OpenAI has introduced new text watermarking technology to comply with the European Union AI Act. This regulation mandates that providers of generative artificial intelligence ensure that AI-generated content is identifiable through machine-readable methods.
Starting immediately, OpenAI is allowing global API customers to opt into text watermarking for specific models, though the feature remains disabled by default. In the coming weeks, the company will also implement invisible watermarks for eligible ChatGPT and Codex outputs within the EU.
The company is also accepting applications for its text watermark detector, restricting access to select researchers and expert organizations. OpenAI noted that these tools are in early development and have significant limitations regarding accuracy and the ability to distinguish between human and AI authorship.
These new provenance tools currently apply only to text and do not impact existing verification systems for audio and images. The move follows similar compliance measures from Anthropic, which began labeling content from its Claude models as AI-generated in August to meet EU standards.
