OpenAI Tags ChatGPT Texts to Ensure Traceability

OpenAI introduces a statistical watermark in the texts generated by its models, aimed at certifying their origin. The company seeks compliance with the European transparency law but acknowledges blind spots regarding short texts and modified content.

Short texts and edits reduce detection, and the option will not be enabled by default

According to OpenAI, it is more complex to identify texts containing fewer than 200 words. Modifying a text can also affect the watermark: the company states that for a 400-word text, changing 10% of the content reduces the detection rate from 99% to 66%. For now, only developers have access to this feature. When it is rolled out to users, it will not be enabled by default.

An invisible statistical signal embedded in word choice

The watermark takes the form of an invisible signal integrated into the choice and order of words in a text. OpenAI describes its technology, textGrain, as the addition of an invisible statistical signal to the model's word choices. A detector searches for this signal to determine if a passage contains an OpenAI digital watermark. This watermark pertains to texts generated by ChatGPT and Codex.

Deployment in Europe, transparency required, and targeted uses

OpenAI links this initiative to the new European transparency law, which mandates marking any content generated by AI. The marking, already applied to images and videos, will now extend to text, with deployment planned for ChatGPT users in the European Union. The goal is to enhance transparency regarding the use of AI text generators. The issue extends beyond the technical realm: writer Thélysson Orélien has been accused of using AI for his novel based on a detection tool, despite these systems not being infallible and the debate remaining open. Cited use cases include applications received by employers and assignments submitted to teachers, with the ability to identify the origin of a text being presented as an increasingly important issue.