Implementation under European Union rules

OpenAI announced on 5 October that it will introduce invisible text watermarks to ChatGPT and Codex outputs across the European Union over the coming weeks. The policy responds to Article 50 of the European Union AI Act, which took effect on 2 August and requires generative artificial intelligence providers to mark synthetic content in a machine-readable format. Customers using the OpenAI application programming interface can enable watermarking immediately for selected models. While competitor Anthropic introduced text marking globally in August across its Claude models, OpenAI is limiting its automated deployment to European Union users while it evaluates performance before deciding on a global rollout.

The technology, named textGrain, operates during token generation rather than attaching visible tags, hidden unicode characters, or extra whitespace. Large language models calculate probability distributions for subsequent tokens (words, sub-words, or punctuation marks). TextGrain subtly adjusts these statistical probabilities according to a secret mathematical pattern to favor specific token choices. OpenAI explained the purpose of the tracking system in an official release.

“Content provenance helps people understand where content comes from, how it was created or modified, and whether it contains signals associated with our models.”

Detection mechanisms and technical constraints

A specialized detector analyzes text passages to locate the statistical regularities created by the model. Changing fonts, copying passages, or pasting plain text does not strip the watermark, because the pattern resides within the specific sequence of words. OpenAI stated that textGrain matches or exceeds alternative methods, including Google SynthID-Text. Furthermore, company benchmarks indicated that watermarked generation did not reduce response quality, addressing concerns raised by users when Anthropic introduced watermarks to Claude.

Detection tools face technical limitations. Shorter texts contain fewer tokens, making statistical identification difficult. Alterations such as paraphrasing, manual sentence editing, and translation across languages degrade or destroy the embedded pattern. The detector can generate false positives by incorrectly identifying human writing as synthetic, as well as false negatives by missing AI-generated passages. OpenAI restricted detector access to approved researchers and expert organizations during the testing phase. The watermark does not identify the specific user, does not record the degree of human editing, and does not certify factual accuracy or content ownership.

Regulatory deadlines and sector competition

The European Union AI Act established transparency standards for generative AI on 2 August, with a transitional compliance deadline of 2 December 2026 for models that entered the market prior to that date. Tech expert Bert Hubert noted that commercial competition accelerates adoption across the industry.

“Otherwise you almost have to explain as a company why you do not have this, which can be bad for your sales process.”

Hubert explained that machine-readable markers could help security systems intercept deceptive materials, such as fraudulent billing emails from impersonated utility providers. Anthropic previously applied open C2PA standards alongside statistical signatures, while Google published research on its SynthID-Text framework in Nature in 2024.