OpenAI's recent announcement of a comprehensive Preparedness Framework underscores a critical shift in AI safety and risk management. This move comes as the tech community acknowledges the current shortfall in managing frontier AI risks.
The Preparedness Framework outlines OpenAI's strategy to track, evaluate, forecast, and mitigate potential catastrophic risks arising from the development of increasingly powerful AI models. The Preparedness team, an integral part of this initiative, is tasked with the monumental responsibility of ensuring the safety of these frontier AI models.
They work in tandem with various safety and policy teams at OpenAI, each focusing on specific aspects of AI safety. The Safety Systems team addresses the misuse of current models and products, including ChatGPT, while the Superalignment team works on foundational safety aspects of future superintelligent models.
This collaborative effort aims to create a comprehensive safety net around the burgeoning field of AI.
Strategic Risk Management and Oversight
A critical aspect of the Preparedness Framework involves defining risk thresholds that trigger essential safety measures.
OpenAI has outlined four initial risk categories: cybersecurity, CBRN (chemical, biological, radiological, nuclear threats), persuasion, and model autonomy. The system specifies four safety risk levels, allowing only those models with a post-mitigation risk score of "medium" or lower to be deployed.
Models with a "high" post-mitigation score can continue development but under stricter scrutiny. The Preparedness team will also spearhead technical evaluations and report synthesis, informing OpenAI's decision-making for safe model development and deployment.
This effort is bolstered by the establishment of a cross-functional Safety Advisory Group, which will review these reports and provide feedback to both the Leadership and the Board of Directors. The Board holds the authority to reverse any decision, ensuring an additional layer of oversight.
Furthermore, OpenAI is committed to developing protocols for enhanced safety and external accountability. Regular safety drills, urgent issue responses, and independent third-party audits are part of this rigorous approach.
Collaborations with external entities and continuous processes to identify emerging risks, including the elusive "unknown unknowns," demonstrate OpenAI's dedication to preemptively addressing AI safety challenges.