OpenAI announced the rollout of textGrain, an invisible machine-readable watermarking system for text generated by its ChatGPT and Codex models. The company stated that its approach matched or exceeded comparable watermarking technologies, including Google DeepMind’s SynthID and Anthropic’s recently announced system, both similarly designed to address requirements under the European Union’s AI Act.
The watermarking deployment will follow a phased approach. Starting with the European Union, eligible ChatGPT and Codex users across all subscription plans will receive watermarked text outputs over the coming weeks. OpenAI indicated it would not establish watermarking as a global default immediately, citing a preference to gather real-world usage data and feedback before broader expansion.
Simultaneously, OpenAI introduced an opt-in watermarking option for API customers worldwide, allowing them to determine how the feature aligns with their transparency requirements and user experiences. The company indicated it is coordinating with cloud partners to extend watermarking availability for OpenAI model outputs accessed through their platforms in the subsequent weeks.
OpenAI acknowledged certain limitations of the watermarking technology, noting that textGrain does not guarantee reliable detection and cannot verify content accuracy, establish ownership, measure human contribution levels, or confirm human authorship. The company also provided AI benchmark scores demonstrating comparable performance between watermarked and non-watermarked text outputs.
To support ongoing evaluation and refinement of text provenance technologies, approved researchers and expert organizations began applying for detector access. OpenAI stated it would grant access on a case-by-case basis in accordance with relevant standards, with the detector reporting watermark presence without disclosing user identity or conversation details. The company refrained from making the detector publicly available at launch, citing risks associated with missed watermarks and false positives.
Article Attribution | Read More at Article Source
Article summary produced by Claude AI