Hidden Codes in Text Generated by ChatGPT, Claude and Gemini

Jakarta, CNBC Indonesia — Think twice before copying text created by AI. Some chatbot models are now equipped with invisible watermark codes to identify AI-generated content.

The European Union’s AI Act requires AI-generated text to carry specific identification. One way to do this is by embedding a watermark in the text.

Anthropic introduced this feature in Claude earlier this year. It is available globally, not only in Europe.

OpenAI has made a similar decision. Unlike Anthropic, however, ChatGPT and Codex will provide invisible watermarks only for users in the European Union.

“Over the next few weeks, we’re adding invisible watermarks to eligible ChatGPT and Codex text in the European Union,” OpenAI said in an announcement earlier this month, as quoted by TechRadar on Wednesday (14 October 2026).

OpenAI uses a system called textGrain to generate these identifiers. The system adds an invisible statistical signal to text generated by its AI models.

TechRadar reported that a specialised tool is required to detect the watermark. Unfortunately, the tool is not yet publicly available.

So far, OpenAI has given access to the detector only to selected researchers and expert organisations. They must apply to use it.

OpenAI has also said the detector is prone to errors. It may struggle to identify short or limited text, as well as text that has been edited.

The company added that it would not release the tool publicly at launch because of the risk of detection errors.

“The tool will not identify users or reveal the prompts they used or their conversations,” OpenAI explained.