OpenAI is adding invisible textGrain watermarks to ChatGPT in the EU, but unlike Anthropic, it will allow API customers worldwide to opt out. Tests show detection rates as high as 95 percent, dropping to 17 percent when a quarter of the words are replaced.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
OpenAI will enable the textGrain invisible watermark for ChatGPT and Codex users in the EU in the coming weeks to comply with the EU's machine-readable labeling requirement for generated text; global API users can choose whether to enable it. But according to OpenAI's own data, this watermark's detection rate is highly dependent on text length and topic, and can be substantially reduced by simple rewriting.
OpenAI will enable an invisible watermark called textGrain for ChatGPT and Codex users in the EU in the coming weeks to comply with the EU's requirement to label generated text in a machine-readable format. The direct driver of this move is compliance, not product feature iteration.
Unlike the mandatory enablement in the EU, OpenAI's API watermark is opt-in worldwide and will be available through cloud partners such as Microsoft Azure in the coming weeks. This means the same company's watermark strategy is not consistent across the two lines of end users and developers.
By contrast, Anthropic's watermark for Claude is mandatory and applies globally regardless of how users access the model. OpenAI's choice to retain choice on the API side creates a clear divergence from Anthropic's approach.
textGrain works by embedding an invisible statistical signal in the model's word choices, similar in principle to the SynthID watermark that Claude uses from Google's open-source technology. This is a statistical-level marking, not the insertion of visible characters into the text.
OpenAI says textGrain matches or outperforms other approaches in internal testing, including Google's SynthID text watermark, and plans to open-source the technology. However, these comparison results come from OpenAI's own internal testing and have no third-party verification.
Whether a watermark can be detected first depends on text length. When the detector's false positive rate is set to 1%, about 95% of watermarks can be identified in a 400-token psychology text, dropping to about 80% at 200 tokens. Halving the length reduces the identification rate by about 15 percentage points.
Topic is equally critical. OpenAI says the detection rate for mathematical content is "significantly lower" because the model has less freedom in word choice. A chart in the source shows that the identification rate for a 400-token mathematical text is about 60%, a stark contrast with the about 95% for psychology text.
OpenAI suggests that longer text may strengthen the statistical signal and improve detection, but provides no data to verify this hypothesis. Actual results may still vary by content, a point that cannot be ignored when assessing watermark reliability.
The watermark has limited resistance to editing. OpenAI says that replacing just 10% of words with synonyms drops the detection rate for a 400-token text from about 92% to 66%.
Replacing a quarter of the words drops the detection rate to 17%, making the watermark easy to evade. This means any user intent on hiding AI-generated traces only needs moderate rewriting to substantially reduce the probability of detection.
By comparison, Anthropic has not published the detection rate for Claude's watermark, only saying its system has good resistance to editing. The difference in transparency between the two companies makes it difficult for outsiders to directly compare the actual strength of the two watermarks.
OpenAI says the watermark does not harm output quality, citing no significant performance difference with the watermark on and off for its frontier model Astra across eight benchmarks including GPQA Diamond, BrowseComp, and DeepSWE.
But these benchmark results cannot determine whether the watermark affects writing quality, which is also the question critics have raised about Claude's watermark. Benchmarks cover capability metrics, not style, fluency, or creative expression.
OpenAI also clearly delineates the boundaries of what the watermark means: detecting a watermark does not indicate how much human creativity or editing the text contains, nor does it establish ownership, assign responsibility, identify users, or indicate whether the text is accurate. Conversely, failure to detect a watermark does not prove the text was written by a human, because the text may be too short, edited or translated, or from a model the detector does not support.
The textGrain detector is initially open only to selected researchers and professional institutions, who can apply via an application form; OpenAI will grant access on a case-by-case basis under the EU Code of Practice. This is similar to Anthropic's approach to its detection API.
The tool only reports whether an OpenAI watermark is detected; it does not identify users or disclose their prompts and conversations. This design is intended to isolate detection capability from user privacy.
OpenAI says access is restricted because the detector may mislabel unwatermarked text or miss watermarks, and plans to expand access "when it believes results can be responsibly interpreted," but gives no specific timeline. The risks of mislabeling and missed detection are precisely why detection results need to be interpreted with caution.
Existing verification tools for images and audio remain publicly available, including openai.com/verify and the Content Provenance API. The controlled opening of the text watermark detector contrasts with the public availability of the image and audio tools.
This article is based on reporting by THE DECODER, in which OpenAI's deployment plan, API opt-in, Anthropic's mandatory policy, and the controlled opening of the detector are confirmed facts; the detection rate figures, the decline after rewriting, the benchmark conclusion of no quality harm, and the boundaries of the watermark's meaning all come from OpenAI's statements and are source claims without independent third-party verification. OpenAI provided no data to verify the hypothesis that "longer text improves detection," nor did it specify when detector access will be expanded.
textGrain is a compliance tool deployed by OpenAI to meet the EU's machine-readable labeling requirement, not a reliable means of content provenance. Its detection rate declines as text gets shorter and topics turn mathematical, and it can be substantially weakened by simple synonym replacement; the controlled opening of the detector and the stated boundaries of its meaning further show that the watermark can only serve as a limited signal and cannot alone be used to judge text origin or the degree of human involvement.
Primary report
Primary source