The implementation of the EU AI Act in August 2026 will require developers to implement methods for labeling generated content; however, experts warn of the critical vulnerability of such solutions to simple bypass methods.
What Happened
According to an analysis by Sean Goedecke, the primary methods of text watermarking, such as statistical sampling (e.g., Google SynthID) and the use of invisible Unicode homoglyphs, can be easily neutralized. Text watermarks are removed by paraphrasing the content with another language model or simply replacing special characters with standard ones.
Context
EU regulatory requirements (EU AI Act) aim to ensure the transparency of AI content. This forces companies to integrate detection mechanisms directly into the token generation processes, creating new technical standards for the industry.
Why It Matters for the Industry
For the industry, this means the creation of a technical challenge and the beginning of an "arms race" between developers of watermarking methods and creators of tools to bypass them. Companies are forced to design complex and potentially ineffective watermarking systems under regulatory pressure, which increases the complexity of inference pipelines.
Why It Matters for Users
Ordinary users can easily "clean" any text of its AI digital footprint simply by running it through another chatbot to change the style. This deprives text watermarks of their primary function—guaranteeing content authenticity.
What Is Not Yet Known / Limitations
There is a risk that due to the technical impossibility of guaranteed text watermarking, EU AI Act standards may require revision toward more complex verification methods, such as cryptographic signatures.
Sources
Author
Look at AI, Editorial Team