executive summary
The productive impact of cutting-edge machine learning (ML) has been, and will continue to be, significantly limited by the difficulty or inability to integrate it with safety-critical applications (those where the failure of an operational system could harm individuals, the public, or the environment). The safety-critical domain is the interface where issues plaguing state-of-the-art ML in low-risk domains meet, primarily revolving around issues of reliability, human interpretability, and the ability to intervene in the internal mechanisms of operational systems. Current general-purpose ML models fail to meet adoption standards in safety-critical areas, contributing to the ever-present sense that their social and economic impact is modest.
ML is forcing the world to adapt in its most versatile form, but it must change to succeed in the most sensitive application areas. Such changes should continue the lineage of some of the most influential engineering wonders in recent history by offering: Guaranteed performance in safety-critical areasminimize the possibility of harmful output and reduce its severity.
Policy recommendations
- The National Artificial Intelligence Initiative Office (NAIIO), in collaboration with the Networking and Information Technology Research and Development (NITRD) Subcommittee, should direct the National Artificial Intelligence Research and Development Interagency Task Force to coordinate a total investment of $1.5 billion over five years in safety-critical neural symbolic AI research and development.
- The Center for AI Standards Innovation (CAISI) at the National Institute of Standards and Technology (NIST) should inform, but not exclusively guide, NAIIO’s investment coordination by convening a working group to establish a research agenda for new metrics specific to safety-critical neural semiotic systems.
- A separate working group convened by NIST’s CAISI must determine whether autonomous readiness levels (ARLs) for non-defense, safety-critical applications are warranted at their current stage of development. These are evaluation metrics specific to models that are expected to perform autonomously over a range of potentially changing situations in performing human inputs. AI “agents” may be considered part of this target group.
- The National Science Foundation (NSF) should parallel NAIIO’s efforts by establishing a National AI Institute dedicated to specific application areas of neurosemiotic AI. It requires an initial investment worth $80 million to $100 million over five years.
The views expressed in this article are those of the author and do not represent the official policy or position of New Lines Institute.
