Summary of Key Points
Following the implementation of Article 50 of the EU's Artificial Intelligence Act, which imposes transparency obligations (requiring labeling and detectability of AI-generated content) in August 2026, Anthropic was among the first to adopt watermarks and metadata tagging on a global scale, although it has not made its detection tools public. A key requirement of the EU rules is "interoperability"—that is, detection tools from different companies must be able to recognize each other's watermarks. Companies that sign the Code of Conduct by February 2027 must comply with these standards. The act has a "long-arm jurisdiction" effect, meaning that Chinese companies, even if they do not operate overseas, must comply if their AI content is used by EU users (including through open-source models). There are differences between China's current AI labeling regulations and those of the EU (for example, watermarks for text are encouraged rather than mandatory in China, and there is a lack of interoperability requirements). Companies like Google and Anthropic are competing for global standards in watermark detection, and Chinese companies need to be vigilant about compliance risks and participate in this standard-setting process.
I. EU's Strict Requirements: AI Labeling Is Not Enough; Detection and Universality Are Also Necessary
Do many people think that adding a watermark to AI-generated content automatically makes it compliant? The EU doesn't see it that way. Article 50 clearly stipulates two requirements: first, the content must have machine-readable markers (such as watermarks); second, there must be tools available to detect these markers, and these tools must be interoperable—meaning that a company's detection tool should be able to recognize another company's watermark, and vice versa.
Why are the requirements so strict? Because simply adding a watermark without providing a way to verify its authenticity is meaningless. The EU has set a deadline: by February 2027, companies that have signed the Code of Conduct must ensure that their detection tools are interoperable. This is not just a suggestion; it is a mandatory compliance requirement. Failing to meet this standard means you cannot operate in the EU market.
II. Can Chinese Companies Be Regulated by the EU Even If They Don't Operate Overseas? Yes, Due to "Long-Arm Jurisdiction"
Don't assume that Chinese companies are off the hook just because they do not actively expand overseas. The extraterritorial scope of the EU's AI Act is even broader than that of the GDPR (EU Data Protection Regulation). As long as your AI content is used by EU users (for example, if EU visitors access your website or use your open-source models), you must comply with Article 50, regardless of whether you have a company in the EU.
What's more problematic is that open-source models are not exempt from these regulations. While the EU generally exempts open-source AI, generative AI systems (such as chatbots and image generators) that are used within the EU still fall under the regulation, and the providers (i.e., Chinese large-scale AI companies) are responsible for compliance. For example, if you make your model available on GitHub and EU users download and use it, you must comply with the EU rules—this is a point that many Chinese companies often overlook.
III. How Do China's AI Labeling Regulations Compare to Those of the EU?
China introduced requirements for labeling AI-generated content as early as 2023, but there are two significant differences:
1. Text Watermarks: Encouraged vs. Mandatory
In China, adding implicit watermarks to text is encouraged (due to high technical costs), while the EU requires explicit, machine-readable watermarks for texts related to public interest (such as news and political content) that must meet specific standards of effectiveness and reliability. Therefore, content that complies with Chinese regulations may be illegal in the EU.
2. Detection Tools: Internal vs. Open Access + Interoperability
Chinese companies have the capability to conduct internal detections, but the rules do not require them to make these tools available externally. The EU, on the other hand, requires that detection tools be freely provided to regulatory agencies, media, and fact-checking organizations, and they must be interoperable with those of other companies. This poses a new challenge for Chinese companies.
IV. How to Achieve Interoperability? Four Approaches, with Google and Anthropic Competing for Standards
The EU has outlined four ways to achieve interoperability:
1. Standardized APIs: Use a unified interface to send detection requests to the relevant company's tools.
2. In-Content Indications: Embed signals within the content to specify which detection tool should be used.
3. Alliance Sharing: Multiple companies can collaborate on shared detection solutions (e.g., the SynthID alliance formed by Google, Apple, and OpenAI).
4. Equivalent Alternatives: Provide a way to verify content without having to test each tool individually.
Google and Anthropic are both making moves in this area. Google's SynthID alliance is closed, while Anthropic is focusing on global labeling as a step towards future interoperability. The company that gains EU recognition first may define the global standard for watermark detection—this is a significant issue, as control over standards means control over the industry.
V. Shortcomings and Opportunities for Chinese Companies
China's existing regulations provide a foundation for compliance, but there are gaps in terms of interoperability. However, since Google and Anthropic's approaches are still evolving, and the EU has not yet clarified the specific interoperability requirements, Chinese companies have an opportunity to:
- Fill in compliance gaps (e.g., by making text watermarking mandatory).
- Actively participate in international standard-setting processes (e.g., by joining alliances or proposing their own solutions).
- Avoid being forced to follow others' standards, which could result in additional costs.
In summary, this is not just a matter of compliance; it is also a global competition for AI content traceability standards. Chinese companies cannot afford to be absent from this debate.
Conclusion: The EU's AI transparency rules may seem like a technical compliance issue, but they are actually about competing for standard-setting power. Chinese companies need to be aware of both the compliance risks and the opportunities presented by standard development. After all, those who control the standards will have the say in the future direction of the AI industry.