If you could leave an invisible mark on sentences ChatGPT writes and look for that signal later, how much could you learn? OpenAI's text watermarking technology textGrain doesn't attach extra symbols to sentences; it holds an invisible signal in the statistical pattern of how the model chooses words. Where it applies differs between ChatGPT and Codex in the EU and the opt-in API setting worldwide, and a detector result doesn't identify who wrote a piece of text.
OpenAI introduced textGrain as part of its text provenance approach under the EU AI Act. Along with Content Credentials and SynthID for images and audio, it is one of several signals that help people understand where generated content came from. You can see an example of marking the source of video content in the post on Google Vids and SynthID labeling as well.
textGrain puts a statistical signal in word choice
When generating a sentence, a model picks one among several words and word fragments that could come next. textGrain applies a tiny adjustment, set by a secret key, to the probability distribution of that choice. It isn't a mark you'd notice from one or two words; it looks for a statistical signal in the pattern of choices across the whole text.
So when you copy the text, no hidden spaces or special characters come along, and the watermark isn't visible to readers. The detector checks for this pattern and reports whether an OpenAI watermark is detected. OpenAI's help article explains that for supported models, a watermark can be put in during generation without adding any separate mark to the API response.
In its official announcement, OpenAI said textGrain showed detection performance equal to or higher than the other methods it tested. It added, though, that performance under ideal conditions is not guaranteed to carry over with the same reliability to everyday text.
Source: X @OpenAI
EU ChatGPT and Codex differ from the API
OpenAI's announcement distinguishes how it applies by product. In the EU, it said it will add the invisible watermark to eligible ChatGPT and Codex text output over several weeks. This is a phased rollout aimed at the EU at launch, and it doesn't mean it is applied by default in ChatGPT and Codex worldwide.
The API is separate. API customers worldwide can choose supported models and turn on text watermarking, but it is off by default. According to OpenAI's help article, you can allow it in project settings or organization settings and select models. Which models are currently supported is something you check in the model list on the settings screen.
Turning on text watermarking and getting permission to use the detector are separate things. Detector applications have opened for researchers and professional organizations, and at first access is provided by reviewing approved applications case by case.
Source: X @OpenAI
The published detection figures are results under specific test conditions
Detection performance is affected by text length and subject. In one evaluation OpenAI presented, with the false positive rate set to 1%, it found the watermark in psychology-related content about 80% of the time in 200-token texts and about 95% of the time in 400-token texts. For content with a narrow range of word choices, such as math, the detection rate was considerably lower.
Changes from editing were tested on 400-token English responses too. When 10% of the words were replaced with synonyms, the detection rate fell from about 92% to 66%, and when 25% were replaced, it dropped to 17%. These values come from that length, language, and set of evaluation questions. They can't be treated as universal rates that apply as they are to other lengths, languages, subjects, or editing methods.
Short texts give less data to evaluate the signal, and math answers or requests to reproduce a given sentence verbatim leave little room for word choice. Heavy rewriting or translation can also weaken the pattern. So a result of "not detected" alone can't tell you a person wrote it, and "detected" doesn't let you conclude that the whole text was written by AI.
Source: X @OpenAI
What the detector does and doesn't tell you
The detector's result is limited information about whether an OpenAI watermark signal was detected. OpenAI explains that the result may indicate the possibility that its systems generated or processed part of that text. But it doesn't measure how much human judgment, editing, or creative contribution there was.
The result also doesn't identify a user, an organization, a prompt, or a conversation. It doesn't determine ownership of the text, the legality of its use, disclosure obligations, or whether it is factual. OpenAI cited the possibility of false positives and false negatives and the need for responsible interpretation of results as reasons it started detector applications and access with researchers and professional organizations.
Provenance technology isn't only for text. Just as each product has a different use, like the agent platform introduced in the OpenAI dots post, content provenance combines different signals and tools depending on the format. textGrain is the signal carried in a text's word choice, and what the detector answers is limited to that.
What to remember when reading provenance signals
textGrain is a statistical watermark that readers can't see. API customers can choose to turn it on for supported models, and it is planned to roll out in phases to eligible ChatGPT and Codex output in the EU. The detector is provided on a limited basis after application and review, and the published detection figures assume specific evaluation conditions.
Above all, a watermark is not a tool for determining authorship or measuring how much a person contributed. Understand detection as one signal about origin, and look at the text's writing process and context along with it.