OpenAI is gradually adding invisible, machine-readable text watermarking to ChatGPT and Codex for users in the European Union, covering all plans over the next few weeks. API users worldwide can opt in to watermarking starting today on certain models.
This watermark does not confirm that the text is accurate, who owns it, or whether it was written by a human. It also cannot indicate how much human involvement was involved. The detector will initially be available only to approved researchers and qualified organizations because it may still miss detections and produce false positives.
What’s happening to text from ChatGPT and Codex
OpenAI is gradually adding an invisible, machine-readable text watermark called textGrain for eligible ChatGPT and Codex users in the European Union, covering all plans over the next few weeks.
API customers worldwide can opt in to watermarking today on certain models, allowing each service to decide how to use watermarks in line with its transparency requirements and user experience.
Why invisible watermarking has become important
As AI-generated text is used in real-world work, users and organizations need to know where the text came from. The textGrain watermark helps with text provenance and addresses the transparency requirements of the EU AI Act.
However, the watermark does not confirm that the text is accurate, who owns it, or how much human assistance was involved in writing it. OpenAI also states that it may miss detections or produce false positives, so it should be treated as a supporting signal for verification rather than a final answer.
Where textGrain fits into OpenAI’s approach
textGrain is an invisible, machine-readable watermarking system for text outputs from OpenAI. It is initially being used with ChatGPT and Codex in the European Union.
On the API side, customers worldwide can opt in on some models. It is not yet enabled by default globally for general users because OpenAI wants to collect feedback from real-world use first.
From ordinary text to watermarked text
OpenAI states that results from AI benchmark tests for watermarked and non-watermarked text showed similar performance, suggesting that watermarking does not clearly reduce text quality.
However, textGrain does not guarantee consistently reliable detection because it may miss watermarks or produce false positives. The watermark therefore helps establish the origin of text but does not confirm its accuracy or the degree of human involvement.
What it looks like in practice across different channels
ChatGPT and Codex users in the European Union will begin receiving textGrain over the next few weeks, covering all plans. It is not yet enabled by default worldwide.
API customers worldwide can opt in to watermarked text starting today, with support limited to certain models. This is suitable for organizations that need to manage their own transparency requirements.
Users accessing OpenAI models through cloud partner services will begin gaining access to watermarking over the next few weeks, depending on support from each partner.
OpenAI compared with SynthID and Anthropic’s approach
OpenAI states that textGrain “matched or exceeded” other approaches such as Google DeepMind’s SynthID, which forms the basis of the text watermarking system Anthropic announced in August.
Both OpenAI and Anthropic link the addition of text watermarking to the requirements of the EU AI Act. However, this does not mean that textGrain can detect watermarks accurately in every case. OpenAI itself states that the system does not guarantee consistently reliable detection.
Strengths and limitations to understand before interpreting the results
textGrain supports the verification of text provenance and allows API customers to choose whether to enable watermarking based on their needs.
However, it should not be interpreted as conclusive evidence because it may miss watermarks or produce false positives. It also does not confirm accuracy, ownership, the amount of human involvement, or whether the text was written by a human.
Pros
- +Supports verification of text provenance
- +API customers can choose whether to enable watermarking
Cons
- −Does not guarantee reliable detection
- −May miss watermarks or produce false positives
- −Does not confirm accuracy, ownership, or whether the text was written by a human
The trade-offs of a system like this
| Cost/Risk | Impact |
|---|---|
| Missed detections or false alerts | Results may be uncertain and cannot be used to conclusively determine provenance |
| Limited access to the detector | Researchers and qualified organizations must apply and receive approval on a case-by-case basis |
| Burden on API customers | They must decide whether watermarking fits their transparency obligations and user experience |
| Information reported by the detector | It only indicates whether an OpenAI watermark was found, without revealing the user or conversation |
Therefore, the real cost is not a purchase price but managing uncertainty and designing an appropriate usage process for each service.
Who will get access to the detector, and what happens next
Approved researchers and qualified organizations can apply for access to the detector. OpenAI will review applications case by case for evaluating and improving text provenance verification.
The results will only indicate whether an OpenAI watermark was found. They will not identify the user or disclose the prompt or conversation. It is not currently available to the general public because of the risk of missed detections or false positives.
What the watermark can and cannot tell you
textGrain can help determine whether text contains an OpenAI watermark. Benchmark results for watermarked and non-watermarked text are similar, so it can be used as part of a text provenance review.
However, the watermark does not guarantee successful detection, confirm accuracy or ownership, indicate the proportion of human involvement, or prove that the text was written by a human.
The newer perspective is that textGrain should be viewed as a tool supporting provenance verification, not as conclusive evidence of a text’s quality, ownership, or human origin. We will need to continue monitoring real-world results and feedback from the European Union.