OpenAI Opens Optional Text Watermarking for Select API Models
The opt-in textGrain signal is available globally through the API, while EU ChatGPT and Codex coverage and a restricted detector will roll out separately.
Edited by Tyronne Panaino
OpenAI made opt-in text watermarking available to API customers worldwide on October 5 for select models. The company calls the system textGrain and says the option is off by default, giving developers a new provenance control without changing every API response.
The release matters to teams that publish or process AI-generated text because it adds a machine-detectable signal at generation time. It does not make reliable identification automatic: OpenAI's own reported evaluations show that short passages, constrained subject matter and later editing can substantially weaken detection.
API customers get the first live option
The OpenAI announcement says API customers can opt in globally starting October 5. Availability is limited to select models, and the watermark remains disabled unless a customer turns it on. That makes the immediate change a developer-controlled capability rather than a universal marker across OpenAI text.
OpenAI has not paired the API opt-in with open public access to the text detector. Applications are open, but initial detector access is restricted to approved researchers and expert organizations that the company says will help evaluate reliability and responsible use. Public verification tools for supported OpenAI images and audio remain separate and continue to be available.
ChatGPT and Codex coverage is an EU rollout
OpenAI says eligible ChatGPT and Codex text output in the European Union will receive an invisible watermark over the coming weeks. The company presents that phased rollout as a response to text-provenance requirements under the EU AI Act, while acknowledging that text watermarking and detection remain early technologies.
The timing language is important. API opt-in is available now, but the consumer and coding-product change is still rolling out. The announcement also describes only eligible output, so it should not be read as evidence that every model, response type, language or product surface is already covered.
textGrain adds a statistical signal to word choices
OpenAI describes textGrain as an invisible statistical signal introduced through a model's word choices. A detector then looks for that signal in a passage. The company says it plans to make the technology open source, but the fetched announcement does not set a release date for that code.
This design differs from attaching visible labels or metadata to a document. The signal is carried in the generated text itself, which may help it survive copying, but it also depends on the model having enough freedom to choose among words. OpenAI says detection is harder when text is short or constrained and explicitly notes that detectors can produce both false positives and false negatives.
OpenAI's tests show sharp limits
At a target false-positive rate of 1%, OpenAI reports that its detector found the watermark in about 80% of 200-token psychology passages and about 95% of 400-token passages. The company says detection was substantially lower for mathematics, where word choice is more constrained. These are vendor-reported results, not an independent validation in the evidence reviewed for this article.
Editing was another major weakness in OpenAI's evaluation. For 400-token passages, replacing 10% of words with synonyms reduced reported detection from about 92% to 66%. Replacing 25% reduced it to 17%. Those results mean a missing signal cannot establish that text was written by a person, while a detected signal still needs to be interpreted within the system's stated false-positive rate and scope.
What affected teams should watch
API developers now have a choice about whether to add the signal, but they still need policies for disclosure, retention and detector interpretation. EU users of ChatGPT and Codex should expect eligible output to change over the coming weeks rather than assume the watermark is already present. Researchers and expert organizations can apply for detector access, while broader text-verification availability remains unresolved.
The next verifiable checkpoints are the exact model and language coverage, the start and completion of the EU product rollout, access terms for the detector, publication of the planned open-source technology and independent testing against paraphrasing, translation and real-world editing.
Status
Confirmed. OpenAI has opened the API opt-in and announced the phased EU rollout. Internal confidence is medium because the evidence is a single first-party announcement and the performance results have not been independently verified in the fetched record.
Sources
Update note: Last reviewed 2026-10-05. We will revise this post when OpenAI identifies exact model coverage, begins the EU product rollout, releases textGrain code or expands detector access.
Sources
- OpenAI — EU text provenance rules — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.