OpenAI Rolls Out Text Watermarking for ChatGPT and Codex

OpenAI has announced the rollout of an invisible, machine-readable text watermarking feature, dubbed 'textGrain,' for its popular AI models, ChatGPT and Codex. This new capability is initially being deployed for users within the European Union, marking a significant step towards enhancing transparency and accountability in AI-generated content. The primary driver behind this implementation is the impending EU AI Act, which mandates greater clarity regarding the origin of AI-generated materials. OpenAI's move mirrors similar initiatives from other leading AI developers; Google DeepMind has introduced its SynthID for text, a technology also adopted by Anthropic for its watermarking efforts announced in August. OpenAI stated that its textGrain approach has "matched or exceeded" the performance of these alternative methods. While the introduction of watermarking aims to address regulatory concerns and potentially combat misinformation, it has not been universally welcomed. The company also released scores from various AI benchmarks, demonstrating that the performance of watermarked text remains largely consistent with unwatermarked output. However, OpenAI prudently noted that 'textGrain' does not offer an absolute guarantee against the misuse or undetected alteration of AI-generated content, underscoring the ongoing challenges in this evolving field.