Abstract
This report was developed in the context of the 'Responsible AI for Social Media Governance' Project, with the steering of the Project Co-Leads and the guidance of the Project Advisory Group, supported by the GPAI Responsible AI (RAI) Working Group. The GPAI RAI Working Group agreed to declassify this report and make it publicly available.
The new generation of general-purpose 'foundation AI models' such as ChatGPT and MidJourney are dramatically more powerful and useful than earlier AI systems. But their use also introduces a range of new risks, which have prompted an ongoing conversation about possible regulatory mechanisms. This paper contributes to this conversation. We propose a specific principle that should be incorporated into legislation—namely, that any organisation developing a new, state-of-the-art foundation model must demonstrate a reliable detection mechanism for the content that model produces, as a condition of release. The detection mechanism should be made publicly available in a tool that allows consumers to query, for an arbitrary item of content, whether the item was generated (wholly or partly) by the model. We argue this requirement is technically feasible and would play an important role in reducing certain risks from foundation models in many domains.