OpenAI rolls out weak watermarking for generated texts

OpenAI has started offering its API customers the option to apply watermarks to the outputs of certain models as part of its efforts to comply with the AI Act in Europe. In the coming weeks, the company plans to implement a form of digital watermarking on text generated by ChatGPT and Codex within the European Union. Unlike its competitor Anthropic, OpenAI will not apply this technique to content generated outside of Europe, although API users can choose to include these marks on texts from any location.

## Why it matters The introduction of watermarks on AI-generated texts is a significant step towards regulating the use of artificial intelligence in Europe. This move aims to enhance transparency but also raises questions about the effectiveness of these marks in identifying AI-generated content.

## What we know OpenAI's watermarking is designed to meet the requirements of the AI Act, which mandates that generative AI service providers identify their outputs in a machine-readable manner. However, the technology employed by OpenAI has limitations, as it can miss a considerable amount of AI-generated text and is easily tricked.

## What remains unclear It has not yet been confirmed how effective this solution will be in practice, especially since OpenAI will only implement the watermarking system in Europe and leave the decision to apply it globally to API customers. This raises concerns about the consistency and reliability of identifying AI-generated texts in varying contexts.