OpenAI to Default to Watermarking ChatGPT Output in EU

OpenAI will implement automatic watermarking for text produced by its ChatGPT service within the European Union, the company announced. This decision comes as a direct response to the EU AI Act, which mandates that content generated by artificial intelligence must be identifiable. While the watermarking feature will be available globally, it will not be enabled by default outside of the EU.

The EU AI Act, effective from August, requires AI-generated content to be marked in a way that allows for its detection. OpenAI's proprietary method, called textGrain, embeds imperceptible patterns into the text that can be identified by a specialized detector. This approach aims to provide a traceable element to AI-generated content without significantly altering the output quality for human readers.

OpenAI plans to grant access to its textGrain detection tool to a select group of researchers and organizations. The company also intends to establish a process for other entities to request access over time. However, the effectiveness and security of such watermarking technologies remain a subject of discussion, as existing methods have been shown to be circumventable with sufficient technical knowledge.

The rollout of default watermarking in the EU highlights the growing regulatory landscape for artificial intelligence and the challenges companies face in balancing compliance with technological feasibility. The long-term impact of this measure on AI content creation and detection is yet to be determined.