AI Safety Subproblems for Software Engineering Researchers

AI Safety is underrepresented in software engineering research: The paper's quantified literature review found that AI Safety discussions are uncommon at SE venues, suggesting a significant gap between the two research communities despite their interconnection.

Software engineering practices will need to evolve as AI capabilities increase: The authors conjecture that software development itself will change substantially as AI systems approach or exceed human-level capabilities, requiring SE researchers to anticipate and prepare for these shifts.

AI Safety challenges map onto traditional software engineering problems: The paper categorizes AI Safety as a collection of "subproblems" that fit into established SE areas, suggesting that existing SE expertise in areas like testing, verification, and system design could be directly applied to making advanced AI systems safer.

The path to advanced AI (AGI/HLMI) creates urgent stakes for software quality: The authors frame the challenge as avoiding harms from AI systems as they meet or exceed human capabilities, positioning software engineering quality and practices as critical to responsible AI development during this transition period.

Interdisciplinary collaboration between SE and AI Safety communities is needed: By identifying how traditional software engineering work could improve AI safety outcomes, the paper implicitly calls for greater engagement between these communities to address long-term AI risks.