AI Article Thieves
Under the EU AI Act, which mandates full compliance by all companies by December 2, companies using artificial intelligence and other interested observers will be able to identify text generated by AI applications. The regulation requires the inclusion of a “hidden watermark” in every piece of generated text. These watermarks are not visible letters or symbols; rather, they are typically hidden “mathematical and statistical patterns” embedded in the way words are selected. While the resulting text will appear completely natural, approved detection tools can accurately identify it as AI-generated rather than human-written. Consequently, it will become increasingly difficult to continue directly copying and pasting AI-generated content for school assignments, academic research, work emails, or drafting official books, articles, and reports, as their AI origin will be detectable.
The European Union issued the EU AI Act concerning the detection of AI-generated written content. The transparency provisions (Article 50) began to take effect on August 2, 2026, requiring providers of generative AI systems to make their outputs machine-readable and detectable as AI-generated content, wherever technically feasible. Anthropic, the company behind the Claude model, has already begun embedding hidden watermarks in all its outputs as of August 2, 2026. This watermarking will be applied globally, not just to users in Europe, because there is currently no reliable method to restrict application geographically.
But how does this method work technically? AI companies embed invisible symbols or characters into every piece of text. The technique operates through a subtle, hidden directive during word selection. Instead of choosing the next word purely randomly from among plausible options, the model introduces a slight statistical bias toward a specific set of “preferred” words (determined by a secret key). This allows detection tools to confidently determine, with high statistical confidence, whether the text was generated by AI.
Although the compliance deadline is December 2, 2026, the three major companies are not yet fully compliant, with the exception of Claude, owned by the AI giant Anthropic. All companies have previously signed the “European Union Code of Conduct on the Transparency of AI-Generated Content” and declared their intention to implement similar technologies in the future. Currently, OpenAI applies watermarks only to images, not to text. The law mandates compliance only “wherever technically feasible,” meaning the obligation is not absolute in every case but depends on actual technical applicability. Nevertheless, there are methods to evade AI detection.
Ahmed Al-Saraf
