The system, called Text Grain, creates a specific pattern in word selection that remains invisible to readers but can be detected by computerised tools. OpenAI says the watermark can remain intact even after text is copied and pasted.

The technology is initially being introduced in the European Union, where the AI Act requires providers of generative AI systems to make AI-generated content identifiable by machines.

OpenAI says there are currently no plans to make the system mandatory worldwide. The company will use the initial rollout to study how the technology performs in real-world use and how users respond to it.

The company says the watermark does not identify the user or reveal the prompt used to generate the text. It only indicates that an AI watermark may be present in the content.

However, OpenAI does not consider the system 100 percent reliable. Detection can be more difficult in shorter texts and in areas such as mathematics, where there is less flexibility in word choice.

OpenAI's tests found that watermark detection was around 80 percent in psychology texts containing about 200 tokens, rising to around 95 percent in texts of about 400 tokens.

The watermark can also become weaker if the text is edited. According to OpenAI's tests, replacing around 10 percent of the words with synonyms reduced detection from about 92 percent to 66 percent, while changing 25 percent of the words reduced detection to around 17 percent.

Translation into another language can also make the watermark harder to detect.

This means that the absence of a detected watermark does not necessarily prove that a text was written by a human.

OpenAI says Text Grain is intended only to help identify potentially AI-generated text. It does not determine whether the information in that text is accurate or false, nor does it establish how much of the content was written or edited by a human.

The move comes as AI-generated articles, academic work and social media content continue to increase, intensifying the need for reliable ways to distinguish AI-generated material from human writing.