OpenAI Launches Hidden Watermark for ChatGPT Text in Europe

In a move aimed at enhancing transparency around AI-generated content, OpenAI has announced that it will begin embedding an invisible watermark within the text produced by both ChatGPT and its Codex coding tool within European Union countries. This decision comes in response to the requirements of the European AI Act, which obliges tech companies to include clear, detectable markers that distinguish automatically generated content from human-created content.

What Does the European Law Require?

The transparency rules set out in the European AI Act came into effect at the beginning of last August. These rules require AI system developers to mark the content produced by their tools in a way that allows other systems to recognize it. With this, OpenAI joins the list of major companies that have committed to the European Code of Practice on generated content, including Google, Meta, Microsoft, and Anthropic.

How Does the Watermark Work?

Contrary to what one might imagine, this watermark is not a visible symbol added to the text. Instead, it works by subtly influencing the model's word choices, leaving behind a hidden pattern that a reader cannot notice but that a specialized detector can identify. Because this pattern resides within the words themselves, it travels with the text when it is copied and pasted elsewhere.

The company confirmed that the watermark does not reveal the user's identity, and that it has not observed any noticeable decline in the performance of its models when it is activated. It explained that it will not make this feature a default setting at the global level upon launch.

Availability and Scope of Application

The feature is expected to be rolled out gradually over the coming weeks to eligible ChatGPT and Codex users across all plans, but only within the European Union. As for developers who use the company's API anywhere in the world, they will be able to enable it for specific models starting now, while it remains disabled by default.

Technical Report in Collaboration With Two Universities

OpenAI published a technical report explaining its methodology, called textGrain, which it prepared in collaboration with researchers from the universities of Pennsylvania and Yale. The report presents an example illustrating the use of a secret key to arrange the predictions for the next word in a sentence. With hundreds of these subtle adjustments accumulating, the detector becomes able to identify generated content based on the text and the key alone.

Limitations of the Technology and Its Removability

OpenAI acknowledged that the watermark may weaken or disappear when the text is modified. Its tests showed that replacing 10% of the words with synonyms reduced detection accuracy from about 92% to 66%. It also noted that some types are difficult to detect, including:

• Very short texts.
• Mathematical answers.
• Translated texts.

Because of these limitations, the company decided to make the detection tool available in its initial phase only to accredited researchers and expert institutions, to help assess its reliability and responsible uses.

Absence of the Watermark Does Not Mean Human Authorship

OpenAI cautioned that the absence of the watermark does not prove that the text was written by a human, since the text may be extremely short, have undergone extensive modifications, or been produced by an AI system belonging to another company. It added that the watermark may indicate that one of its systems produced or processed part of the text, but it does not determine the extent of human intervention, creativity, or editing involved.

A Broader Context Among Competitors

This announcement comes two months after Anthropic announced its intention to watermark the text of its assistant Claude worldwide, a decision that drew objections from some users who felt that they are the ones providing the instructions, context, and decisions, while the model remains merely a tool. It is worth noting that OpenAI had previously developed a text watermark but delayed its release, partly out of fear that users would switch to competitors that do not adopt this technology.

✦ بقلم فريق دروب أيديا