In a significant response to the EU AI Act’s new transparency rules, OpenAI will soon begin embedding an invisible watermark in text outputs from ChatGPT and Codex for users residing in the European Union. This initiative, designed to help satisfy the August 2 transparency requirements that mandate AI-generated content be clearly identifiable, was unveiled by OpenAI in a blog post on Monday. The update affects all EU-based users—across both free and paid tiers—and ensures content is marked so downstream systems can verify its AI origin, in line with current regulations.
Technology Rollout and Mechanism
With the new watermarking system, the watermark will be automatically enabled for users located within the EU. Outside the European Union, those utilizing OpenAI’s API have the choice to turn this feature on for certain models, but by default, it stays off unless manually activated.
The watermark works by subtly influencing word selection through algorithmic tweaks that introduce a detectable, yet invisible, statistical signal into the text. As these signals are based on wording patterns and not visual markings, humans cannot perceive them, but specialized tools can reliably identify the watermark. When copied and pasted, the hidden marker stays intact. OpenAI emphasized that this invisible watermark does not store or disclose user identity and has demonstrated no notable impact on model output quality after being integrated.
For transparency, OpenAI published the technical framework for its method in a technical report detailing their system, referred to as textGrain, created in collaboration with Yale University and the University of Pennsylvania researchers. Their approach utilizes a ‘secret key’ to guide word prediction choices, embedding consistent patterns that can later be picked up by their detection tool.
Detection, Challenges, and Access
Despite its design, OpenAI noted that even moderate edits can compromise the watermark’s detectability. Experiments found that changing just 10% of the words with synonyms dropped detection accuracy from about 92% down to 66%. Detection is further hampered with brief passages, translation, or mathematical content—where the watermark presence is harder to verify.
Given these constraints, initial use of the detection tool will be limited to “approved researchers and expert organizations,” allowing time for broader validation and the establishment of ethical standards. OpenAI clarified that negative detection does not confirm human authorship, as the sample may be too heavily revised, too short, or produced by other AI systems. The watermark serves only to show that “an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it,” according to their statement.
Regulatory Context and Industry Trends
This update places OpenAI alongside other industry leaders adopting transparency measures as the EU tightens its oversight of AI-generated material. Back in June, Anthropic implemented a similar watermark policy for Claude outputs globally, provoking some debate from users over whether AI-assisted content marked in this way remained their own.
Originally, OpenAI delayed rolling out watermarking for fear of losing users to less restrictive competitors; this reasoning was reported by the Wall Street Journal in 2024. Now, the move reflects the wave of transparency initiatives being adopted. OpenAI, along with Anthropic, Google, Meta, and Microsoft, has committed to the EU’s voluntary code of practice requiring clear labelling for AI-generated content.
Future Outlook
By introducing watermarking for AI-generated outputs, OpenAI is setting a new benchmark for transparency and accountability in the EU. As regulatory guidelines and industry best practices evolve, OpenAI’s approach—alongside continuing improvements in watermark detection—illustrates the ongoing efforts to responsibly mark synthetic text as enforcement of the EU AI Act advances and the landscape continues to shift.