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

Travis LaCroix, Yoshua BengioarXiv 2020 · arXiv

Introduction

In both artificial intelligence and economics, we are increasingly confronted with the challenge of managing powerful, adaptive agents. Whether these agents are deep learning models or multinational corporations, they share a fundamental characteristic: they are optimisers designed to achieve specific objectives under a set of constraints. Historically, society has attempted to control these entities through explicit, rule-based systems—hard-coded logic for AI and rigid legal frameworks for corporations.

This paper proposes a novel theoretical framework by drawing an analogy between the evolution of machine learning and the future of economic regulation. Just as the field of AI shifted from brittle, rule-based systems to flexible, implicit deep learning models to solve complex problems, the authors argue that society must move toward implicit principles and social norms to ensure that corporate behavior aligns with the common good.

The Problem of Explicit Constraints

When an agent’s primary goal is to maximise a reward (such as profit or a specific loss function) within a set of rigid constraints, mathematical optimisation dictates a specific outcome: the agent will naturally find solutions that exist on the very boundary of those constraints.

Drawing on Optimisation Theory, the authors reference linear programming theorems to demonstrate that optimal solutions inherently reside on the vertices of feasible sets. In a real-world context, this manifests as "reward hacking" in AI or the exploitation of legal loopholes by corporations. If a corporation can increase its profit by navigating around the literal text of a regulation, it proves that its optimal behavior was already riding the absolute edge of the law. This "Boundary Effect" suggests that explicit rules, no matter how detailed, will always invite agents to seek out and exploit gaps between the letter and the spirit of the law.

Morality as an Implicit System

The failure of classical, rule-based AI to achieve human-level performance in tasks like image recognition or natural language translation provides a vital lesson. These tasks are "System 1" activities—intuitive, context-dependent, and highly complex. Human morality and social norms function similarly; they are not easily reducible to a manageable set of "if-then" rules.

Deep Learning Theory offers a solution to this complexity. Deep neural networks represent piecewise-linear functions with an exponentially growing number of regions, allowing them to learn implicit representations of complex data. The authors argue that morality is an implicit representation that humans navigate intuitively. Because this knowledge cannot be compressed into a small set of explicit rules without losing its essence, any attempt to govern agents solely through explicit laws will inevitably fail to capture the nuance of moral behavior.

The Backward Analogy: Learning from AI

While AI researchers have long looked to economics for solutions to the "alignment problem" (e.g., the principal-agent problem), this paper suggests a "backward analogy." Economics and legal theory should look to the success of deep learning. To ensure corporations act morally, we require regulatory environments that exert implicit pressures—norms, principles, and human judgment—rather than just mechanical, explicit rules.

Implicit rules require human intervention to evaluate. While this introduces a degree of subjectivity, it makes the system significantly harder to "game." Unlike explicit rules, which can be checked mechanically, implicit principles require the agent to align with the intent of the system.

Comparative Analysis: Implicit vs. Explicit Frameworks

The paper highlights three qualitative results where implicit rules have been successfully applied in economic and legal frameworks, often contrasting the approaches of the United States and Canada:

  1. Spirit vs. Letter of the Law: While US law often prioritises textualism (the explicit letter), Canadian law (e.g., the Interpretation Act of 1985) requires that laws be interpreted according to their "true spirit, intent, and meaning."
  2. Stakeholder vs. Shareholder Theory: US case law, such as eBay v. Newmark, often enforces strict, explicit profit maximisation for shareholders. In contrast, Canada’s Bill C-97 allows corporate directors to consider the implicit, complex interests of the environment, employees, and consumers.
  3. Accounting Principles: The US generally relies on rule-based Generally Accepted Accounting Principles (GAAP), which are frequently gamed. Canada utilises principle-based GAAP, which requires human judgment to ensure the substance of a transaction is recorded, rather than just its technical form.

Limitations and Future Directions

The authors acknowledge that a shift toward implicit rules presents practical challenges. Relying on human judgment makes it more difficult to quantify value and verify compliance mechanically. Furthermore, human-driven systems can be slower and are potentially vulnerable to bias or corruption if not supported by high levels of transparency.

Future work should focus on practical policy shifts that empower civil servants to exercise judgment based on high-level principles. The authors propose modifying corporate charters to mandate the maximisation of the "common good." This would require a holistic auditing of a company’s societal impact—such as its effect on the environment and employee well-being—moving beyond simple financial metrics and legal compliance toward true alignment with human values.

Summary updated

Abstract

There is an analogy between machine learning systems and economic entities in that they are both adaptive, and their behaviour is specified in a more-or-less explicit way. It appears that the area of AI that is most analogous to the behaviour of economic entities is that of morally good decision-making, but it is an open question as to how precisely moral behaviour can be achieved in an AI system. This paper explores the analogy between these two complex systems, and we suggest that a clearer understanding of this apparent analogy may help us forward in both the socio-economic domain and the AI domain: known results in economics may help inform feasible solutions in AI safety, but also known results in AI may inform economic policy. If this claim is correct, then the recent successes of deep learning for AI suggest that more implicit specifications work better than explicit ones for solving such problems.
BibTeX
@misc{lacroix2020learning,
  author = {Travis LaCroix and Yoshua Bengio},
  title = {{Learning from Learning Machines: Optimisation, Rules, and Social Norms}},
  year = {2020},
  eprint = {2001.00006},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2001.00006},
}
Figure