OpenAI will begin watermarking ChatGPT’s textual content within the EU

OpenAI will begin including an invisible watermark to textual content generated by ChatGPT and Codex within the European Union to adjust to the EU AI Act, the corporate stated Monday in a blog post.

The EU AI Act’s transparency rules, which took impact on August 2, require AI firms to mark AI-generated content material in a method different programs can establish.

OpenAI stated the watermark will roll out over the approaching weeks to eligible ChatGPT and Codex customers on all plans, however solely within the EU. Developers utilizing OpenAI’s API wherever within the globe can flip it on for choose fashions beginning at this time; it’s off by default. OpenAI stated it’s not making textual content watermarking a world default at launch.

The watermark isn’t an precise image, however works by subtly shaping the mannequin’s phrase decisions, leaving a sample readers can’t see, however a detector can decide up. Because it lives within the phrases themselves, it travels with the textual content when it’s copied and pasted. OpenAI stated the watermark doesn’t establish the consumer, and that it noticed no significant change in its fashions’ efficiency with it switched on.

OpenAI additionally printed a technical report for its methodology, known as textGrain, alongside the announcement. Co-written with researchers from the University of Pennsylvania and Yale, it walks by an instance of utilizing a secret key to kind next-word predictions to complete the sentence. Add lots of of those nudges collectively, and the detector can spot AI-generated content material utilizing solely the textual content and the important thing.

Can the watermark be eliminated by modifying? OpenAI’s assessments counsel sure. In one take a look at, changing 10% of phrases with synonyms dropped detection from about 92% to 66%. The firm additionally stated brief passages, math solutions, and translated textual content are more durable to detect.

“These limitations contribute to our choice to offer preliminary detector entry solely to permitted researchers and knowledgeable organizations, who may help us consider reliability and accountable makes use of,” stated the corporate.

OpenAI additionally cautioned {that a} lacking watermark “doesn’t show human authorship.” The textual content could possibly be too brief or too closely edited, or it may come from one other firm’s AI.

“[Watermarks] can point out that an OpenAI system generated or processed a part of a passage, however not how a lot human judgment, modifying, or creativity went into it,” the corporate stated.

The announcement comes two months after Anthropic said it would watermark textual content generated by Claude, a transfer it’s making use of worldwide. That choice drew backlash from some Claude users, who argued that they had equipped “the directions, context, selections” whereas Claude was simply “the device.”

OpenAI had constructed a textual content watermark earlier than however held off on releasing it, partly over issues that customers would swap to rivals that didn’t watermark, The Wall Street Journal reported in 2024.

Anthropic, Google, Meta, Microsoft, and OpenAI are among the many firms which have dedicated to following the EU’s code of practice on AI-generated content material.