Affordance Analyses of AI Safety Policies: A Proof-of-Concept using OpenAI's Preparedness Framework

Abstract

Prominent AI companies are producing risk management frameworks as a type of voluntary self-regulation. As interventions, these policies purport to establish risk thresholds and safety procedures for the development and deployment of highly capable AI. Understanding which AI risks are covered and what actions are allowed, refused, demanded, encouraged, or discouraged by these risk management policies is a foundational step in assessing how they might govern the development and deployment of AI systems in practice. To investigate how such policies are operationalised, we introduce a transferable AI policy analysis method based on the Mechanisms & Conditions model of affordances (M&C) and the MIT AI Risk Repository. We illustrate the utility of this method by applying it to OpenAI's "Preparedness Framework Version 2"(April 2025). We find that OpenAI's safety policy requests evaluation of a small minority of AI risks, encourages deployment of systems with "Medium"capabilities for unintentionally enabling "severe harm"(which OpenAI defines as > 1000 deaths or > $100B in damages), and allows OpenAI's CEO to deploy even more dangerous capabilities. These findings suggest that effective mitigation of AI risks requires more robust governance interventions beyond current industry self-regulation - of which there are well-established models in other domains. In addition, we illustrate how our affordance analysis provides a replicable method for evaluating what AI policies permit versus what they claim. Applied broadly, our AI policy analysis method will help clarify what is needed to mitigate AI-enabled harms and to facilitate trustworthy and socially beneficial AI systems. Show more