OPENAI TO ADD WATERMARK TO CHATGPT TEXT IN EU

OpenAI said it will begin adding an invisible watermark to text generated by ChatGPT and Codex for users in the European Union, according to a blog post from the company. OpenAI said the move is intended to comply with the EU's AI Act, whose transparency rules took effect on 2 August and require providers of generative AI systems to make outputs identifiable in a machine-readable way. The company said it is using a proprietary system called textGrain, which it described as adding an invisible statistical signal to the model's word choices that a separate detector can then identify. OpenAI said the rollout will happen over the coming weeks for eligible ChatGPT and Codex users across all plans in the EU, and that watermarking will not be a "global default" at launch. Anthropic announced a similar watermarking move in August to meet the same EU requirement, OpenAI said.

OpenAI said textGrain "matched or exceeded" the performance of other approaches, including Google DeepMind's SynthID for text, citing its own benchmark scores. The company cautioned, however, that "strong performance under ideal conditions does not guarantee reliable detection in everyday use," and said shorter passages and mathematical content are harder to detect, with roughly an 80% success rate in some cases; editing watermarked text also reduces detection accuracy, OpenAI said. The company added that a watermark does not verify accuracy, establish ownership, measure human contribution, or prove human authorship. OpenAI said it plans to release the watermarking technology as open source so others can adapt it. Pre-existing providers, including OpenAI, Anthropic, Microsoft, Google and Meta, have until 2 December to comply with the relevant AI Act provision, according to Report B.

From today, OpenAI said, API customers worldwide can opt in to watermarked outputs for select models, and it is working with cloud partners to extend watermarking to its models accessed through their services in the coming weeks. Access to the detection tool will initially be limited to approved researchers and expert organisations, which can apply from today under what OpenAI described as its Code of Practice commitments, with approvals granted on a case-by-case basis. OpenAI said the detector does not identify users or reveal their prompts or conversations.