OpenAI has introduced a new feature to its text generation tools, ChatGPT and Codex, for users in the European Union. The invisible, machine-readable watermark, called textGrain, is designed to match the performance of other watermarking methods such as Google DeepMind’s SynthID. However, OpenAI is cautious, stating that textGrain does not guarantee reliable detection nor does it verify accuracy or ownership.

‘We are not making text watermarking a global default at launch,’ the company notes, indicating a regional approach to gather real-world feedback. This move is in line with the EU’s AI Act, but it has raised eyebrows in the tech community, who wonder about the implications of such surveillance on creativity and privacy.

For API customers, OpenAI is offering an opt-in option, allowing them to decide how watermarking fits their needs. The company is partnering with cloud services to make this feature available in the coming weeks. Meanwhile, access to the watermark detection tool is being provided on a case-by-case basis to approved researchers, with no public release at this stage.

This step towards increased transparency and accountability in AI-generated content is a significant development. But it also raises questions about the extent of AI oversight and the potential for misuse of such technology.