Introduction
As the European Union's AI Act is set to be enforced in August 2026, a new requirement has sparked debate: the need to make AI-generated content detectable through watermarks. Article 50 stipulates that all outputs from language models must include a hidden signature to identify them as artificially generated. But how does this process work, and why do these text watermarks remain so easy to remove?
Challenges of Text Watermarking
Unlike images, where digital watermarking is facilitated by the amount of noise the human eye doesn't perceive, text is a highly compressed medium. Any modification to incorporate a watermark can be immediately noticeable. Adding specific characters or subtly altering the text can compromise the quality and flow of the message. For instance, a watermark like "every fifth letter is an ‘e’" would make the text obviously riddled with errors.
Technically Imperfect Solutions
Companies like Google with SynthID or OpenAI are exploring solutions such as using discreet Unicode characters to insert watermarks, but these methods have their limits. These solutions must balance between watermark effectiveness and text integrity. For example, allowing an AI model to decide how to fit a watermark consumes computational resources that could be better used to solve the user's problem.
The Necessity of Watermarks
Some might question whether these watermarks are truly necessary. Companies like Anthropic need to verify if their models produced a specific text. However, verification based on the similarity of each model's predictions is ineffective, as it generates too many false positives for human-written texts that appear AI-generated.
Regulatory Implications
With the AI Act, the EU imposes strict requirements for watermarking, but the technical challenges and imperfect solutions raise questions about the effectiveness of these measures. Companies will need to invest in innovative research to navigate these constraints while complying with regulations and preserving user experience.
Conclusion
As the EU's AI Act comes into effect, watermarking AI-generated texts poses an ongoing challenge. Current solutions, while creative, fail to offer foolproof protection against removal. However, the importance of these measures for transparency and regulation cannot be ignored.
Let's discuss your project in 15 minutes.