OpenAI is preparing to launch its digital watermark technology to comply with the European Union’s new transparency rules. This step aims to make text and code generated by generative AI models identifiable. The company announced that it will utilize a special system called textGrain in this process. This technology helps determine the source of the content by adding invisible statistical signals to the models’ word choices.

textGrain Technology and Working Principle

The textGrain system, developed by OpenAI, stands out as a method for detecting content generated by artificial intelligence. The system proves that the content is an AI creation by embedding an invisible statistical signal into the model’s word preferences.

However, the company warns that strong performance under ideal conditions does not always guarantee reliable results in daily use. Short texts and mathematical content, in particular, exhibit a more difficult detection rate.

According to OpenAI data, the success rate for this type of content remains at approximately 80 percent. Edits made to the text are also among the factors that reduce the success rate of watermark detection.

EU AI Law and Implementation Process

Article 50 of the European Union’s AI Law requires generative AI providers to make their text outputs identifiable in a machine-readable format. This regulation covers all existing organizations as well as new market entrants.

Companies like OpenAI and Anthropic, as well as giants like Microsoft, Google, and Meta, are required to comply with these rules. The deadline for existing organizations to comply is December 2nd.

The watermark feature will be enabled by default for eligible ChatGPT and Codex outputs in the European Union. Users in other regions will be able to opt in for this feature for suitable models.

OpenAI is initially taking a more cautious approach regarding access to the detection software. Initially, the use of the software will be limited to approved researchers and specialist organizations.

The company emphasizes that this detector does not identify the user or reveal any random requests or conversations. Furthermore, it is planned to make the developed technology open source, allowing others to work on it as well.

OpenAI has not yet shared clear details about which models will be made available for general use. This technological development is considered a significant milestone in the transparency of AI content.

Do you think these kinds of invisible watermarks added to AI-generated content will be sufficient to ensure security in the digital world?

OpenAI is deploying its digital watermark technology under the EU’s new transparency rules. A new era for AI content detection.