OpenAI Follows Anthropic With Watermarked Text in the EU
OpenAI will use a new tool called textGrain to watermark ChatGPT and Codex text in compliance with EU rules, following Anthropic's move to add an invisible watermark to text generated by Claude. But the companies' methods differ vastly.
OpenAI says its watermark hides a special code within generated text, allowing an AI detector to tell whether an OpenAI system produced or edited it. textGrain prepares several slightly tweaked versions of the model’s word probabilities, keeps their average close to the original so quality stays the same, and then uses the key to pick which version to use at each step. Meanwhile, Claude's method works like a keyed lottery for each word: It gives every candidate word a secret score based on the key and recent text, then picks the word with the highest score so the pattern shows up over long passages.
AI detectors' ability to identify textGrain-watermarked text...
Copyright of this story solely belongs to www.extremetech.com. To see the full text click HERE
