🤖 TechnologyHotCovered by 3 outlets

OpenAI begins text watermarking for ChatGPT and Codex users in the EU

Oct 7, 1:16 AM2 min read

Drafted by AI from the sources below, fact-checked by the Ajako Taja Editorial Team · How we use AI

AI Summary

OpenAI is launching invisible text watermarking for EU users to meet regulatory requirements, though tests show the system can be bypassed by simple editing.

OpenAI has announced it will automatically watermark text generated by ChatGPT and Codex in the European Union to comply with the EU AI Act.

The new feature, which utilizes a proprietary method called textGrain, will roll out to all ChatGPT and Codex plans in the EU over the coming weeks. According to reports from TechCrunch and The Verge, the watermark is invisible to human readers and works by subtly influencing the model’s word choices. This creates a machine-readable pattern that remains even if the text is copied and pasted elsewhere.

While the watermarking is the default in the EU, OpenAI stated it will not be a global default at launch. Outside of the EU, API customers can choose to opt in to the feature for select models starting today. Ars Technica noted that while competitor Anthropic enabled similar watermarking globally in August, OpenAI is currently limiting the default setting to regions where regulators require it.

Technical data published by OpenAI shows that the system is not foolproof. In tests, the detector had a 92 percent success rate, but that figure dropped significantly with minor edits. TechCrunch reported that replacing just 10 percent of the text with synonyms reduced detection to 66 percent, while Ars Technica noted that changing 20 percent of the text could lower the success rate by nearly 75 percent. The company also cautioned that shorter passages, translations, and math answers are harder to detect.

OpenAI is not making the detection tool public yet, citing risks of false positives and missed watermarks. Instead, access will be granted to approved researchers and expert organizations on a case-by-case basis. The company emphasized that the watermark does not identify specific users or verify the accuracy of the content, but merely indicates that an OpenAI system generated or processed part of the passage.

Background

The move is a response to the EU AI Act, which took effect in August and requires AI companies to mark generated content so that other systems can identify it. OpenAI published a technical report on the textGrain method co-written with researchers from the University of Pennsylvania and Yale.