Beware of Being Caught: Using ChatGPT for Work Now Leaves Clear Traces
The text outputs from ChatGPT and Codex will be augmented with invisible watermarks. OpenAI has confirmed this addition is specifically for users within the European Union.
This AI modification serves as a method to ensure compliance with European AI regulations, which are set to be implemented on 2 August 2026. The regulations from the European continent require AI-generated content to utilise specific identifiers.
The watermark will be rolled out across OpenAI’s products over the coming weeks. This applies to all eligible users across all subscription tiers, as reported by TechCrunch on Tuesday.
For developers, the implementation can be activated immediately, although the feature will remain disabled by default. OpenAI also clarified that the implementation of this watermark will not be available globally.
The watermark in question will remain invisible to readers but is engineered to be detectable by detection software. TechCrunch reports that the watermark will reside within the text itself, meaning it will persist even when the text is copied to another location.
OpenAI has also released a technical report regarding the method, known as ‘textGrain’, which was co-authored with researchers from the University of Pennsylvania and Yale. The report discusses the use of a secret key; the system sequences word predictions to form a sentence, allowing detectors to identify AI-generated content using the text and the secret key.
OpenAI confirmed that editing the text can remove the watermark and the company has conducted tests regarding this. One test attempted to replace 10% of the text with synonyms, which resulted in detection rates dropping from 92% to 66%.
However, the system may struggle to detect certain types of text, such as short passages, mathematical answers, and translations. Even if a watermark is absent, it does not necessarily mean the text was human-authored, as the lack of a watermark could result from the text being too short, heavily edited, or produced by another company.
“The watermark indicates that OpenAI’s system generated or partially processed the content, but it does not indicate the extent of human judgement, editing, or creativity involved,” stated OpenAI.
OpenAI’s implementation follows a similar move by Anthropic two months prior, though Anthropic’s watermark for Claude-generated text is applied globally.