FamCheck. Everyone who looks after your kids stays in the loop. $9.99 a month for the whole household, first week free, no card. famcheck.coSPONSORED
The Miami Art Journal

Tech

OpenAI To Watermark ChatGPT Text In Europe

Invisible markers added to generated content to meet EU regulations

The Editors ·

OpenAI To Watermark ChatGPT Text In Europe

OpenAI announced that it will begin adding an invisible watermark to text generated by ChatGPT and Codex within the European Union. This move is designed to comply with the transparency rules established by the EU AI Act. The company detailed the implementation in a blog post released on Monday, confirming that the feature will roll out over the coming weeks to eligible users on all plans in the region. The measure applies specifically to text outputs and does not currently extend to other forms of media generated by the models.

The EU AI Act’s transparency provisions took effect on August 2. These rules require artificial intelligence companies to mark content in a manner that allows other systems to identify it as machine-generated. OpenAI’s approach involves subtly shaping the model’s word choices rather than inserting a visible symbol. This technique creates a pattern within the text itself. The pattern is invisible to human readers but can be detected by specific software tools. Because the watermark is embedded in the linguistic structure, it travels with the text when it is copied or pasted into other documents.

Developers using OpenAI’s API globally have the option to enable this feature for select models starting today. The setting is off by default for API users worldwide. OpenAI stated that it is not making text watermarking a global default at the time of launch. The company published a technical report alongside the announcement to explain the methodology. The report, titled textGrain, was co-written with researchers from the University of Pennsylvania and Yale University. It describes a process that uses a secret key to sort next-word predictions, allowing detectors to identify AI-generated content using only the text and the key.

Testing indicates that the watermark can be removed or weakened through editing. In one test, replacing ten percent of words with synonyms reduced detection rates from approximately ninety-two percent to sixty-six percent. The company also noted that short passages, mathematical answers, and translated text are more difficult to detect. Due to these limitations, OpenAI decided to provide initial detector access only to approved researchers and expert organizations. These groups will help evaluate the reliability and responsible uses of the technology.

The company cautioned that the absence of a watermark does not prove that text was written by a human. Content may be too short, heavily edited, or generated by a different company’s AI system. OpenAI emphasized that the watermark does not identify the user who requested the generation. The company reported seeing no meaningful change in its models’ performance with the feature switched on. This technical adjustment represents a significant step in the integration of compliance measures for large language models operating in regulated markets.

Takeaway: Developers can enable the watermark for select models via the API starting today, though it remains off by default.

Source: TechCrunch AI