OpenAI plans to watermark ChatGPT and Codex text output in EU amid AI transparency push
OpenAI says the watermarks are embedded in the text and designed to be invisible to the human eye, while AI detection tools can stop ChatGPT- or Codex-generated content.
- Oct 6, 2026,
- Updated Oct 6, 2026 10:14 AM IST
OpenAI has announced that it will add an invisible watermark to text content generated by ChatGPT and the coding tool Codex. The decision comes to comply with the EU AI Act, which means the watermarking will be limited to users in the European Union, and no global release is planned as of now.
OpenAI says the watermarks are embedded in the text and designed to be invisible to the human eye, while AI detection tools can stop ChatGPT- or Codex-generated content. Therefore, if someone copies and pastes ChatGPT-generated text elsewhere, the watermark remains with the text.
Must read: ‘Not nearly careful enough’: Former OpenAI employee quits amid concerns over AI safety
How does OpenAI’s text watermark work?
OpenAI revealed that it uses a watermarking method called textGrain, which uses a secret key to subtly influence which words ChatGPT chooses when generating text, according to a separate technical report. Therefore, an AI detector will need s only the text and the secret key to check for the watermark.
But is there any way to remove the text watermark? Well, OpenAI said that editing the text can make the watermark harder to detect. For example, changing just 10% of the words to synonyms reduced detection accuracy from around 92% to 66%. The company also revealed that the watermarking system struggles with very short pieces of text, math answers, and translated text.
“These limitations contribute to our decision to provide initial detector access only to approved researchers and expert organizations, who can help us evaluate reliability and responsible uses,” said the company. However, the watermarking system does not impact the performance of the AI model.
Must read: OpenAI faces lawsuit over Hugging Face system breach incident: What you need to know
The company further added, “[Watermarks] can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it.”
What the EU’s AI Act says about text watermarking?
The EU AI Act requires AI companies to make AI-generated content identifiable in a way that machines can read and detect. To comply, OpenAI will start rolling out its textGrain watermarking system to ChatGPT users in the EU over the coming weeks. In August, Anthropic also released a watermarking system for Claude texts after the AI Act came into force on August 2, 2026.
Initially, the text watermark will be available in the EU, but OpenAI is also making the technology available globally to API customers as an opt-in feature starting today, but only on certain models.
For Unparalleled coverage of India's Businesses and Economy – Subscribe to Business Today Magazine
OpenAI has announced that it will add an invisible watermark to text content generated by ChatGPT and the coding tool Codex. The decision comes to comply with the EU AI Act, which means the watermarking will be limited to users in the European Union, and no global release is planned as of now.
OpenAI says the watermarks are embedded in the text and designed to be invisible to the human eye, while AI detection tools can stop ChatGPT- or Codex-generated content. Therefore, if someone copies and pastes ChatGPT-generated text elsewhere, the watermark remains with the text.
Must read: ‘Not nearly careful enough’: Former OpenAI employee quits amid concerns over AI safety
How does OpenAI’s text watermark work?
OpenAI revealed that it uses a watermarking method called textGrain, which uses a secret key to subtly influence which words ChatGPT chooses when generating text, according to a separate technical report. Therefore, an AI detector will need s only the text and the secret key to check for the watermark.
But is there any way to remove the text watermark? Well, OpenAI said that editing the text can make the watermark harder to detect. For example, changing just 10% of the words to synonyms reduced detection accuracy from around 92% to 66%. The company also revealed that the watermarking system struggles with very short pieces of text, math answers, and translated text.
“These limitations contribute to our decision to provide initial detector access only to approved researchers and expert organizations, who can help us evaluate reliability and responsible uses,” said the company. However, the watermarking system does not impact the performance of the AI model.
Must read: OpenAI faces lawsuit over Hugging Face system breach incident: What you need to know
The company further added, “[Watermarks] can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it.”
What the EU’s AI Act says about text watermarking?
The EU AI Act requires AI companies to make AI-generated content identifiable in a way that machines can read and detect. To comply, OpenAI will start rolling out its textGrain watermarking system to ChatGPT users in the EU over the coming weeks. In August, Anthropic also released a watermarking system for Claude texts after the AI Act came into force on August 2, 2026.
Initially, the text watermark will be available in the EU, but OpenAI is also making the technology available globally to API customers as an opt-in feature starting today, but only on certain models.
For Unparalleled coverage of India's Businesses and Economy – Subscribe to Business Today Magazine