OpenAI has deployed an invisible text watermarking system across the European Union for ChatGPT and Codex coding output to meet transparency requirements under the EU AI Act. The technology, named textGrain, adjusts model word selection to create a hidden statistical pattern, though company tests show that simple editing drastically reduces detection rates.
How OpenAI textGrain System Performs Under Testing
Unlike traditional digital tagging methods, textGrain alters vocabulary selection directly within the output rather than inserting invisible Unicode characters, spaces, or custom punctuation. This design prevents users from stripping the watermark through simple text-cleansing scripts.
Detection accuracy depends heavily on response length and text modifications. OpenAI measures text in tokens, where one token equals roughly 0.75 English words. According to OpenAI test data:
- Unedited 300-word passages yielded a 92% detection rate.
- Unedited 150-word passages dropped to a 66% detection rate.
- Swapping 10% of words (1 in 10) with synonyms reduced detection on 300-word passages from 92% to 66%.
- Swapping 25% of words (1 in 4) with synonyms dropped detection accuracy to 17%.
Despite embedding statistical shifts into generated text, OpenAI noted that textGrain had almost no impact on benchmark evaluation scores for its frontier model, GPT-6 Astra.
Limited Access and Global Opt-In Settings
OpenAI is restricting access to its detection tool rather than making it public. The company is taking applications exclusively from approved researchers and specialized organizations. OpenAI explained that a positive test result does not identify the specific user, prompt, or account, while a negative result fails to prove human authorship because translated, short, or edited text easily bypasses detection.
By contrast, OpenAI relies on the SynthID framework for image and audio media, which allows the public to verify file origins through an online tool. While textGrain is mandatory within the EU to comply with enforcement rules that began in August, it remains disabled by default for users outside Europe. However, global API subscribers can manually enable the feature across organization and project settings.
The regulatory pressure comes as criminal adoption of AI tools grew 40% year-on-year, according to research from TRM Labs. As regulatory bodies press for stricter oversight, agencies like the IRS have already signaled that automated finance requires AI-driven regulatory supervision.
Why It Matters
The gap between regulatory requirements and technical capabilities presents a central challenge for AI compliance. While the EU AI Act mandates machine-readable identification for generated content, OpenAI's internal figures confirm that basic editing easily neutralizes text watermarks. Until detection tools achieve higher resilience against simple modifications, financial institutions and compliance teams cannot rely solely on watermarking systems to detect synthetic text or code.
