Text Watermarks in the EU: OpenAI’s Response to Regulation

In the European Union, claiming ChatGPT’s generated content as your own is about to become more challenging. OpenAI announced on Monday that it will begin embedding an invisible watermark—undetectable by the human eye but identifiable by machines—into all text produced by its AI models, including ChatGPT and Codex, its coding assistant.

This watermark is a significant step for OpenAI to comply with the EU AI Act, which mandates that AI-generated text must include a machine-readable identifier. Initially, this watermark will be an opt-in feature available for global users starting today. in the upcoming weeks, the watermark will be automatically applied to all eligible ChatGPT and Codex text outputs within the European Union, with no option to disable it.

Who will have access to this watermark? OpenAI states that researchers and expert organizations will receive a tool to detect the unseen marker. The company has decided not to release this detection tool to the public at the outset due to concerns about false positives. Instead, access will be granted on a case-by-case basis to facilitate evaluation.

OpenAI joins other companies in implementing watermarking technology. Anthropic introduced a similar machine-readable marker in August, which prompted users to seek methods to remove AI indicators from their text. Google DeepMind has been embedding a watermark in its outputs for several years. Interestingly, Google recently reversed its stance, announcing in August that users could eliminate watermarks from images, videos, and music in regions where they are not mandated.

It’s important to note that the watermarks from these other companies primarily target users in the United States. In contrast, OpenAI’s watermark does not apply there, despite increasing worries about the rise of AI-generated content. A recent survey from the Rainey Center indicated that nearly 80% of respondents support the implementation of AI watermarking requirements, including 82% of Democrats and 75% of Republicans.

Public support does not equate to policy change, and the United States continues to fall behind in establishing federal regulations concerning AI. Without a requirement for leading AI companies to incorporate a watermark that aids users and platforms in identifying AI-generated content, OpenAI lacks the incentive to comply with a non-existent law. This illustrates that regulations can be effective if they are enacted and enforced.