OpenAI is putting invisible watermarks into the text it generates for ChatGPT and Codex, but only for users in the European Union at first. The feature is called textGrain, and it is rolling out over the coming weeks. OpenAI says it will not become the default setting globally at launch.
The rollout is part of the company’s response to the EU’s AI Act. OpenAI is not alone in this — Anthropic also announced a watermarking plan in August, based on the same underlying technology.
How the rollout works
The watermark is invisible to the naked eye but readable by detection software. OpenAI says it is introducing textGrain to eligible ChatGPT and Codex users across all plans in the EU only. The company is not making text watermarking a global default at launch. Instead, it is taking a regional approach so it can learn from real-world use and feedback.
Over the coming weeks, the feature will roll out to eligible users. OpenAI says the EU-only start gives it room to gather data from actual use before deciding whether to expand.
The comparison to SynthID
OpenAI says its textGrain watermarking “matched or exceeded” other approaches like Google DeepMind’s SynthID for text. That comparison matters because SynthID is also the basis for the watermarking Anthropic announced in August.
Both companies made the move to meet the requirements of the EU’s AI Act. However, not everyone was happy to learn about the addition.
What the watermark does not do
OpenAI is quick to say what the watermark does not do. The company notes that textGrain “does not guarantee reliable detection,” and it lists several things the watermark cannot prove:
- It does not verify the accuracy of the text.
- It does not determine who owns the text.
- It does not measure how much a human contributed.
- It does not prove human authorship.
“textGrain does not guarantee reliable detection”
OpenAI frames the watermark as a transparency tool, not a verification system. That distinction matters for anyone who might want to use it to settle ownership disputes.
Who gets to opt in
API customers around the world will be able to opt in to watermarked text outputs for select models starting today. That means enterprise users can choose whether to include the mark in documents produced by OpenAI’s models.
Cloud partners are also being brought into the picture. OpenAI is working with them to make watermarking available for OpenAI model outputs accessed through their services in the coming weeks.
Detector access for researchers
Approved researchers and expert organizations can apply for access to the detection tool starting today. The tool reports whether it detects an OpenAI watermark without identifying the user or revealing their prompts or conversations.
Access is being granted on a case-by-case basis to support evaluation and improvement of text provenance. OpenAI is not making the tool publicly available at launch.
The practical stakes
The rollout is a live test of whether text watermarks hold up in real use. OpenAI is gathering data from actual use before deciding whether to expand.
If the watermark fails to detect itself consistently, the whole exercise raises questions rather than answers.
For now, the watermark is a curiosity. It is a technical solution to a legal requirement.
Source material: “OpenAI is adding text watermarking in ChatGPT and Codex,” The Verge.
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.