ChatGPT Output Will Carry Hidden Watermarks: How textGrain Works and Who It Affects

OpenAI has begun rolling out hidden marking of text generated by ChatGPT and its Codex development environment in the European Union. The company’s textGrain technology embeds a special statistical trace in the model’s responses, letting algorithms detect the presence of machine-generated content. The move is a direct result of the European Union’s AI Act taking effect.

Unlike earlier, crude watermarking methods, textGrain does not rely on hidden characters, invisible spaces or technical metadata that can easily be wiped with a plain text editor. As OpenAI explains in its blog, the system intervenes directly in the generation process, subtly shifting the probabilities of choosing particular words and tokens.

The result is a hidden mathematical pattern that forms across a sufficiently long passage of text. To the human eye, the phrasing looks entirely natural and well-formed, but a specialized detector can confirm with high accuracy that the wording was generated by OpenAI models.

Vulnerability to Paraphrasing and a Closed Detector

For all the sophistication of the approach, its creators openly warn about its limits. In the company’s help center, they stress that the watermark is not strict legal proof of AI involvement — it only indicates a high probability that AI was used.

The mark’s resilience to editing has its limits:

• heavy manual rewriting or automatic translation into another language can blur or completely destroy the statistical trace;
• short replies, program code and scientific texts full of formulas are nearly impossible to mark reliably without degrading the quality of the response itself;
• the textGrain detector will not become a public service: initially, access will go only to accredited academic researchers and expert organizations, to reduce false accusations and keep spammers from quickly testing ways to evade it.

A Law for Europe, Freedom for Developers

The urgency of the integration stems from strict European rules. Official European Commission guidance states that under Article 50 of the AI Act, all developers of generative models must attach machine-readable marking to synthetic content. Those rules took effect back on August 2, 2026, but developers were given a transition period until early December.

Global users, however, will not be affected by mandatory marking for now: in the consumer versions of ChatGPT and Codex, the system activates only for residents of EU countries. For developers using the OpenAI API worldwide, textGrain will be available as a separate option starting October 5, but will be off by default. The company plans to open-source the technology later to set a common industry standard.