OpenAI Adds Invisible Watermarks to ChatGPT in EU, Leaves API Users a Way Out

OpenAI is rolling out invisible watermarks on text generated by ChatGPT for users in the European Union, marking its first regional deployment of the provenance technology. The marks, built on a system called textGrain, are designed to let detectors identify machine-written passages without altering how the text reads. API customers worldwide will be able to turn the watermarking off, a decision that separates OpenAI from rival Anthropic.

The watermarking push responds to EU transparency rules that increasingly expect AI providers to label synthetic content. textGrain works by subtly biasing word choices during generation so a statistical detector can later flag the pattern, leaving no visible trace for a human reader. OpenAI has not said whether the EU rollout will expand to other regions or become the default everywhere.

Detection performance is uneven, and the company's own tests show why. textGrain identified watermarked text at rates as high as 95 percent under clean conditions, but the figure fell to 17 percent once roughly a quarter of the words were swapped out. That gap matters because paraphrasing tools, translation and light editing are routine, and a watermark that survives only untouched output offers limited protection.

The opt-out for API users reflects a commercial calculation. Developers building products on OpenAI's models often need raw, unmarked output for fine-tuning, evaluation or downstream processing, and a mandatory watermark could interfere with those pipelines or raise liability questions for customers. Anthropic, by contrast, has taken a firmer line, and the divergence gives enterprises a concrete factor to weigh when choosing a model provider.

Watermarking has become one of the more contested technical problems in generative AI. Proposals range from token-level statistical signatures to metadata attached at the API layer, and none has proven both durable and invisible under real editing. A detector that misses most paraphrased text could create false confidence, while an aggressive one risks flagging human writing as synthetic.

For now, the practical effect is limited to EU ChatGPT users, with the rest of the world's API traffic unaffected unless customers opt in. The open question is whether regulators will accept a system that can be disabled by the very developers whose output is most likely to circulate at scale. OpenAI has not published a timeline for extending textGrain beyond the EU or for tightening the API policy.