OpenAI Introduces Invisible Text Watermarking for ChatGPT in the EU
To comply with the EU AI Act, OpenAI is rolling out invisible text watermarking for ChatGPT and Codex in Europe, while offering developers global opt-in access.
OpenAI has announced that it will begin embedding an invisible watermark into text generated by ChatGPT and Codex across the European Union. The decision was detailed in a company blog post on Monday and serves to align the company's services with the European Union AI Act. Under transparency obligations that took effect on August 2, the EU legislation requires artificial intelligence providers to mark synthetic content so external systems can detect it.
The deployment will take place over the coming weeks for qualifying ChatGPT and Codex users across all subscription tiers in the EU. OpenAI confirmed that the watermarking mechanism will not be activated as a global default for consumers at launch, remaining restricted to European users for now.
How the textGrain watermarking system works
Rather than embedding visual characters or visible markers, the watermarking technique functions at the linguistic level. The system subtly influences the model's token distribution, guiding its word choices to create an underlying statistical pattern. While ordinary human readers cannot notice the adjustment, a specialized detector can identify the output.
Because the watermark is embedded directly within the phrasing and vocabulary of the response, it stays intact when text is copied and pasted elsewhere. According to OpenAI, the implementation does not embed user identifiers or compromise user privacy. The company also stated that internal evaluations showed no meaningful degradation in the quality or performance of its models when the watermark is active.
Alongside the deployment, OpenAI published a technical report detailing the underlying framework, named textGrain, developed in collaboration with academic researchers from Yale University and the University of Pennsylvania. The system operates by using a secret cryptographic key to organize and bias next-word predictions during sentence completion. Over a span of hundreds of biased token decisions, the cumulative shifts allow an external detector armed with the corresponding secret key to verify AI generation.
Technical limitations and detection robustness
OpenAI's empirical testing highlights notable constraints in how resilient the watermarking signal remains when text undergoes revision. The company's data indicates that substituting just 10% of the words in a watermarked passage with synonyms decreased the detection rate from approximately 92% to 66%.
Beyond manual editing, detection accuracy falls off in several specific contexts. The company noted that short snippets of text, mathematical answers, and translated passages prove significantly harder to identify reliably. Due to these vulnerabilities, OpenAI is restricting access to the detection tool during its initial stage, opening it exclusively to vetted researchers and expert organizations to examine its dependability and safe application.
OpenAI emphasized that the absence of a watermark does not confirm that a piece of text was written by a human, noting that the content could originate from another vendor's model, be too brief to register, or be substantially rewritten. Furthermore, the company clarified that while watermarks can verify that an OpenAI model processed or generated a portion of a document, they cannot determine the amount of human judgment, editing, or creative input involved.
Watermarking initiatives have drawn industry-wide focus following commitments by Anthropic, Google, Meta, Microsoft, and OpenAI to adhere to the EU's code of practice for artificial intelligence content. Anthropic rolled out watermarking globally for Claude text two months prior, which prompted pushback from users who maintained that their prompts, context, and editorial direction constituted the primary work. OpenAI had previously engineered text watermarking technology but held back on a public rollout over concerns that users might migrate to competing platforms that did not mark output, as reported by The Wall Street Journal in 2024.
What it means for developers
For software engineers and API customers, OpenAI's watermarking rollout introduces a distinct implementation path compared to consumer tiers. While the watermark is automatically applied to EU users of ChatGPT and Codex, the system is turned off by default within the API. However, OpenAI has made the feature immediately available to developers globally on an opt-in basis for select models.
Teams building software for European audiences must now consider how compliance with the EU AI Act intersects with their data pipelines, downstream processing, and content delivery architectures. Because the watermarking signal can degrade rapidly when downstream workflows alter vocabulary, run automated paraphrasing, or combine short fragments, developers cannot treat detection as an absolute guarantee of origin.
As major providers introduce differing technical standards for compliance, many engineering teams are assessing multiple model families side by side. Developers can try top AI models cheaply through one API at https://apixoai.online to evaluate how different systems handle generation, latency, and compliance requirements across varied development environments.
Ultimately, API developers maintain control over whether to embed textGrain markers outside the EU, allowing organizations to independently balance regulatory transparency with workflow flexibility.
Source: OpenAI will start watermarking ChatGPT’s text in the EU — TechCrunch AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
