Granting Non-AI Experts Creative Control Over AI Systems

Abstract

Many harmful behaviors and problematic deployments of AI stem from the fact that AI experts are not experts in the vast array of settings where AI is applied. Non-AI experts from these domains hold promising potential to contribute their expertise and directly design the AI systems that impact them, but they face substantial technical and effort barriers. Could we redesign AI development tools to match the language of non-technical end users? My research develops novel systems allowing non-AI experts to define AI behavior in terms of interpretable, self-defined concepts. Monolithic, black-box models do not yield such control, so we introduce techniques for users to create many narrow, personalized models that they can better understand and steer. We demonstrate the success of this approach across the AI lifecycle: from designing AI objectives to evaluating AI behavior to authoring end-to-end AI systems. When non-AI experts design AI from start to finish, they notice gaps and build solutions that AI experts could not—such as creating new feed ranking models to mitigate partisan animosity, surfacing underreported issues with content moderation models, and activating unique pockets of LLM behavior to amplify their personal writing style.

Ähnliche Arbeiten