ChatGPT and Codex to Feature New Text Watermarking Technology

A new, invisible watermark for text output is being introduced for ChatGPT and Codex, initially available only to users in the European Union. OpenAI claims that its textGrain watermarking technology “matched or exceeded” other methods, such as Google DeepMind’s SynthID, which also serves as the foundation for the watermarking announced by another company in August. This move aligns with the EU’s AI Act requirements, though not everyone is thrilled about the change.

OpenAI has shared performance scores from AI benchmarks indicating that watermarked and unwatermarked text perform similarly. However, it emphasizes that textGrain “does not guarantee reliable detection” and that text watermarks do not confirm accuracy, ownership, the extent of human contribution, or human authorship.

> The rollout of text watermarking in the EU will begin in the coming weeks, targeting eligible ChatGPT and Codex users across all plans. This feature will not be a global default at launch, allowing for real-world feedback and learning.

> API customers worldwide can opt in to receive watermarked text outputs for select models starting today. This option enables customers to align watermarking with their transparency needs and user experiences. OpenAI is also collaborating with cloud partners to offer watermarking for model outputs accessed through their services soon.

> Access to detection tools will be available for approved researchers and expert organizations starting today. In line with the Code of Practice, access will be granted on a case-by-case basis to aid in evaluating and improving text provenance. The tool will indicate whether an OpenAI watermark is detected without revealing user identities or their prompts. Due to the potential for missed watermarks and false positives, public access will not be available at launch.