Meta has published its new Frontier AI Framework, which outlines the company's approach to evaluating and mitigating risks associated with advanced AI models. The framework details the criteria Meta will use to classify AI systems based on risk levels and the corresponding actions it will take, including halting development, restricting access, or not releasing systems.

Meta's new AI policy addresses growing safety concerns surrounding AI, especially regarding the potential for misuse and harmful outcomes. The Frontier AI Framework outlines how Meta categorizes AI models into "high-risk" and "critical-risk" groups and how it plans to mitigate the risks associated with each category to "tolerable levels".

AI models classified as "critical risk" are those that could uniquely enable the execution of a defined threat scenario. In such cases, Meta would halt development, restrict access to a small group of experts, and implement security measures to prevent hacking or data exfiltration.

High-risk AI models are those that could significantly increase the likelihood of a threat scenario but do not enable its execution17. For these models, Meta would limit access and implement measures to reduce the risk to moderate levels.

The threat scenarios considered by Meta include:

  •       The proliferation of high-impact biological weapons.
  •       Widespread economic damage to individuals or corporations through large-scale fraud and scams.
  •       Automated compromise of corporate-scale environments.

Meta acknowledges that the list of potential catastrophes is not exhaustive but includes those it believes are the most urgent and likely to arise from releasing a powerful AI system.

Meta's risk assessment process involves internal and external experts from various disciplines and company leaders. The company emphasizes that its approach to evaluating and mitigating risks will evolve over time. Meta also plans to focus on improving the robustness and reliability of its evaluations to ensure that testing environments accurately reflect how the models will perform in real-world scenarios.

If a system is determined to be high-risk, Meta will limit its access internally and will not release it to the public until risk mitigation strategies are implemented to reduce the risk to moderate levels. For critical-risk systems, Meta will halt development and implement security protections to prevent misuse.

The new framework applies only to Meta's most advanced models and systems that match or exceed current capabilities. Meta hopes that sharing its approach will promote transparency and encourage discussion and research on improving AI evaluation and risk quantification.