OpenAI on October 5, 2026, set out its strategy to textual content watermarking beneath the EU AI Act, opening opt-in watermarking to API clients worldwide and asserting that invisible watermarks will probably be added to eligible ChatGPT and Codex textual content output within the European Union over the approaching weeks.
The EU AI Act requires generative AI suppliers to make generated textual content identifiable in a machine-readable means. OpenAI described textual content watermarking and detection as early applied sciences with important limitations, and mentioned its phased strategy displays each the authorized necessities and people limits.
Decide-In API Entry and Regional EU Rollout
From October 5, 2026, API clients globally can decide in to textual content watermarking for choose fashions, and the setting stays off by default. OpenAI mentioned the opt-in design lets clients resolve how watermarking suits their transparency obligations and the experiences they supply to customers, and that it’s working with cloud companions to make watermarking out there for OpenAI mannequin outputs accessed via their providers within the coming weeks.
For its personal merchandise, OpenAI will introduce textual content watermarking for eligible ChatGPT and Codex customers throughout all plans within the European Union solely, additionally within the coming weeks. The corporate mentioned it isn’t making textual content watermarking a world default at launch and that the regional strategy will let it be taught from real-world use and suggestions.
OpenAI additionally opened purposes for entry to its textual content watermark detector on October 5, 2026. Entry will initially be restricted to authorized researchers and professional organizations, granted case by case in accordance with the EU’s Code of Observe on transparency of AI-generated content material, to assist analysis and enchancment of textual content provenance. The announcement applies solely to textual content: OpenAI mentioned its verification instruments for audio and pictures, together with the openai.com/confirm net software and the Content material Provenance API, stay publicly accessible.
How textGrain Works
OpenAI’s watermarking know-how, textGrain, embeds an invisible statistical sign within the phrases a mannequin selects, and the corporate’s detector exams a passage for that sign to guage whether or not it carries an OpenAI watermark. A technical report titled “textGrain: Entropy-Calibrated Watermarking for Language Mannequin Textual content,” dated October 5, 2026, lists authors affiliated with the College of Pennsylvania, Yale College, and OpenAI.
In response to the report, the strategy ties token technology to keyed randomness by fixing an optimum transport drawback whose prices derive from Gumbel random variables, with Kullback–Leibler regularization penalizing departures from independence. An entropy funds caps the common sampling entropy given up in alternate for the watermark sign, and the report states that the KL penalty equals precisely the common entropy the watermark removes, giving the energy parameter a direct information-theoretic which means. Making use of the transport step to keyed blocks of vocabulary tokens cuts the dimensions of the optimization whereas maintaining relative token possibilities inside every block unchanged. The detector wants solely the generated textual content and the key key, and never the entropy funds used throughout technology.
OpenAI mentioned it plans to launch the know-how as open supply so others can construct on it, and that the report will probably be up to date with extra particulars within the coming weeks.
Detection Efficiency and Said Limitations
OpenAI experiences that in its evaluations textGrain matched or exceeded the efficiency of different approaches it examined, together with SynthID for textual content, whereas cautioning that sturdy efficiency beneath ideally suited situations doesn’t assure dependable detection in on a regular basis use.
At a goal false constructive fee of 1%, the detector recognized watermarks in about 80% of 200-token passages and about 95% of 400-token passages for content material reminiscent of psychology, OpenAI reported, with considerably decrease charges for content material reminiscent of arithmetic, the place phrase selection is much less versatile. In an analysis of 400-token passages, changing 10% of phrases with synonyms lowered detection from about 92% to 66%, and changing 25% of phrases lowered it to 17%. The evaluations used watermarked English responses to questions from the ELI5 dataset.
On output high quality, OpenAI mentioned that throughout the benchmarks it makes use of to evaluate Astra, which it describes as its newest frontier mannequin, it doesn’t see significant efficiency variations with and with out watermarking. Its printed desk experiences, for unwatermarked and watermarked textual content respectively, 49.57 and 49.76 factors on the Synthetic Evaluation Intelligence Index and 94.44% and 93.94% on GPQA Diamond. OpenAI mentioned the detection limitations contributed to its choice to limit preliminary detector entry to authorized researchers and professional organizations.
What a Watermark Does Not Set up
OpenAI states {that a} textual content watermark gives a restricted sign in regards to the position its programs performed in a passage. A watermark doesn’t measure human contribution, doesn’t set up possession or accountability, doesn’t determine the consumer, and doesn’t confirm accuracy, the corporate states. It additionally states that the absence of a detected watermark doesn’t show human authorship, as a result of textual content could also be too quick, edited, or translated for dependable detection, could come from an unsupported mannequin, could predate watermarking, or could have been generated by one other firm’s instruments.
The detector will report whether or not it detects an OpenAI watermark with out figuring out the consumer or revealing their prompts or conversations. Given the chance of missed watermarks and false positives, OpenAI isn’t making the software publicly out there at launch.
EU Transparency Obligations
The transparency obligations beneath Article 50 of the AI Act, protecting marking and detection of AI-generated content material and labelling of deepfakes and sure AI-generated publications, apply from August 2, 2026. The Code of Observe on Transparency of AI-generated Content material was drawn up by impartial consultants in a multi-stakeholder course of facilitated by the AI Workplace, with the ultimate code printed June 10, 2026.
Adherence to the code is voluntary, however the Article 50 transparency necessities are authorized obligations. The Fee and the AI Board have confirmed the code is an sufficient voluntary software to exhibit compliance, and about 190 firms and organizations had signed the code by the tip of July 2026, in response to the Fee.
OpenAI mentioned it should proceed enhancing detection, learning how watermarks stand up to modifying and translation, and exploring methods to tell apart AI help from AI authorship, and that it’s going to increase detector entry when it believes outcomes might be interpreted responsibly.
