Meta Plans to Limit Release of Risky AI SystemsMeta Plans to Limit Release of Risky AI Systems

Meta has published a new risk policy framework outlining how it plans to evaluate and mitigate the risk posed by new frontier AI models and when it would stop development, restrict access or not release systems.

In a bid to address growing safety concerns around AI, the Frontier AI Framework describes how Meta could categorize AI models into high risk and critical risk groups and then act accordingly to mitigate the associated risk to “tolerable levels.”

For example, critical risk is defined as being able to uniquely enable the execution of an outlined threat scenario. High risk means the model could provide a significant uplift towards execution of a threat scenario but does not enable execution.

Threat scenarios include the proliferation of high impact biological weapons, with capabilities equivalent to known agents and widespread economic damage to individuals or corporations via scaled long form fraud and scams.

For the models that reach the critical-risk threshold, Meta would stop development, restrict access to the model to a small number of experts, and input security protection to prevent hacking or exfiltration “insofar as is technically feasible and commercially practicable.”

For the high risk, it would limit access and implement mitigations to reduce risk to moderate levels, whereby the model would not provide significant uplift towards execution of a threat scenario. 

Meta said the risk assessment process involves multi-disciplinary engagement, including internal and, “where appropriate,” external experts from various disciplines and company leaders from multiple disciplines.

The new framework relates only to the company’s most advanced models and systems that match or exceed current capabilities. 

“We hope that sharing our current approach to development of advanced AI systems will not only promote transparency into our decision-making processes but also encourage discussion and research on how to improve the science of AI evaluation and the quantification of risks and benefits,” Meta said.

Its approach to evaluating and mitigating risks would evolve and mature over time. However, it added, due to the nascent nature of AI evaluation that is still very much in development, 

Meta said it plans to focus on improving the robustness and reliability of evaluations, including working to ensure that its testing environments produce results that reflect how the model will perform once in production, the company said.

“Our decision-making process for developing and releasing frontier AI is guided by our internal AI governance program, our risk thresholds, and the rigorous program of evaluation and mitigation that underpins them,” it added. 

This article first appeared in IoT World Today's sister publication AI Business.