OpenAI Introduces Invisible Text Watermarking in EU to Comply with AI Act

OpenAI is introducing invisible watermarking for text generated by ChatGPT and Codex in the European Union, complying with the EU Artificial Intelligence Act. The new technology, called textGrain, embeds statistical signals into the word choices made by the model.

Invisible Signals and Word Choice

According to OpenAI, textGrain uses a cryptographic-like key that leads the model to prefer certain words over others during generation. These subtle shifts create hidden statistical patterns that remain invisible to the human eye while ensuring the text appears completely natural. However, the system's detection capabilities vary depending on the length and nature of the content. In psychological texts, the tool successfully identified watermarks in roughly 80 percent of paragraphs containing 200 tokens, jumping to 95 percent for paragraphs with 400 tokens. Conversely, mathematical content offers models less lexical flexibility, resulting in lower detection rates.

"The system relies on statistical signatures embedded directly into the text generation process, though editing and paraphrasing can significantly diminish detection reliability," OpenAI noted.

Limitations and Regulatory Scope

OpenAI emphasizes that text watermarking is not foolproof. When users substitute 10 percent of the model-selected words with synonyms, the detection rate drops from 92 percent to 66 percent. Replacing 25 percent of the words causes the detection accuracy to plummet to 17 percent. For now, the rollout is limited to the European Union to satisfy regulatory compliance, though the company plans to use the deployment to refine its detection mechanisms.