OpenAI Adds Text Watermarking to ChatGPT and Codex in EU
OpenAI is implementing an invisible, machine-readable watermark in text generated by ChatGPT and Codex. The feature is initially rolling out to users exclusively within the European Union.
AI company OpenAI has begun deploying an invisible, machine-readable watermark within text output from its ChatGPT and Codex services. Initially, this feature will be available only to users located within the European Union.
OpenAI states that its textGrain watermarking technology meets or exceeds the performance of other approaches, such as Google DeepMind's SynthID for text. Anthropic announced a similar system in August, also based on SynthID.
These moves by both OpenAI and Anthropic are in response to the requirements set forth by the EU's AI Act, which aims to enhance transparency in AI-generated content. However, the introduction of this watermarking has faced some criticism.
OpenAI has shared results from AI benchmark tests indicating that watermarked and unwatermarked text perform comparably. The company notes, however, that textGrain does not guarantee identification in all circumstances.