A socio-technical lens on foundation models

Introduction

The rapid roll-out, commercialisation and adoption of foundation models (general purpose AI applications based on broad data) across domains has rightfully triggered a range of AI governance initiatives. These interventions seek to tackle the harmful impacts exacerbated – and even generated – by these technologies. However, these governance initiatives, largely emerging from the global north, have so far focused on concerns of technical safety and speculative risks. This narrow lens fails to account for the growing body of evidence highlighting the range of sociotechnical risks and harms posed by the proliferation of foundation models.  

Drawing on a review of multi-disciplinary foundation models literature, we developed a typology which accounts for the socio-technical impacts of these technologies. We identify 14 categories of risks and harms across three core areas: individual, social and biospheric. This typology provides a more expansive framework through which the impacts of foundation models can be identified, and technical and normative interventions can be undertaken.

Alongside this impacts mapping, we are conducting an in-depth qualitative analysis of emerging governance initiatives related to foundation models and generative AI. This ongoing work aims to evaluate how the various risks posed by foundation models are being addressed globally by different policy and technical interventions, across their entire value chain.