OpenAI Rolls Out Invisible Text Watermarks for ChatGPT and Codex, Starting in the EU

OpenAI has begun adding an invisible, machine-readable watermark to text produced by ChatGPT and Codex, with the first rollout limited to users in the European Union. The mark is designed to let automated systems detect whether a passage was generated by the company's models, without changing what users see on screen. The EU-first timing lines up with the bloc's AI Act, which imposes transparency and labeling duties on providers of general-purpose AI systems.

The technique, which OpenAI calls textGrain, was benchmarked against rival approaches and performed at least as well, the company says. That comparison includes Google DeepMind's SynthID for text, the same underlying method Anthropic drew on for the watermarking it announced in August. Two of the three largest US model developers now point to the same family of detection techniques, giving the work a shared technical baseline rather than a set of incompatible proprietary schemes.

Text watermarking is harder than watermarking images or audio. A picture carries millions of pixel values that can absorb tiny statistical nudges; a paragraph is a short string of tokens where almost any alteration is visible to a reader. TextGrain works by biasing the model's choice among near-equivalent next words according to a key, so the resulting prose reads normally but carries a statistical signature that a detector holding the key can recover.

That design also exposes the limits. Watermarks survive copying and pasting but can be weakened by paraphrasing, translation, or running the text through another model, and they cannot travel with facts that a person retypes by hand. Detection is only available to parties with the key, which means OpenAI, not the public, controls who can check a given document. The scheme also covers only OpenAI's own outputs, so text from other vendors, open-weight models, or a locally run fine-tune carries no such signal.

For developers, the practical questions are where the mark is applied and how it interacts with downstream tools. Codex generates code, where a watermark could complicate license review, diffing, or ingestion into training pipelines if detectors start flagging machine-written source files. Teams building detection or provenance features should test against watermarked ChatGPT output now, before the scheme reaches users outside the EU, because a detector tuned to unmarked text will behave differently once the signal is present.

OpenAI has not said when the watermark will expand beyond the European Union or whether it will apply to API traffic, which matters most to companies piping model output into products. Anthropic's August rollout and Google's SynthID work give the field three overlapping implementations, but no shared standard for reading them. The next thing to watch is whether the EU's transparency rules push other providers toward the same detection key formats, or whether users get a patchwork of marks that only their creators can read.