OpenAI has officially announced the rollout of new strict European Union requirements regarding the transparency of artificial intelligence systems. The developer of the popular ChatGPT chatbot is introducing a special invisible labeling technology for text content and program code generated by users within European countries. This move responds to the entry into force of key provisions of the EU Artificial Intelligence Act, which requires tech giants to ensure machine-readable identification of generated materials to avoid massive penalties.
The textGrain Technology and Watermark Principles
To implement the new requirements, OpenAI engineers utilized their proprietary development called textGrain. This system embeds an imperceptible statistical signal directly into the language model's word selection algorithm, ensuring that the visual perception of the text or code remains completely unchanged for humans. According to the developers, recognition efficiency reaches 95 percent on large text fragments of around 400 tokens. However, for shorter messages containing up to 200 tokens, this figure drops to about 80 percent, imposing certain limitations on the rapid analysis of short replies.
Contradictory Data and Watermark Resilience
Technical limitations and the real-world resilience of the implemented watermarks are actively debated in expert communities and specialized tech publications, with sources providing varying estimates. While OpenAI's official baseline documentation states that heavy paraphrasing or replacing a quarter of words with synonyms can significantly weaken the watermark—lowering detection accuracy down to 17 percent—independent technical analysis portals such as yellow.com indicate that even relatively light manual editing can reduce recognition accuracy to 66 percent. Additionally, complex domains like mathematical computations remain a serious challenge, as rigid logic limits symbol variability.
Legal Framework, Deadlines, and Massive Fines
These innovations are being implemented in strict compliance with Article 50 of the EU Artificial Intelligence Act. This rule obliges all generative systems to label content, including text, photos, audio, and video, with a special "AI-Generated" marker, especially when dealing with potential deepfakes or news reports. New developers have been subject to the rules since August, while established market players including OpenAI, Anthropic, Microsoft, Google, and Meta have been given an adaptation period until December 2. Violations carry unprecedented financial sanctions: fines can reach up to 15 million euros or up to 3 percent of the company's annual global turnover, underscoring the EU regulator's firm stance.
