OpenAI will begin watermarking ChatGPT and Codex text in the European Union to meet AI Act transparency rules

What happened

OpenAI announced Monday that it will start placing an invisible watermark in text produced by ChatGPT and Codex, but only for users in the European Union at first, TechCrunch and The Verge reported. The company said the move is meant to comply with the EU AI Act. OpenAI said it is not making text watermarking a global default at launch.

According to TechCrunch, the AI Act's transparency rules took effect on August 2. They require AI companies to mark AI-generated content in a way that other systems can identify.

Who gets the watermark and when

ChatGPT and Codex users in the EU

OpenAI said the watermark will roll out over the coming weeks to eligible ChatGPT and Codex users on all plans in the EU. The company described the regional approach as a way to learn from real-world use and feedback, The Verge reported.

Developers and cloud partners

API customers anywhere in the world can opt in to watermarked output for select models starting Monday, according to both outlets. TechCrunch reported the feature is off by default. OpenAI also said it is working with cloud partners to offer watermarking for its model outputs accessed through their services in the coming weeks.

Limited access to the detector

Approved researchers and expert organizations can apply for access to a detection tool, which OpenAI said will be granted case by case. The tool reports whether it finds an OpenAI watermark without identifying the user or revealing prompts or conversations, The Verge reported. OpenAI said it is not making the detector public at launch because of the risk of missed watermarks and false positives.

How the watermark works

TechCrunch reported that the watermark is not a visible symbol. Instead, it subtly shapes the model's word choices, leaving a pattern readers cannot see but a detector can pick up. Because the pattern lives in the words themselves, it stays with the text when it is copied and pasted.

OpenAI published a technical report on the method, called textGrain, co-written with researchers from the University of Pennsylvania and Yale. It describes using a secret key to sort next-word predictions, and TechCrunch reported that hundreds of these small nudges let a detector spot AI-generated content using only the text and the key. OpenAI said the watermark does not identify the user and that it saw no meaningful change in model performance with it turned on.

The Verge reported that OpenAI says textGrain "matched or exceeded" other approaches, including Google DeepMind's SynthID for text.

The limits OpenAI acknowledged

OpenAI's own tests suggest editing can weaken the mark. In one test reported by TechCrunch, replacing 10% of words with synonyms dropped detection from about 92% to 66%. The company said short passages, math answers and translated text are also harder to detect.

According to The Verge, OpenAI said textGrain "does not guarantee reliable detection" and that watermarks do not verify accuracy, establish who owns a text, or measure how much a human contributed. OpenAI also told readers that a missing watermark "does not prove human authorship," TechCrunch reported, since text could be too short, heavily edited, or produced by another company's AI.

How rivals have handled it

The announcement follows Anthropic's August decision to watermark text from Claude, which TechCrunch reported the company is applying worldwide. The Verge noted Anthropic's approach is based on SynthID. That move drew backlash from some Claude users who argued they supplied the instructions and decisions while Claude was only the tool.

TechCrunch cited 2024 reporting by The Wall Street Journal that OpenAI had built a text watermark earlier but held off on releasing it, partly over concern that users would move to rivals that did not watermark. Anthropic, Google, Meta, Microsoft and OpenAI are among the companies that have committed to the EU's code of practice on AI-generated content, TechCrunch reported. More coverage of AI policy and products is collected in our AI and tech section.

What to watch next

• The rollout of watermarking to eligible ChatGPT and Codex users in the EU over the coming weeks.
• Whether cloud partners begin offering watermarking for OpenAI model outputs in the coming weeks, as the company said it is working toward.
• How OpenAI handles case-by-case detector access for approved researchers and expert organizations.
• Whether OpenAI changes its position that text watermarking will not be a global default at launch.