Artificial Intelligence

OpenAI Adding Invisible Watermarks to EU Text and Code

OpenAI is preparing to deploy invisible watermarks for text generated by ChatGPT and Codex within the European Union, aligning with new regional transparency requirements.

Radwan AbdulhadiEditor and technology writer

Published: October 6, 2026 · Updated: October 6, 2026 · 3 min read

OpenAI Adding Invisible Watermarks to EU Text and Code — Image: BleepingComputer

Complying with EU Transparency Rules

OpenAI is preparing to introduce invisible watermarks to text and code generated by ChatGPT and Codex within the European Union, as [explained] in recent company documentation. This move is designed to comply with the European Union's new AI transparency rules, specifically addressing mandates that require generative AI system providers to make text outputs identifiable in a machine-readable way.

While the technology is being automatically implemented for eligible regional outputs, OpenAI is not making watermarking a global default. However, starting immediately, API developers worldwide can opt in to watermarking for supported models.

How textGrain Technology Works

The underlying mechanism, called textGrain, does not alter the appearance of the text when read or copied by users. Instead, it slightly adjusts the model's word choices to generate a distinct statistical pattern that specialized detectors can later recognize.

OpenAI notes that the proprietary system matches or exceeds the performance of alternative frameworks such as [SynthID for text], though the company emphasizes that ideal lab conditions do not guarantee seamless everyday reliability.

OpenAI will add a digital watermark to text and code generated in the EU — Image related to the report from Engadget · Source: Engadget

Limitations and Editing Vulnerabilities

Despite its utility under controlled evaluations, text watermarking remains vulnerable to basic text manipulation. OpenAI's internal tests indicate that normal editing significantly diminishes detection capabilities. For instance, in an assessment of 400-token passages, substituting just 10 percent of words with synonyms dropped detection rates from approximately 92 percent down to 66 percent.

When 25 percent of the words were replaced, the detection rate fell sharply to 17 percent. Shorter passages also pose challenges; detection stood at roughly 95 percent for 400-token psychology responses, but dropped to about 80 percent for 200-token texts. Furthermore, performance suffered in academic domains like mathematics, where models have fewer alternative phrasing options.

Image related to the report from Engadget · Source: Engadget

Detector Access and Privacy Safeguards

OpenAI is opening applications for its watermark detector, but initial availability will be strictly limited to approved researchers and expert organizations. The organization has also clarified privacy boundaries regarding the scanning process.

According to the company, a detected watermark does not expose the user's account, specific prompts, conversation history, or the exact proportion of human versus machine contribution in the final text. OpenAI also cautions that the absence of a positive watermark detection does not definitively prove human authorship.

Sources

BleepingComputerOpenAI is adding invisible watermarks to ChatGPT and Codex text in the EU
EngadgetOpenAI will add a digital watermark to text and code generated in the EU