AI
“OpenAI will begin adding an invisible watermark to eligible ChatGPT and Codex text output in the European Union over the coming weeks, as part of its response to the EU AI Act”
Plain restatementOpenAI has stated that, within the coming weeks, it will apply an invisible watermark to eligible ChatGPT and Codex text output for users in the European Union, as part of its response to EU AI Act transparency requirements.
This claim is accurate. OpenAI published a statement on October 5, 2026 saying it will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union over the coming weeks, as part of its response to the EU AI Act. The post's wording closely matches OpenAI's own. The system is called textGrain and it works by nudging the model's word choices to carry a hidden statistical signal that OpenAI's detector can look for. OpenAI itself published the limitations: detection is far weaker on short text and on content like mathematics, and in one of its tests replacing a quarter of the words with synonyms cut detection from about 92 percent to 17 percent. Two things to keep in mind: the EU rollout had not yet happened at the time of the announcement, it was a stated plan with no fixed date, and every performance number comes from OpenAI's own testing, with detector access currently limited to approved researchers, so no independent party has verified how well it works.
OpenAI will [drifted from the evidence:] begin adding an invisible watermark to eligible ChatGPT and Codex text output in the European Union [drifted from the evidence:] over the coming weeks, as part of its response to [drifted from the evidence:] the EU AI Act
OpenAI [added by the neutral restatement:] has stated that, within the coming weeks, it will [added by the neutral restatement:] apply an invisible watermark to eligible ChatGPT and Codex text output [added by the neutral restatement:] for users in the European Union, as part of its response to EU AI Act [added by the neutral restatement:] transparency requirements.
Red-tinted words in the claim drifted from the evidence. Green-tinted words are what a neutral restatement needs.
The trace / claim to source
- The core sentence matches OpenAI's own published wording almost word for word. OpenAI said it will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union over the coming weeks.
- The EU AI Act framing is correct and is OpenAI's own framing, not an outside interpretation. OpenAI presents the move as a response to EU AI Act requirements on machine-readable identification of generated text.
- The post's quoted excerpt of the three bullet points is an accurate reproduction of the official statement, including that API watermarking is opt-in and off by default globally, and that detector access starts with approved researchers and expert organizations.
- The system name textGrain is correct, and the description of it as adding a statistical signal to word choices that OpenAI's detector can identify matches OpenAI's own description.
- The claim that detection works better on longer passages is supported. OpenAI reports roughly 80% detection on 200-token passages versus roughly 95% on 400-token passages for psychology content at a 1% target false positive rate.
- The figure cited in the caption is correct. OpenAI reports that replacing 25% of words with synonyms reduced detection from about 92% to 17%.
- The statement that the post's scope is text only, with image and audio verification tools remaining publicly accessible, matches the official post.
- The caption gives the 92% to 17% figure without the conditions attached to it in the source. OpenAI's numbers apply to 400-token English passages, drawn from the ELI5 dataset, at a target false positive rate of 1%. The post does say "in one test," which preserves the fact that this is a single evaluation condition, so the omission narrows context without changing the meaning.
- The caption's phrasing that "ChatGPT watermarking will remain limited to the EU" is a reasonable reading of the announcement but is slightly firmer than the source. OpenAI states where it is rolling the watermark out, not that it will never extend it elsewhere. The word "remain" implies a settled permanent boundary that the official post does not assert.
- Whether the EU rollout to ChatGPT and Codex has actually begun. As of 2026-10-06 the announcement is one day old and OpenAI gave no specific date beyond "over the coming weeks," so this is a stated commitment rather than a shipped feature.
- Which exact model versions count as "eligible" for ChatGPT and Codex watermarking, and which "select models" support the API opt-in. OpenAI's announcement does not enumerate them.
- Whether the reported detection rates hold outside OpenAI's own evaluation. All performance figures come from OpenAI's internal testing, and no independent evaluation has been published. Because detector access is restricted to approved researchers and expert organizations, independent verification is currently constrained by design.
- Whether textGrain genuinely matched or exceeded Google's SynthID for text. That comparison is OpenAI's own and was not run by a neutral third party.
- Whether this implementation satisfies EU AI Act transparency obligations as a legal matter. OpenAI describes it as its response to the requirement, which is a company position, not a regulator's determination.
The claim is a near-verbatim restatement of OpenAI's own published announcement. OpenAI's October 5, 2026 post states that the EU AI Act requires generative AI providers to make generated text identifiable in a machine-readable way, that text watermarking and detection remain early technologies with significant limitations, and that its phased approach reflects both the EU AI Act requirements and the technology's limitations. The post sets out three steps: starting that day, API customers globally can opt in to text watermarking for select models, with watermarking off by default in the API; over the coming weeks, an invisible watermark will be added to eligible ChatGPT and Codex text output in the European Union; and applications are open for access to the text watermark detector, with access initially limited to approved researchers and expert organizations. On the system itself, OpenAI states that its text watermarking technology, textGrain, adds an invisible statistical signal to the model's word choices, and its detector looks for that signal to assess whether a passage contains an OpenAI watermark. The accompanying technical report describes the mathematical details of the textGrain watermark, specifying generation and detection procedures and explaining how to set a budget for sampling entropy lost in exchange for watermark signal. On performance, OpenAI reports that at a target false positive rate of 1%, its detector identified watermarks in about 80% of 200-token passages compared with about 95% of 400-token passages for content such as psychology, with detection rates substantially lower for content such as mathematics where there is less flexibility in word choice. On editing, OpenAI reports that in an evaluation of 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%, and replacing 25% of words reduced it to 17%. OpenAI also states that these limitations contributed to its decision to provide initial detector access only to approved researchers and expert organizations, and that the results shown are for watermarked responses to mathematics and psychology questions from the ELI5 dataset at a target false positive rate of 1%.
Complete reasoning
The reply is formatted for pasting into the thread where the claim is circulating.
Compact share page: ai.trueseeker.com/s/af576091b1e9/HaC3h9Dy5ubQhoZOK7POirrmoj4