OpenAI adds invisible text watermarks to ChatGPT in the EU

In short: OpenAI is starting to add invisible “watermarks” to text made by ChatGPT and Codex, beginning with users in the European Union.

What happened

OpenAI says it will roll out a new system called textGrain that adds an invisible, machine-readable watermark to some text outputs from ChatGPT and Codex. “Machine-readable” means people usually cannot see it, but a computer tool can try to spot it (like a hidden stamp on a document).

The EU rollout will happen over the coming weeks, and it will apply to eligible ChatGPT and Codex users across all plans in the European Union. OpenAI says it is not making watermarking the global default at launch, and it wants to learn from real-world feedback first.

Outside the EU, OpenAI says API customers can opt in to watermarked text starting today for select models. An API is a way for other apps and services to connect to OpenAI, kind of like ordering from a menu instead of using the ChatGPT website.

OpenAI also says approved researchers and expert organizations can apply for access to a detection tool. The tool will only say whether it detects an OpenAI watermark, and it will not reveal the user, prompts, or conversations.

Why it matters

Watermarks are meant to help people tell when text probably came from an AI system. But OpenAI warns the watermark will not always be detected, and it does not prove the text is true, who owns it, or whether a human wrote it. The move also lines up with the EU AI Act, which pushes for clearer labeling of AI-generated content.

Source: The Verge AI