Legal, Ethical, Customizable AI: Beyond Hard-Coded Morality

TL;DR

In the pursuit of truly safe and accountable AI, Joseph A. Blass (Northwestern University) argues that we must move away from rigid, hard-coded rules. By leveraging Analogical Reasoning and Structure Mapping, this research develops systems that learn ethical and legal principles from cases and precedents—much like a human judge or an ethical individual—allowing AI to be both legally compliant and culturally customizable.

The Motivation: Why Rules Fail

Most current AI safety frameworks attempt to "bottleneck" behavior through strict guardrails or reward functions. However, the author identifies two fatal flaws in this approach:

1. The Common Law Problem: Legal systems (especially in the US) rely on judicial interpretation and precedent. These "rules" are often unwritten or evolve through specific scenarios that defy simple Boolean logic.
2. The Ethical Pluralism Problem: Ethics aren't universal. What is considered a "polite white lie" in one culture might be seen as "impermissible deception" in another. AI should not impose its creator's ethics on the user; it should be customizable.

Methodology: The Power of Analogy

The core of this research rests on three sophisticated components: MoralDM, SME, and SAGE.

1. Structure Mapping Engine (SME)

Unlike traditional machine learning that looks at statistical correlations, SME looks at structural alignments. It cares about the relationships between actors (e.g., "Who did what to whom and why?") rather than just the keywords.

2. Sequential Analogical Generalization Engine (SAGE)

SAGE takes a library of examples and performs "analogical abstraction." If it sees ten cases where "causing unnecessary pain" leads to a "bad" outcome, it builds a generalization that emphasizes the core relationship (Pain $\to$ Bad) while ignoring irrelevant details (e.g., the name of the person or the color of the room).

3. Analogical Chaining

A major breakthrough here is the ability to fill in "commonsense gaps." Human stories assume the reader knows that "pain is bad." Analogical Chaining uses micro-stories to teach the AI these basic truths, which it then chains together to understand complex legal scenarios.

(Note: Representation of the structural mapping and reasoning framework discussed in the paper)

Experiments & Results: Human-Consistent Judgments

The author’s research demonstrates that:

• Generalization > Exemplars: Systems that generalize principles from multiple cases (via SAGE) consistently make more human-like moral decisions than those that try to match a new situation against a single previous example.
• Defeasible Logic: By using logic that allows for "defaults and exceptions," the AI can handle situations where two rules conflict—a necessity for legal reasoning where "self-defense" might override the "prohibition against violence."

(Note: Quantitative results often illustrate the alignment between SAGE-based generalizations and human normative responses)

Critical Analysis & The Road Ahead

While the logic is sound, the author admits a significant hurdle: Natural Language Understanding (NLU). Legal language is notoriously dense. Currently, these systems often require "experts" to translate English law into CycL (a logic language).

The next frontier involves:

• Automating Rule Synthesis: Converting generalized structures directly into Horn-clause rules.
• Attention Mechanisms: Teaching the AI which facts in a legal case are "material" (important) and which are "immaterial" (noise).

Conclusion

This work is a vital pivot from "AI that follows instructions" to "AI that understands principles." By treating legal and ethical reasoning as an analogical task rather than a classification task, we pave the way for AI that can truly navigate the nuances of human society.