The 2025 OpenAI Preparedness Framework does not guarantee any AI risk mitigation practices: a proof-of-concept for affordance analyses of AI safety policies

The framework permits deployment of systems with significant harm potential: OpenAI's Preparedness Framework allows deployment of AI systems rated as having "Medium" capabilities for causing severe harm, which the company defines as incidents resulting in over 1,000 deaths or more than $100 billion in damages.

Only a small fraction of known AI risks are addressed: The analysis found that the framework requests evaluation of only a minority of AI risks documented in the MIT AI Risk Repository, leaving many potential harms unmonitored or unmanaged.

CEO override authority undermines safety guardrails: The framework permits OpenAI's CEO to deploy AI systems that are even more dangerous than those classified as "Medium" risk, creating a loophole that could bypass established safety procedures.

Self-regulation alone may be insufficient: The researchers conclude that "effective mitigation of AI risks requires more robust governance interventions beyond current industry self-regulation," suggesting that voluntary company frameworks are inadequate safeguards.

There's a gap between stated commitments and actual practices: The study demonstrates a replicable method showing that what safety frameworks claim to do often differs significantly from what they actually permit, highlighting the importance of scrutinizing the specific language and loopholes in AI safety policies.