OpenAI Rolls Out Invisible Watermark for ChatGPT and Codex in the EU

OpenAI announced it will embed an invisible watermark into text produced by ChatGPT and Codex for users in the European Union, aligning with the EU AI Act’s transparency requirements that took effect on 2 August.

The watermark, called textGrain, subtly biases word‑choice probabilities so that a detector can recognise AI‑generated passages without any visible marker. The pattern travels with the text when copied, yet OpenAI says it does not reveal the identity of the user and does not measurably degrade model performance.

OpenAI published a technical report co‑authored with researchers from the University of Pennsylvania and Yale, describing how a secret key orders next‑token predictions to embed the signal. Tests show that modest editing—replacing roughly ten percent of words—can cut detection rates from about 92 % to 66 %, and short, mathematical, or translated snippets are harder to flag.

Because the detector is initially limited to approved researchers, OpenAI stresses that the absence of a watermark does not prove human authorship. The move follows Anthropic’s global rollout of a similar system, which sparked debate over creator attribution.

Why it matters for GPU / AI infrastructure

Watermarking adds a lightweight compliance layer that runs on the inference pipeline; for cloud GPU providers such as AiGpu it means ensuring that the serving stack can optionally attach the statistical signature without latency penalties. As regulators expand transparency rules, infrastructure that can toggle watermarking per‑region will become a competitive differentiator for enterprise customers who need audit‑ready AI outputs.