OpenAI is introducing a text-based watermarking system called textGrain for its ChatGPT and Codex services. The rollout begins in the European Union in the coming weeks and will apply to all subscription tiers. The company intends to gather user feedback during this regional launch before proceeding with a global release.
The textGrain technology is designed to match or exceed the performance of existing solutions like Google DeepMind's SynthID. Anthropic has also adopted similar labeling methods to comply with the EU AI Act. OpenAI notes that the watermarks do not guarantee factual accuracy, identify the user, or serve as proof of human authorship.
Starting today, global API users can opt to enable watermarking for specific models. OpenAI is also working with cloud partners to integrate this feature into their services. Additionally, researchers and relevant organizations can now apply for individual access to evaluate and improve text provenance tracking methods.
