OpenAI Adds Text Watermarking to ChatGPT and Codex in EU

OpenAI has begun deploying an invisible, machine-readable text watermark across its ChatGPT and Codex platforms, marking a strategic move to align with emerging regulatory standards. The feature, internally designated as textGrain, is initially rolling out to eligible users across all subscription tiers in the European Union. This phased, regional launch allows the company to gather real-world usage data and feedback before considering a global default. The implementation directly addresses compliance requirements under the European Union’s AI Act. OpenAI’s textGrain system reportedly matches or exceeds the performance of comparable solutions developed by Google DeepMind and Anthropic, which utilize similar underlying frameworks. While the move signals industry-wide adoption of AI transparency measures, it has sparked mixed reactions among users and developers concerned about potential privacy and workflow implications. According to OpenAI, benchmark tests demonstrate that textGrain maintains performance parity with unwatermarked outputs, showing no measurable degradation in model accuracy or fluency. However, the company explicitly cautions that the technology is not a definitive verification tool. Watermarked outputs cannot guarantee reliable detection, nor do they authenticate factual accuracy, establish content ownership, quantify human involvement, or confirm human authorship. The watermark serves strictly as a machine-readable flag rather than a provenance certificate. The rollout strategy extends beyond consumer applications. API customers worldwide can immediately opt into watermarked outputs for select models, enabling enterprises to integrate the feature according to their specific transparency obligations. OpenAI is also collaborating with major cloud providers to make watermarking accessible through their respective AI service layers within the coming weeks. In parallel, OpenAI is opening a controlled access channel for its detection tool, initially limiting approval to vetted researchers and expert organizations. Access will be granted on a case-by-case basis to support the evaluation and refinement of text provenance technologies. The detector will only indicate the presence or absence of an OpenAI watermark and will not expose user identities, prompts, or conversation history. Public availability has been deliberately delayed to mitigate risks associated with false positives and missed detections during the technology’s early deployment phase. This structured approach reflects OpenAI’s balancing act between regulatory compliance, technical transparency, and ecosystem scalability. By prioritizing regional testing and opt-in enterprise controls, the company aims to refine its watermarking infrastructure while establishing a standardized framework for AI-generated text identification across competing platforms.