TL;DR — Key Takeaways
– OpenAI is rolling out textGrain, an invisible, machine-readable watermark for eligible ChatGPT and Codex text in the EU.
– The watermark alters token selection in subtle ways that create a statistical signature detectable with specialized tools.
– The move is aimed at complying with Article 50 of the EU AI Act, which requires certain AI-generated content to be machine-detectable.
OpenAI says it has begun embedding an invisible, machine-readable watermark into text generated by ChatGPT and its coding assistant, Codex, to comply with strict European Union regulations.
The system, called textGrain, alters how underlying large language models (LLMs) select words by factoring in pseudorandom values tied to a secret key. While invisible to human readers, the process leaves a distinct statistical signature that allows specialized tools to confirm whether a passage was produced by artificial intelligence (AI).
The rollout comes as technology giants race to satisfy Article 50(2) of the landmark EU Artificial Intelligence Act.
The regulation mandates that AI-generated outputs be clearly marked and detectable in a machine-readable format. Companies that launched models prior to Aug. 2 have until Dec. 2 to comply or face steep penalties, up to 3% of global annual turnover. With OpenAI’s annualized recurring revenue approaching $70 billion, potential non-compliance fines could exceed $2 billion.
While the watermark will soon be mandatory for all ChatGPT and Codex users across all service tiers in the EU, OpenAI is taking a more cautious approach elsewhere. Outside Europe, the feature remains off by default, available only as an opt-in setting for global API customers and cloud partners.
OpenAI noted that its regional deployment allows room to gather real-world feedback before considering wider rollouts.
Unlike competitors Anthropic and Google DeepMind, which implemented similar watermarking technologies globally, OpenAI is holding back a general release in unregulated markets like the United States, where federal requirements for AI tracking currently do not exist.
The company also cautioned that text watermarking remains an emerging technology with significant technical constraints. In testing, OpenAI’s detector successfully identified watermarks in only 80% of 200-word passages. Reliability dropped further on topics with rigid technical lexicons, such as mathematics, and on shorter text samples.
Human editing also severely degrades detection accuracy. According to OpenAI, altering just 10% of words in a 400-word passage reduced the detection rate to 66%, while modifying 25% of the text dropped it to 17%.
Because of these limitations and the risk of false positives, OpenAI is withholding its detection tool from the public. Access is currently restricted to approved researchers and expert organizations on a case-by-case basis.
OpenAI emphasized that textGrain is designed strictly as a machine identifier and cannot perform broader authentication tasks. The watermark cannot establish copyright ownership, verify facts, calculate human contribution, or identify the specific user who prompted the AI. Furthermore, the absence of a watermark does not serve as definitive proof of human authorship.
Despite these hurdles, regulatory pressure is rapidly reshaping how tech firms handle synthetic media. With Europe enforcing its rules under threat of massive financial penalties, tech developers are finding that compliance is no longer optional.
