OpenAI has begun rolling out an invisible, machine-readable text watermarking system named textGrain for ChatGPT and Codex users located within the European Union. The localized deployment aims to align with transparency requirements established by the EU AI Act while keeping global default behavior unchanged for now.
Regional Rollout and Developer Opt-Ins
The textGrain technology is expanding over several weeks to all ChatGPT and Codex service tiers exclusively within the EU. OpenAI stated that keeping the feature regionally restricted gives the company room to collect operational feedback before considering broader deployment. Outside the EU, API developers globally can voluntarily opt in to watermarked outputs for select models, with support across cloud partner platforms expected in the following weeks.
Detection Tooling and System Boundaries
To evaluate text provenance under the regulatory Code of Practice, OpenAI is offering select researchers and accredited organizations case-by-case access to a detection tool. The system checks if text originated from OpenAI models without sharing individual user identities, prompts, or conversation history. OpenAI is holding back public release of the detector due to risks of missed signals and false positives. The company also emphasized that textGrain does not guarantee foolproof detection, nor does it verify factual accuracy, determine copyright ownership, or quantify human creative input.
Performance Claims and Regulatory Context
According to OpenAI, benchmark assessments showed that textGrain maintains model performance comparable to unwatermarked text outputs. The company claimed the implementation matched or exceeded alternative mechanisms such as Google DeepMind’s SynthID, which Anthropic announced adopting in August to fulfill EU AI Act obligations. The rollout underlines how major AI laboratories are implementing localized compliance mechanisms to satisfy European regulatory deadlines.
AI-generated editorial illustration; not a documentary photograph.
Source: OpenAI is adding text watermarking in ChatGPT and Codex
SIGNAL TITANS
Leave a Reply