OpenAI is preparing to add invisible watermarks to text generated by ChatGPT and Codex for users in the European Union. The move is part of the company’s effort to comply with transparency requirements under the EU AI Act.
The watermark will not appear as a visible label, symbol or special character. Instead, OpenAI’s technology, called textGrain, changes the statistical pattern of the words selected by the AI model. A dedicated detector can then look for that pattern and determine whether the text likely came from an OpenAI system.
OpenAI said the watermarking system will be introduced to eligible ChatGPT and Codex users in the EU over the coming weeks. It will apply across ChatGPT plans, while the company is keeping text watermarking optional for API customers outside the EU.
OpenAI Will Add Watermarks to ChatGPT and Codex Text
OpenAI’s announcement comes as the company expands its work on content provenance, which aims to help users and organizations understand where AI-generated content came from.
Under the new approach, eligible text produced by ChatGPT and Codex in the European Union will contain an invisible watermark. Readers will not see the watermark when they read, copy or paste the text.
Instead, textGrain subtly influences the model’s choices between possible words or tokens. Across a longer passage, those choices create a statistical pattern that OpenAI’s detector can search for. The system is different from traditional AI-content detection tools.
Rather than trying to guess whether text was written by AI based on writing patterns, OpenAI’s detector looks for a signal that was embedded when the text was generated. OpenAI says it developed textGrain to balance detection with the quality and variety of model responses.
EU AI Rules Push OpenAI Toward Text Watermarking
The European Union’s AI rules are a major reason behind the rollout. The EU AI Act requires providers of generative AI systems to make AI-generated content identifiable in a machine-readable way. OpenAI said its new text watermarking system is intended to meet this requirement for eligible text generated by its models.
The company is initially limiting automatic text watermarking to the EU rather than making it a worldwide default. This means users in other regions will not automatically receive the same watermark on ChatGPT or Codex responses at launch.
OpenAI said the regional rollout will allow it to collect feedback and learn how the technology performs in real-world conditions before deciding how its approach should evolve.
ALSO READ: OpenAI’s Latest AI Agent Incident Could Trigger New EU Scrutiny
How OpenAI’s textGrain Watermarking Technology Works
textGrain works by modifying the statistical pattern behind the model’s word choices. When an AI model generates text, it calculates probabilities for possible words or tokens that could come next.
OpenAI’s watermarking system creates a pattern in those choices using a secret key. A detector can later analyze the text and look for that pattern. The system does not insert hidden characters, invisible spaces or unusual punctuation into the response. This is important because users should not notice any visible difference in the text.
OpenAI also says its testing found no meaningful impact on model performance, while the effect on model speed is negligible. OpenAI’s testing of its latest frontier model, Astra, showed similar results across several benchmarks with and without watermarking.
OpenAI Says Watermarks Cannot Prove Who Wrote the Text
OpenAI is also warning users not to treat the watermark as definitive proof of authorship. A detected watermark can indicate that an OpenAI model likely generated or processed the text. However, it does not show who used the model, who owns the content or how much of the final text was written by a person.
The watermark also does not determine whether the information in the text is accurate. OpenAI says the absence of a watermark should not be treated as proof that a human wrote the content either.
Text may not be detected if it is too short, substantially edited, translated or generated by an unsupported model. This limitation is particularly important for people who use ChatGPT to draft articles, reports, academic material or other long-form content.
Editing Can Make AI Watermarks Harder to Detect
OpenAI’s own testing shows that text watermarking has significant limitations. The company found that longer passages are easier to detect than shorter ones. At a target false-positive rate of 1%, its detector identified watermarks in about 80% of 200-token passages and around 95% of 400-token passages in one set of tests.
Detection was lower for subjects such as mathematics, where there is less flexibility in word choice. Editing can also weaken the signal. In OpenAI’s evaluation of 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%. Replacing 25% of words reduced detection to about 17%.
These results are one reason OpenAI is not making its detection system publicly available at launch.
OpenAI Will Restrict Access to Its Detector
OpenAI is opening applications for access to its text watermark detector, but the company will initially restrict access to approved researchers and expert organizations. The goal is to allow outside experts to test the technology, study its limitations and help improve text provenance systems.
OpenAI said the detector will report whether it detects an OpenAI watermark. It will not identify the user or reveal their prompts or conversations. The company plans to keep this limited approach because false positives and missed watermarks remain possible.
API Customers Can Choose OpenAI Text Watermarking
OpenAI is also making text watermarking available to API customers around the world. Starting October 5, developers using supported OpenAI models can opt in to watermarked text. However, the feature will remain off by default for API customers.
This gives businesses and developers the option to use watermarking as part of their own AI transparency practices, even when they are outside the EU. OpenAI is also working with cloud partners to make watermarking available for eligible OpenAI model outputs accessed through their services.
How OpenAI’s Watermarks Will Change the ChatGPT Experience
For most EU ChatGPT users, the change should happen without requiring any action. The watermark will be embedded into eligible AI-generated text automatically and will not be visible during normal use. Users will still be able to copy and paste their responses as before.
The bigger change is that approved detection systems will have another way to identify text that likely came from an OpenAI model. However, OpenAI’s own data suggests that the system should not be treated as a perfect AI detector.
Rewriting, translation and other changes can weaken the watermark, while short or highly constrained text can be difficult to identify. OpenAI also says text watermarking will not replace visible AI labels or other disclosures that may be required in specific situations.
OpenAI Plans to Expand Text Watermarking Beyond the EU
OpenAI’s EU rollout marks a new step in its broader effort to add provenance signals to AI-generated content. The company already uses technologies such as Content Credentials and invisible watermarks for supported images and audio. It is now extending that approach to text through textGrain.
OpenAI also plans to release textGrain as open-source technology. This could allow researchers and other developers to study the approach and build additional tools around AI text provenance.
For now, the company is taking a cautious approach. Automatic watermarking will begin with eligible ChatGPT and Codex text in the EU, while API watermarking remains optional globally.
As AI-generated content becomes more common, the OpenAI rollout could give regulators and technology companies another way to identify the origin of digital content without changing how that content appears to users.
