OpenAI to Implement Invisible Watermarking for ChatGPT and Codex Outputs in the European Union

How the Watermarking Technology Functions

The watermarking mechanism is designed to help distinguish AI-generated text from human-written content without altering the user experience. OpenAI emphasized that the technology does not affect the quality, speed, or accessibility of the models. The move aligns with broader efforts to increase transparency around AI use, particularly in contexts where distinguishing synthetic from authentic text is critical, such as education, journalism, and legal documentation.

The invisible watermark works by subtly adjusting word choices during text generation in a pattern detectable only by machine learning classifiers. These adjustments are statistically insignificant to human readers but create a signature that verification tools can identify. OpenAI has not disclosed the specific algorithm used but confirmed it applies uniformly across eligible outputs in the EU region. The company noted that similar techniques have been tested internally for over a year before this public deployment.

What Happens if Users Modify the Watermarked Text?

If users edit or paraphrase AI-generated text, the watermark may degrade or become undetectable, depending on the extent of changes. OpenAI acknowledged that minor edits could preserve the signal, while substantial rewriting might remove it entirely. The company stated it is monitoring how modifications affect detectability and may adjust the system based on real-world performance. However, it stressed that the primary goal is compliance with regulatory expectations, not enforcing usage restrictions.

Frequently Asked Questions

Will the watermark affect how ChatGPT or Codex responds to prompts? No, the watermark is applied only to the output text and does not influence the model’s Can users opt out of the watermarking feature? Currently, OpenAI has not provided an opt-out option for users in the EU, as the measure is tied to legal compliance under the EU AI Act.

Is this watermarking system being rolled out outside the European Union? At this time, the invisible watermark is limited to users located in the European Union. OpenAI has not announced plans to extend the feature to other regions.