Thursday, October 22, 2026
RW|EN
INTERNATIONAL

ChatGPT to Introduce Hidden Watermarks in AI-Generated Text

ChatGPT to Introduce Hidden Watermarks in AI-Generated Text

OpenAI is preparing to introduce invisible watermarks in text generated by its artificial intelligence tools, ChatGPT and Codex, in countries belonging to the European Union (EU).

The move aims to comply with transparency requirements governing the use of artificial intelligence under the EU AI Act, whose relevant provisions began taking effect on August 2, 2026.

Unlike traditional watermarks that can be seen by the human eye, these invisible marks work through subtle changes in the way AI selects and arranges words. The changes are designed to help OpenAI identify text generated by its AI tools using a detection system reportedly called textGrain.

However, the technology is not completely reliable. OpenAI has indicated that even small changes to an AI-generated text can affect the system’s ability to identify it. For example, replacing just 10 percent of the words in a text with other words that have similar meanings could reduce the detection rate from 92 percent to 66 percent.

The system may also struggle to identify watermarks in short texts, translated content and mathematical answers. Because of these limitations, access to textGrain is currently expected to be restricted to authorised experts.

The development comes two months after OpenAI’s competitor, Anthropic, announced plans to introduce a similar watermarking system for AI-generated text worldwide. The announcement attracted mixed reactions from users, with some raising concerns about how such technology could affect their use of AI tools.

OpenAI has reportedly been working on text watermarking for a long time but had previously hesitated to introduce it because of concerns that the move could discourage users from using its products.

The introduction of invisible watermarks could become another step towards making AI-generated content easier to identify. However, the technology’s limitations raise questions about how reliably it can distinguish AI-generated writing from human-written content, particularly when texts are edited, translated or substantially rewritten.

Brenna AKARABO
RADIOTV10

Share this story: Facebook Twitter WhatsApp

Leave a Reply

Your email address will not be published. Required fields are marked *

15 − three =