OpenAI to watermark EU-generated text

The company announced on October 5 that the rollout would begin over the coming weeks across all plans in the EU.

“We are not making text watermarking a global default at launch,” OpenAI said in its official announcement.

Developers worldwide can opt in through OpenAI’s API for selected models, with watermarking switched off by default.

The EU’s transparency obligations, applicable since August 2, require providers to mark AI-generated content in a machine-readable form that enables detection.

OpenAI’s technology, called textGrain, subtly adjusts word choices to embed a statistical pattern invisible to readers. Because the signal is contained in the wording, it can remain when text is copied and pasted.

The company reported no meaningful performance differences in its model benchmarks with watermarking enabled.

However, editing can weaken detection. In one evaluation, replacing 10% of words with synonyms reduced detection from about 92% to 66%. Short passages, mathematical answers and translated text also present difficulties.

Detector access will initially be restricted to approved researchers and expert organisations because of the risk of missed watermarks and false positives.

OpenAI stressed that watermarks cannot identify users, verify factual accuracy or measure human involvement in writing. A detected signal may indicate that its systems generated or processed part of a passage, without revealing how much a person contributed.

Likewise, the absence of a watermark “does not prove human authorship,” the company cautioned.

OpenAI also published a technical report and said it plans to make the technology open source.

Earlier, Qazinform News Agency reported that OpenAI unveiled Dots, a new line of always-on AI agents designed to work independently on users’ tasks and goals, following Meta’s introduction of its Muse AI agent.