OpenAI to Watermark Text in the EU
OpenAI announced that in response to EU regulations, content traceability will expand from images and audio to text, while acknowledging significant limitations in existing text watermarking technology. Tools already exist for verifying whether images and audio were generated by its models. Text will be added to help determine if content may have been generated or edited by OpenAI models.
The EU's Artificial Intelligence Act requires generative providers to identify generated text in a machine-readable format. The transparency rules took effect on August 2. In the coming weeks, eligible ChatGPT and Codex text within the EU will be watermarked invisibly, covering all packages. The company clarified that text watermarking will not be set as a global default at launch.
The API is globally available from today, limited to certain models, and is turned off by default. Customers will decide whether to use it to meet transparency obligations. Applications for detectors have been opened, initially available only to approved researchers and professional institutions for evaluation and improvement. Public verification tools for images and audio and content traceability interfaces remain open to institutions, unaffected by this text limitation.
The watermark is not a visible mark. The method, called textGrain, leaves an invisible pattern detectable by detectors through subtle word modifications. The signal remains even after the text is copied and pasted. The company stated that the watermark does not identify users, and there is no significant change in model performance when activated. The technical report was co-authored with researchers from the University of Pennsylvania and Yale University. The company also claimed that textGrain meets or exceeds Google DeepMind's SynthID text solution. Anthropic also implemented text identification under the same act in August.
The regional launch is described as a means to collect genuine usage feedback. EU users will encounter it by default, while ChatGPT and Codex in other regions will not. API customers, even outside the EU, can choose to activate it voluntarily.
This is a compliance switch, not a new model. Buyers are the EU regulatory subjects requiring machine-readable identification and voluntarily activating API customers, while the seller is OpenAI embedding signals in the wording. The event is driven by the transparency rules that took effect on August 2. Funding will not shift due to the watermark, and detection rights will initially remain with approved institutions. Beneficiaries are the EU deployers needing to meet obligations, while users relying on copy-pasting to erase sources, with signals still embedded in word order, are under pressure.
Source: Public Information
ABAB AI Insight
Previously, OpenAI only applied public verification to images and audio. The text solution was provided on October 5, after the EU rules took effect on August 2. ChatGPT and Codex will have watermarks by default in the EU within weeks, with the API globally optional and turned off by default. Anthropic had already implemented identification under the same act in August, pointing to DeepMind's SynthID. OpenAI responded with textGrain, claiming it meets or exceeds that solution. However, the detector is not publicly available and is initially given to researchers.
Capital has not entered new business. The cost involves adding a key to the word distribution, with the report co-authored with the University of Pennsylvania and Yale University. The motivation is to meet machine-readable obligations while avoiding global defaults affecting non-EU products. User identities are not carried by the watermark, and copying and pasting cannot remove the signal. API customers can treat the switch as their compliance option, rather than having OpenAI activate it globally.
A comparison is made to camera manufacturers embedding C2PA source information into photo files, while text lacks a file header to write into. OpenAI thus modifies word order without changing the page. The industry is in a regulatory-driven identification phase: public tools for images and audio already exist, while text just began with the EU, and global defaults have been explicitly postponed.
Structurally, this is a regulatory change. The pricing power for generated text remains with the model companies, while identifiability has been transformed into a delivery condition by law. The mechanism hides the signal in the choice of the next word, with detection rights not yet handed over to the public. EU users will have the watermark by default, while other regions will have to wait for API customers to activate it themselves.
ABAB News · Cognitive Law
Text lacks a file header, so the mark is embedded in the wording.
Compliance can start with regional defaults, without needing to activate globally.
Copying and pasting can remove formatting, but cannot eliminate signals in word order.