OpenAI Introduces Invisible Text Watermarks in the European Union

OpenAI has announced it will begin adding invisible watermarks to text generated by ChatGPT and Codex for users in the European Union. The move is designed to comply with the European Union's AI Act, which requires artificial intelligence companies to label AI-generated content so that other computer systems can identify it. The new feature will roll out over the coming weeks to eligible users on all subscription plans within the EU.

How the Watermark Works

The watermarking technology, which OpenAI calls textGrain, does not use visible symbols or stamps. Instead, it works by subtly influencing the AI's word choices. This leaves a specific pattern that human readers cannot see, but a specialized detector can easily recognize. Because the watermark is embedded directly into the choice of words, it remains intact even when the text is copied and pasted.

OpenAI developed this method alongside researchers from the University of Pennsylvania and Yale. The system uses a secret key to influence how the AI predicts and selects the next word in a sentence. By combining hundreds of these tiny word adjustments throughout a document, the detector can identify OpenAI's writing using only the text and the secret key. OpenAI states that this process does not hurt the performance of its models and does not identify individual users.

Limits of the Technology

While the technology is sophisticated, OpenAI acknowledges that it is not foolproof. The watermark can be degraded or removed through editing. During testing, replacing just 10 percent of the words in a passage with synonyms caused the detection rate to drop from 92 percent to 66 percent.

Additionally, the watermark is much harder to detect in short passages, mathematical answers, and translated text. OpenAI cautioned that the absence of a watermark does not prove a human wrote the text. The text could simply be too short, heavily edited, or created by a different company's AI. The company also noted that a watermark cannot measure how much human effort, editing, or creativity went into a piece of writing.

Availability and Access

OpenAI is not making this text watermarking a global default. For now, the automatic rollout is strictly limited to the EU. However, developers worldwide who use OpenAI's API—the application programming interface that allows external software to connect to OpenAI's models—can choose to turn the feature on starting today. It remains turned off by default for these developers.

To prevent false positives and missed detections, OpenAI is keeping the detection tool private for now. Only approved researchers and expert organizations can apply for access on a case-by-case basis. When used, the detector will only report whether it finds an OpenAI watermark; it will not reveal the user's identity, prompts, or conversations.

The Broader Industry Context

The decision to implement watermarking comes amid growing regulatory pressure. Other major tech companies, including Anthropic, Google, Meta, and Microsoft, have also committed to the EU's guidelines on AI-generated content. Anthropic recently rolled out its own global watermarking system for its Claude AI, which is based on Google DeepMind's SynthID technology.

However, watermarking has faced pushback. Some users of rival systems have complained that watermarks ignore the human effort, context, and instructions that go into generating AI text. OpenAI itself previously held back on releasing text watermarking due to concerns that users might switch to competitors that do not label their text. By limiting the initial rollout to the EU, OpenAI hopes to learn from real-world feedback before deciding on a global approach.