OpenAI introduces hidden text markers for ChatGPT and Codex outputs
OpenAI is rolling out an imperceptible watermarking system for text produced by ChatGPT and its coding assistant Codex. The method, called textGrain, embeds a statistical pattern in model output so specialized tools can identify AI-written content. It is mandatory for users in the European Union under the AI Act, while API customers elsewhere can opt in.
OpenAI's textGrain works by changing how a language model samples its next token. A secret key produces pseudorandom values that gently shift candidate probabilities, leaving a statistical signature while preserving meaning, tone, and readability. Detection systems holding the key can check whether that signature appears across a passage.
The rollout is phased by region and service. EU users must have it under the AI Act, while API customers elsewhere may choose it. Over coming weeks, it will reach ChatGPT web and mobile, plus Codex.
The watermarking could affect writers, students, developers, and platforms by making AI-assisted text easier to label or verify. EU users may see it as a compliance measure, while API customers elsewhere can decide whether to adopt it. Detection could help with transparency, but it may also raise concerns about false positives or overreliance on imperfect signals. Its social effect may depend on accuracy, access to detection tools, and how institutions interpret watermark results.