Safety Cases: How to Justify the Safety of Advanced AI Systems
This addresses the need for structured safety justifications for companies and regulators making decisions about training and deploying advanced AI systems, though it is incremental as it builds on existing safety case concepts.
The paper tackles the problem of justifying the safety of advanced AI systems by proposing a framework for creating 'safety cases' to argue that these systems are unlikely to cause catastrophes, with results including four categories of arguments and examples evaluated for deployment justification.
As AI systems become more advanced, companies and regulators will make difficult decisions about whether it is safe to train and deploy them. To prepare for these decisions, we investigate how developers could make a 'safety case,' which is a structured rationale that AI systems are unlikely to cause a catastrophe. We propose a framework for organizing a safety case and discuss four categories of arguments to justify safety: total inability to cause a catastrophe, sufficiently strong control measures, trustworthiness despite capability to cause harm, and -- if AI systems become much more powerful -- deference to credible AI advisors. We evaluate concrete examples of arguments in each category and outline how arguments could be combined to justify that AI systems are safe to deploy.