OpenAI has begun rolling out an invisible, machine-readable text watermarking system for its ChatGPT and Codex models, initially targeting users within the European Union. The move, announced on Monday, October 5, 2026, is aimed at complying with the EU AI Act, according to TechCrunch.
The company's textGrain watermarking technology will be introduced to eligible ChatGPT and Codex users across all plans in the EU over the coming weeks, The Verge reported. OpenAI stated that it is not making text watermarking a global default at launch, opting for a regional approach to learn from real-world use and feedback.
For developers worldwide, OpenAI is enabling an opt-in watermarking feature for select models via its API, effective immediately, according to The Verge. This allows API customers to decide how watermarking aligns with their transparency obligations and the user experiences they provide. OpenAI is also collaborating with cloud partners to make watermarking available for its model outputs accessed through their services in the coming weeks.
Details of the textGrain Watermarking System
The textGrain watermark is not a visible symbol. Instead, it subtly shapes the model's word choices, creating a pattern that is imperceptible to human readers but detectable by a specialized tool, TechCrunch explained. Because the watermark is embedded within the text itself, it remains with the content even when copied and pasted.
OpenAI claims its textGrain watermarking "matched or exceeded" other approaches, including Google DeepMind's SynthID for text, which also forms the basis for a watermarking system announced by Anthropic in August, according to The Verge. OpenAI also published a technical report on its textGrain method, co-written with researchers from the University of Pennsylvania and Yale, TechCrunch reported. The report details how the system uses a secret key to sort next-word predictions, with hundreds of these nudges allowing a detector to identify AI-generated content using only the text and the key.
The company stated that the watermark does not identify the user, and its tests showed no meaningful change in the models' performance when the watermarking was active, TechCrunch reported. The Verge also noted that AI benchmarks indicated similar performance from both watermarked and unwatermarked text.
Limitations and Detector Access
Despite its capabilities, OpenAI acknowledges several limitations of its textGrain system. The company noted that textGrain "does not guarantee reliable detection," and text watermarks do not verify accuracy, determine ownership, measure human contribution, or prove human authorship, according to The Verge.
OpenAI's own tests indicate that editing can make the invisible marks harder to detect. For instance, replacing 10% of words with synonyms in one test caused detection rates to drop from approximately 92% to 66%, TechCrunch reported. The company also found that short passages, mathematical answers, and translated text are more challenging for the detector to identify.
Given these limitations and the risk of missed watermarks and false positives, OpenAI is not making its detector publicly available at launch. Instead, approved researchers and expert organizations can apply for access starting today, The Verge stated. Access will be granted on a case-by-case basis to support the evaluation and improvement of text provenance, in accordance with the Code of Practice. The tool will report the detection of an OpenAI watermark without identifying the user or revealing their prompts or conversations.
OpenAI cautioned that a missing watermark does not prove human authorship, as the text could be too short, heavily edited, or generated by another company's AI, TechCrunch reported. The company added that watermarks "can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it."
Context and Background
The rollout of OpenAI's watermarking system comes in response to the EU AI Act's transparency rules, which took effect on August 2, TechCrunch reported. These rules mandate that AI companies mark AI-generated content in a way that other systems can identify.
This announcement follows a similar move by Anthropic, which stated in August that it would watermark text generated by its Claude model, applying the measure worldwide, TechCrunch reported. Anthropic's decision drew criticism from some Claude users, who argued that they provided the "instructions, context, decisions" while Claude served merely as "the tool."
OpenAI had previously developed a text watermark but chose not to release it, partly due to concerns that users might switch to rival platforms that did not implement such features, The Wall Street Journal reported in 2024, as cited by TechCrunch. However, OpenAI, alongside Anthropic, Google, Meta, and Microsoft, are among the companies that have committed to following the EU's code of practice on AI-generated content, TechCrunch noted.
