OpenAI to Introduce Invisible Watermarks for ChatGPT in the EU

OpenAI Introduces Invisible Watermarks for ChatGPT Text in the EU: Why It’s Needed and Whether It Can Be Bypassed

OpenAI, the developer of the popular language model ChatGPT, has begun integrating invisible watermarks into text generated by its systems in the European Union. This new feature, called textGrain, is designed to increase transparency around the use of artificial intelligence (AI) in line with the EU’s new regulatory requirements.

The feature is being gradually rolled out to all ChatGPT and Codex users over the coming weeks, regardless of plan. The rollout is tied to the European Artificial Intelligence Act, which requires AI-generated content to be labeled.

How textGrain will work

The textGrain technology works by introducing imperceptible changes to word choice during text generation. These changes create a unique statistical pattern that can be recognized by special software. Because the watermark is built directly into the structure of the text, it remains even after copying and pasting.

OpenAI says the presence of such a watermark will only indicate that an OpenAI system was involved in creating or processing the text, but will not determine the degree of human involvement.

Sensitivity to editing and access to the technology

OpenAI research indicates that the watermark is somewhat vulnerable to manual editing. For example, replacing about 10% of the words with synonyms can reduce the likelihood of detecting the AI trace from 92% to 66%. The algorithm also has more difficulty detecting the marking in short passages, mathematical formulas, and translations.

Given these limitations, detection tools for textGrain will initially be available exclusively to accredited researchers and expert organizations to assess its reliability.

Regulatory pressure and a competitive trend

The launch of text labeling in the EU is a direct response by OpenAI to the introduction of the European Artificial Intelligence Act. This act sets clear rules for transparency in AI use, including a requirement to label machine-generated content in a format that can be automatically recognized.

Other major technology companies are taking a similar approach. Recently, Anthropic, the developer of the Claude model, announced a comparable rollout of invisible watermarks for its content starting in August 2026. Google has also begun labeling ads created or edited with AI, while Meta already uses the AI info feature in the About this ad section.

These steps point to a broader trend in the artificial intelligence industry toward greater accountability and transparency in the use of its technologies, in line with both regulatory requirements and public expectations.