matters
A visible label can inform the person viewing the content, but it can be removed through cropping, editing or reposting.
Machine-readable marking provides another layer of information that platforms and verification services may be able to detect automatically.
For example, a social platform could examine an uploaded image and determine that it contains an AI watermark even when no visible label appears on the image.
This could support:
- Deepfake detection.
- Election integrity systems.
- Content moderation.
- Copyright and provenance investigations.
- Journalistic verification.
- Protection against impersonation scams.
However, no watermarking technology is perfect.
Signals may be damaged by aggressive editing, screenshots, re-recording or deliberate attempts to remove them. Detection systems may also produce uncertain results.
Watermarking should therefore be treated as one source of evidence rather than absolute proof that content is authentic or artificial.
What This Means for Deepfakes
Deepfakes have become one of the most visible risks associated with generative AI.
Realistic synthetic video and audio can be used to impersonate public figures, executives, family members and ordinary individuals.
Common risks include:
- Financial scams using cloned voices.
- False political statements.
- Fabricated evidence.
- Non-consensual synthetic images.
- Impersonation of company executives.
- Misleading videos of real-world events.
Clear labels and persistent origin information cannot prevent every misuse, but they can make manipulated media easier to investigate and identify.
The effectiveness of the European rules will depend on how consistently companies apply the requirements and how well content markings survive when media moves between platforms.
What This Means for Publishers and Bloggers
Publishers using AI tools should begin documenting how generated content is created, reviewed and approved.
A responsible publishing workflow may include:
- Human review of every AI-assisted article.
- Verification of names, dates, numbers and quotations.
- Editorial responsibility for the final publication.
- Clear disclosures when synthetic media could be mistaken for reality.
- Records of the original image or video source.
- Preservation of available content credentials and metadata.
Using AI to organize notes, correct grammar or assist with a draft does not automatically remove editorial responsibility from the publisher.
The organization publishing the content should still ensure that the final material is accurate, lawful and not misleading.
For a news or educational website, the safest approach is to treat AI as a production tool while keeping a human responsible for research, review and publication.
Will Every AI Image Need a Visible Label?
The exact disclosure requirements depend on how the image is used and whether it could mislead people.
A clearly fictional illustration accompanying a technology article is different from a realistic synthetic photograph presented as evidence of a real event.
Creative, artistic, fictional and satirical content may be treated differently, but users must still receive appropriate disclosure when there is a serious risk of confusion.
Businesses should avoid assuming that a small disclaimer hidden in a privacy policy will satisfy every requirement.
Labels should be understandable, accessible and placed where users can reasonably see them.
Why This Decision Matters Outside Europe
The EU AI Act directly applies within the European regulatory environment, but its influence is likely to extend beyond Europe.
Large technology companies generally prefer to avoid building completely separate AI systems for every country.
When a major market establishes technical requirements for watermarking and disclosure, companies may gradually apply similar systems more broadly.
Publishers outside Europe may also reach European readers, customers or users through their websites and online services.
This means European transparency standards may influence how AI-generated content is labeled around the world.
Final Thoughts
Google’s decision to sign the EU transparency code is an important step toward creating common standards for identifying AI-generated content.
The code provides companies with a practical path for meeting legal obligations that begin applying on August 2, 2026.
Google’s involvement also gives SynthID a larger role in the emerging ecosystem of AI watermarking and content verification.
However, technical labels alone will not solve the problem of deceptive synthetic media.
Effective transparency will require cooperation among AI developers, publishers, social networks, regulators and the people who create and share content.
The challenge is not simply adding an “AI-generated” label. It is creating reliable information that remains attached to content as it is edited, copied and distributed across the internet.
Sources
- Google’s official announcement published July 24, 2026.
- European Commission Code of Practice on Transparency of AI-Generated Content.
- Google DeepMind SynthID documentation.