Brussels/San Francisco, October 7, 2026: OpenAI has announced that it will introduce invisible text watermarks to eligible ChatGPT and Codex outputs in the European Union as part of its compliance with the EU AI Act. The rollout will happen over the coming weeks, while API customers globally can already opt in to watermarked text for select models.
New System Called textGrain
OpenAI’s watermarking technology, called textGrain, adds an invisible statistical signal to the model’s word choices.
A dedicated detector can then analyse a passage and assess whether it contains an OpenAI watermark. The watermark itself is not visible to users reading the text.
Rollout Limited to EU at Launch
OpenAI says the watermark will initially be added to eligible ChatGPT and Codex text outputs only in the European Union.
The company is not making watermarking the global default at launch, saying a phased regional rollout will allow it to study how the technology performs in real-world conditions.
EU AI Act Requires Machine-Readable Identification
The move is linked to transparency requirements under the EU AI Act, which requires providers of generative AI to make AI-generated text identifiable in a machine-readable form.
OpenAI says text watermarking is one way to meet those obligations while acknowledging that the technology still has limitations.
API Customers Can Opt In Globally
Starting this week, OpenAI API customers around the world can choose to enable text watermarking for select models.
Unlike the EU ChatGPT and Codex rollout, watermarking remains off by default for API customers.
Watermark Does Not Prove Authorship
OpenAI has cautioned that the presence of a watermark does not prove who wrote, submitted or published a particular passage.
It only indicates that the text contains a statistical signal associated with supported OpenAI-generated output.
Editing Can Make Detection Harder
Text watermarking is not foolproof.
OpenAI says short text, heavy editing, translation and rewriting can make a watermark more difficult to detect, while detectors may also produce false positives or false negatives.
Detector Access Will Be Restricted Initially
OpenAI is opening applications for access to its watermark detector, but the tool will initially be limited to approved researchers and expert organisations.
The company says it is not launching a public detector yet because watermark detection can still produce errors and needs to be interpreted carefully.
Wider Push for AI Content Provenance
The text watermarking initiative forms part of OpenAI’s broader effort to make AI-generated content easier to identify.
The company already uses provenance technologies such as Content Credentials, C2PA and SynthID for supported image and audio outputs.
Source – OpenAI, The Verge
Supreme News Network
