OpenAI follows Anthropic and complies with the European AI Act. Sam Altman's company will add a machine-readable digital watermark to the text generated by ChatGPT and Codex, to comply with the transparency requirement set out in Article 50 of the EU Regulation on Artificial Intelligence, effective August 2.
"Content provenance helps people understand where content comes from, how it was created or modified, and whether it contains signals associated with our models," the California-based company states in an official post. "Watermarking and text detection technologies are still in their early stages and have significant limitations. Furthermore, the debate about their benefits and possible responsible uses is still evolving. Our phased approach takes into account both the requirements of the AI Act and technological limitations, emphasizing transparency regarding what a text watermark can and cannot communicate to users."
OpenAI will leverage its textGrain technology, which it claims outperforms competing technologies, including Google's SynthID. However, it emphasizes, "detection systems can make two types of errors: reporting the presence of a watermark when none is present (false positive) or failing to detect a watermark that is actually present (false negative)."
After the introduction of the digital watermark for texts in the EU, based on the feedback received, the company will decide whether to extend it globally. Researchers and expert organizations can test the tool. Anthropic embarked on this path in August, declaring it would use the open technology standard C2PA and extend watermark support to older AI models as well.
(Unioneonline)
