OpenAI announced plans to add invisible digital watermarks to text and code generated in the European Union to comply with the region’s new AI transparency rules.. Article 50 of the EU AI Act requires generative AI providers to make text outputs identifiable in a machine-readable way.
The requirement already applies to new market entrants; existing companies, including OpenAI, Anthropic, Microsoft, Google and Meta, have until December 2 to comply.. Watermarking will be enabled by default only for eligible ChatGPT and Codex text output in the EU. Customers elsewhere can opt in for select models, which OpenAI did not identify..
OpenAI said its proprietary textGrain system adds an invisible statistical signal to a model’s word choices that a detector can identify. It plans to release the technology as open source so others can modify it.. The company said textGrain “matched or exceeded” alternatives, including SynthID for text, but cautioned that “strong performance under ideal conditions does not guarantee reliable detection in everyday use.”.
OpenAI acknowledged that shorter text has a detection success rate of around 80 percent. Mathematics is also harder to detect, and editing watermarked text can reduce detection rates.. Anthropic made a similar watermarking move earlier this year..
Access to OpenAI’s detector will initially be restricted to approved researchers and expert organizations. The detector will not identify users or reveal prompts or conversations.. Featured image credit.
Tags: chatgptCodexFeaturedopenAI
