---
title: OpenAI starts embedding watermarks in text from ChatGPT and Codex under EU AI Act
url: https://www.elseif.net/openai-starts-embedding-watermarks-in-text-from-chatgpt-and-codex-under-eu-ai-act
published: 2026-10-08T21:12:18+00:00
language: en
section: Models
source: https://www.blognone.com/node/151829
organizations: openai, anthropic
publisher: elseif
---

# OpenAI starts embedding watermarks in text from ChatGPT and Codex under EU AI Act

OpenAI has begun embedding invisible watermarks in the output text of ChatGPT to comply with the European Union AI Act. This follows Anthropic, which implemented similar technology in August. The technology, known as textGrain, applies statistical "signals" to words chosen by the model, resulting in a slight decrease in output performance.

Detection methods will examine these signals for accuracy. OpenAI claims this technology outperforms Google's SynthID and other approaches in the market. However, OpenAI acknowledges that shorter texts are harder to detect, particularly in contexts with limited vocabulary, such as mathematics. Additionally, any modifications made to the text afterward may reduce detection effectiveness.

Users of ChatGPT and Codex within the European Union will have the watermark embedded without the option to opt-out, while users outside the EU accessing via API will still have the option to opt-in.