OpenAI text watermarking is coming to eligible ChatGPT and Codex output in the European Union over the coming weeks. OpenAI says the move responds to the EU AI Act requirement for generated text to be identifiable in a machine-readable way.
The system is called textGrain. Instead of adding visible labels, hidden characters, unusual spaces, or special punctuation, it changes the statistical pattern of the model's word choices. A detector can then look for that pattern in a passage.
What OpenAI's text watermark actually does
When a model generates text, it has several possible words or word pieces to choose from at each step. textGrain subtly adjusts those choices according to a secret pattern. Across a long enough passage, the choices can form a signal that a matching detector can recognize.
That makes the system different from a normal AI-text classifier. A classifier tries to infer whether text looks AI-generated after the fact. OpenAI's watermark is embedded during generation, so detection is based on a provenance signal created by the model itself.
OpenAI says the watermark does not add hidden characters or watermark-only tokens, and ordinary readers will not see anything different in the text. Its provenance documentation describes textGrain as an invisible change to the randomness used in word selection. OpenAI's provenance guidance explains the mechanism and its limits.
The EU rollout is deliberately limited
OpenAI says eligible ChatGPT and Codex text generated in the EU will receive the watermark over the coming weeks, across all ChatGPT plans. It is not making text watermarking a global default at launch.
API customers are getting a different option. Customers globally can opt in to text watermarking for select models, with the feature off by default. OpenAI also says it is working with cloud partners to extend watermarking to eligible outputs accessed through their services.
| Area | OpenAI's current approach |
|---|---|
| ChatGPT and Codex in EU | Watermarking rolling out |
| API customers | Global opt-in for select models |
| API default | Off |
| Detector access | Approved researchers and expert organizations |
| Public text detector | Not available at launch |
| Technology | textGrain |
OpenAI's own testing shows the watermark has limits
The most important part of the announcement may be what happens when the text is edited.
At a target false-positive rate of 1%, OpenAI says its detector identified watermarks in about 80% of 200-token psychology passages and about 95% of 400-token passages. Detection was substantially weaker for mathematics, where there is less freedom to change wording.
Editing can weaken the signal quickly. In OpenAI's test of 400-token passages, replacing 10% of the words with synonyms reduced detection from about 92% to 66%. Replacing 25% reduced detection to 17%.
OpenAI says a detected watermark can indicate that an OpenAI system generated or processed part of a passage. It does not establish who wrote the text, how much a human contributed, who owns it, or whether the text is accurate.
A missing watermark does not prove human writing
OpenAI is also explicit about the other side of the detection problem. If a detector finds no watermark, that does not prove a person wrote the passage.
Short answers may not contain enough text for reliable detection. Code and highly constrained factual answers can also give the model too little freedom to alter word choices. Translation, substantial paraphrasing, and other major edits can further weaken the signal.
That distinction matters when comparing provenance signals with AI-writing detectors. A provenance watermark answers a narrower question: whether a supported OpenAI signal can be found in the text. It does not provide a complete history of how the passage was written.
Saganote's Claude text watermark explained coverage examines the same provenance problem from Anthropic's side, while Google's SynthID watermark approach shows another model-maker's approach to identifying generated content.
OpenAI says watermarking does not reduce model quality
OpenAI reports that its tests on Astra showed no meaningful performance difference between watermarked and unwatermarked text across the benchmarks it uses to evaluate the model. The company also says text watermarking has a negligible impact on model speed.
| Benchmark | Unwatermarked | Watermarked |
|---|---|---|
| Artificial Analysis Intelligence Index | 49.57 | 49.76 |
| AutomationBench | 34.09% | 34.86% |
| DeepSWE v1.1 | 72.80% | 71.68% |
| Terminal-Bench 4.0 | 53.90% | 56.06% |
| BrowseComp | 87.92% | 87.35% |
| GPQA Diamond | 94.44% | 93.94% |
Text joins OpenAI's broader provenance system
OpenAI already uses provenance signals for other media. Supported images can carry Content Credentials and SynthID, while supported generated audio uses SynthID. The new text system extends that layered approach to another content type.
The broader OpenAI product line matters too. ChatGPT Images 2.5 adds another major generated-content workflow, while GPT-Live shows why provenance increasingly has to cover more than plain text.
OpenAI is also expanding ChatGPT into business workflows. Its Data Agent turns questions into Power BI and Tableau dashboards, showing how generated and transformed information is becoming part of practical work.
The company's wider platform strategy includes advertising as well. Saganote's coverage of ChatGPT Ads tracks that expansion, although advertising and provenance are separate systems.
Provenance is not the same as authorship
OpenAI's most useful warning is that provenance signals should be treated as evidence about a model's involvement, not as a complete authorship record.
A watermark cannot tell a reviewer whether a person wrote an outline and an AI polished it, whether an AI drafted a passage that a person heavily edited, or whether the model generated the whole passage. It also does not identify the user, account, prompt, or conversation associated with the text.
That limitation will matter in schools, workplaces, publishing, and research, where the important question is often not simply whether AI was involved, but how AI was used.
What happens next
OpenAI plans to continue testing how textGrain performs under editing and translation and says it intends to make the technology open source. Detector access will initially remain with approved researchers and expert organizations so OpenAI can study reliability and responsible use before broader access.
For now, the EU rollout is best understood as a provenance signal rather than a universal AI detector. OpenAI is adding a machine-readable marker to eligible generated text, but its own results show that text length, subject matter, and editing can all affect detection.
Frequently Asked Questions
What is OpenAI text watermarking?
Will every ChatGPT response have a watermark?
Can the watermark prove that AI wrote an entire document?
Can editing remove the watermark?
Can people see the watermark?
OpenAI's official EU text provenance announcement provides the technical results and rollout details. OpenAI's provenance guidance explains how textGrain fits with its image and audio provenance systems.
