
OpenAI will start adding hidden watermarks to text and code made in the European Union. This move is meant to follow the EU’s new AI transparency rules.
The new OpenAI tool is called textGrain. It puts a secret statistical signal into the model’s word choices. A detector can then check for that signal. OpenAI says it tested better than other methods, such as Google’s SynthID for text. But the company warns it may not always work in real life.
Tests show some problems. Short text is hard to detect. Passages of about 200 tokens were caught around 80 percent of the time. However, longer passages of 400 tokens were caught about 95 percent of the time. Math content was much harder to detect. Editing also lowers the rate. Swapping 10 percent of words with synonyms cut detection from about 92 percent to 66 percent. Swapping 25 percent brought it down to 17 percent.
The textGrain OpenAI detector will not be open to everyone at first. Only approved researchers and expert groups will get access. OpenAI says the detector does not identify users or reveal prompts or conversations.
Article 50 of the EU’s AI Act says generative AI providers must make text outputs identifiable in a machine-readable way. Companies already in the market have until December 2, 2026 to comply.
Source: OpenAI
This article, OpenAI adds hidden watermarks to code, text generated on its platform, was originally published at NoypiGeeks | Philippines Technology News, Reviews and How to's.
Comments
Post a Comment