OpenAI to Watermark ChatGPT Text in EU to Meet AI Act Transparency Rules

OpenAI announced that it will begin adding an invisible watermark to all ChatGPT and Codex text produced within the European Union, a step taken to comply with the EU AI Act’s new transparency requirements that took effect on August 2. The watermark, called textGrain, will be deployed over the next few weeks to all eligible users on every plan, but it will remain off by default for global API customers.

Key points

• OpenAI will watermark ChatGPT and Codex text in the EU to meet AI Act transparency rules.
• The invisible watermark is embedded in word choices, detectable only with a secret key.
• Editing can diminish detection; 10% synonym replacement drops accuracy from 92% to 66%.
• Short passages, math answers, and translations are harder for detectors to flag.
• OpenAI’s own test shows no measurable impact on model performance.

What happened

In a Monday blog post, OpenAI outlined the rollout plan for its text watermarking initiative. The company explained that the watermark is not a visible symbol; instead, it subtly biases the model’s next‑word probabilities, leaving a pattern that a detector can identify when supplied with the secret key. Because the watermark is embedded in the text itself, it travels with the content when copied or pasted.

OpenAI said developers using its API anywhere in the world can opt in to the watermark for select models starting today, but the feature will be disabled by default. The EU‑only rollout is a direct response to the AI Act’s requirement that AI‑generated content be identifiable by other systems.

The company released a technical report co‑written with researchers from the University of Pennsylvania and Yale that details how the watermark works. The paper describes a secret key that sorts next‑word predictions, and how aggregating hundreds of these nudges allows a detector to flag AI‑generated text.

When tested, the watermark proved robust: detection rates hovered around 92% on unedited passages. However, the company noted that editing could reduce success. Replacing 10% of words with synonyms lowered detection to 66%, and the authors warned that very short, heavily edited, or translated text would be even harder to catch. "OpenAI also cautioned that a missing watermark ‘does not prove human authorship,’" the post states.

OpenAI also clarified that the watermark does not reveal which user generated the text, nor does it indicate the degree of human editing. "These limitations contribute to our decision to provide initial detector access only to approved researchers and expert organizations," the report quotes the company.

The backstory

The EU AI Act, finalized in 2023, introduced the first comprehensive regulatory framework for high‑risk AI systems. Among its provisions, the Act requires that any content produced by an AI system be marked in a way that third‑party tools can identify. In March, Anthropic announced it would watermark Claude’s output worldwide, a move that sparked backlash from users who felt the watermark implied a lack of human agency.

OpenAI had previously developed a watermarking technique but delayed its release, citing concerns that users might switch to competitors that did not watermark. The Wall Street Journal reported in 2024 that the company had held back its system for this reason. The EU Act’s enforcement deadline has prompted several major AI vendors, including Google, Meta, Microsoft, and Anthropic, to commit to the new transparency rules.

Why it matters for businesses

The watermarking policy forces enterprises that rely on ChatGPT or Codex for content generation to reassess how they verify authenticity. For compliance teams, the presence of a watermark will become a new audit marker. For developers, the optional API flag means they can choose to expose watermarking in their own products, potentially affecting user experience and data integrity.

From a legal standpoint, the watermark could serve as evidence that a text was AI‑generated, which may impact liability discussions in intellectual property disputes. Conversely, the fact that editing can erase the watermark introduces a new vector for fraud or misrepresentation.

Technical trade‑offs

The watermark’s design balances transparency with model fidelity. OpenAI reports that enabling the watermark does not noticeably degrade performance or alter output quality. However, the detection algorithm’s reliance on a secret key raises concerns about key management and potential misuse. If the key were compromised, malicious actors could generate text that bypasses detection.

the watermark’s resilience to editing means that human reviewers cannot rely solely on the absence of a watermark to confirm authenticity. The company’s own experiments show that even moderate synonym substitution can reduce detection rates, suggesting that watermarking alone will not be a silver bullet.

Regulatory and competitive dynamics

By choosing to implement the watermark only within the EU, OpenAI signals a pragmatic approach to compliance while preserving flexibility for global customers. This selective rollout may create a market split, with EU users facing a different experience than those outside the region. Competitors that adopt a global watermarking policy might gain an advantage among EU‑centric enterprises that prioritize transparency.

The EU Act’s enforcement deadline adds urgency. Companies must ensure that their content pipelines can handle watermark‑tagged text and that their detection tools are ready to interpret the signal. Failure to do so could expose firms to regulatory penalties or reputational damage.

What builders and businesses should do

• Verify whether your internal or client‑facing AI tools will be subject to EU watermarking and update your data handling policies accordingly.
• Integrate detector libraries that accept the OpenAI secret key to flag AI‑generated text, especially for compliance reporting.
• Educate content reviewers that the absence of a watermark does not guarantee human authorship; incorporate additional verification steps.
• Review your API usage to decide whether to enable watermarking for specific models, balancing transparency requirements with user experience.
• Monitor updates to the EU AI Act and OpenAI’s documentation for changes in watermarking guidelines or key management practices.

The bottom line

OpenAI’s EU‑only watermark rollout signals a broader industry shift toward mandated transparency. Enterprises using ChatGPT or Codex must adapt their compliance frameworks now, as the watermark will become a key audit marker by the end of the year. Keep an eye on key‑management security and detection tooling, as these will define how effectively businesses can leverage the new transparency rule.