OpenAI to Add Invisible Watermarks to ChatGPT and Codex Text in EU

The watermark is not a visible symbol. OpenAI's system, textGrain, adds what the company calls an invisible statistical signal to the model's word choices. A detector looks for that signal to assess whether a passage contains an OpenAI watermark. Because the signal is carried in word choices, it travels with the text when copied and pasted, TechCrunch reported. OpenAI published a technical report about the method, and TechCrunch reported it was co-written with researchers from the University of Pennsylvania and Yale. OpenAI says it plans to make the technology available in open source so others can build on it. Engadget described textGrain as a proprietary system.

Detection access will initially be limited to approved researchers and expert organizations. OpenAI said the detector does not identify the user or reveal prompts or conversations. The company also said a watermark does not measure human contribution, establish ownership or responsibility, identify the user, or verify accuracy, and that the absence of a detected watermark does not prove human authorship. The text could be too short or too heavily edited, or it could come from another company's AI, OpenAI said.

OpenAI acknowledged that detection remains unreliable in some cases. Shorter text is harder to detect, with around an 80 percent success rate, Engadget reported. Content such as mathematics is also difficult because there is less flexibility in word choice. Editing weakens the signal: in an evaluation of 400-token passages, replacing 10 percent of words with synonyms reduced detection from about 92 percent to 66 percent, while replacing 25 percent of words reduced it to 17 percent. Translated text and heavily edited passages are also harder to detect. OpenAI said textGrain matched or exceeded other approaches such as SynthID for text, but cautioned that strong performance under ideal conditions does not guarantee reliable detection in everyday use.

OpenAI said its comparison of watermarked and unwatermarked text found no meaningful difference in overall performance, and that watermarked outputs scored slightly higher on several benchmarks. The company said it expects to revisit its approach as technology, standards, and evidence evolve, and will continue studying how well watermarks survive editing and translation and whether they can more meaningfully distinguish AI assistance from AI authorship.

The EU AI Act's transparency rules took effect on August 2, TechCrunch reported. Article 50 requires providers of generative AI systems to make text outputs identifiable in a machine-readable way. The requirement already applies to new companies entering the market, but pre-existing entities have until December 2 to comply. That group includes OpenAI, Anthropic, Microsoft, Google, and Meta, among others. Anthropic said earlier this year that it would watermark text generated by Claude worldwide, a move that drew backlash from some users who argued they supplied the instructions, context, and decisions while Claude was just the tool. Anthropic also developed a detector for its own watermark, which cannot determine whether text came from OpenAI or another competitor because it uses a secret key specific to Anthropic's system.

OpenAI had built a text watermark before but held off on releasing it, partly over concerns that users would switch to rivals that did not watermark, The Wall Street Journal reported in 2024. Anthropic, Google, Meta, Microsoft, and OpenAI are among the companies that have committed to following the EU's code of practice on AI-generated content. OpenAI's watermarking rollout is narrower than Anthropic's worldwide approach: the ChatGPT and Codex watermark will apply only in the EU for eligible users, while API customers elsewhere may choose to turn it on.

Editor's Summary OpenAI's EU watermarking plan will add invisible signals to eligible ChatGPT and Codex output in the EU and allow optional API use elsewhere. The textGrain system is designed to detect some OpenAI-generated text, but OpenAI says detection is unreliable for short, edited, translated, or math-heavy content and does not establish authorship. The move follows Anthropic's global watermarking decision and comes as major AI providers face a December 2 compliance deadline.