Learning from Learning Machines: Optimisation, Rules, and Social Norms

Explicit rules may be less effective than implicit learning for complex behavior: The paper suggests that deep learning's success indicates that implicit specifications (where systems learn patterns rather than follow programmed rules) work better than explicit ones for achieving morally good decision-making in AI systems—a finding that challenges traditional approaches to AI safety that rely on hard-coded rules.

Economic systems offer lessons for AI safety: The authors propose that machine learning systems and economic entities are analogous as adaptive systems, suggesting that established economic principles and policies could inform how we design AI systems to behave responsibly, creating a potential bridge between two traditionally separate fields.

The challenge of defining moral behavior in machines remains unsolved: The paper identifies that while we can specify what we want AI systems to do in economic or technical terms, "it is an open question as to how precisely moral behaviour can be achieved in an AI system"—highlighting a fundamental gap in responsible AI development.

Bidirectional knowledge transfer between AI and economics: The authors argue that insights flow both ways—not only can economics inform AI safety solutions, but advances in AI and machine learning may also provide new insights applicable to economic policy and understanding human economic behavior.