Designing Transparency: A Conceptual Blueprint for Legal XAI

TL;DR

As AI moves from simple document filtering to complex legal reasoning, the "Black Box" nature of neural networks becomes a liability. This paper argues that explainability in law shouldn't be a post-hoc analysis but a core design constraint. By introducing the KORA design methodology and the IntERMeDIUM research strategy, the authors provide a bridge between abstract legal ethics and executable code.

The "Black Box" Liability in Law

The legal domain is fundamentally built on reasoning and justification. While Deep Learning excels at detecting unfair clauses in contracts, it fails at the "Why." Modern AI researchers are now moving toward Computational Models of Legal Reasoning (CMLR), where the goal isn't just a prediction, but a transparent path to a decision. The problem is that engineers and lawyers speak different languages—one speaks in "parameters," the other in "provisions."

Methodology: The Two Routes to Explainability

1. KORA: From Legal Code to Source Code

The KORA method is a design pipeline that ensures legal requirements are "compiled" into technical functions.

• Phase 1-2: Take a legal provision (e.g., Transparency) and derive specific "Criteria."
• Phase 3-4: Map these criteria to Functional Requirements and finally to Implementation Choices.

This ensures that when an AI model is chosen (e.g., a decision tree versus a neural network), the choice is justified by a specific legal requirement identified at the start.

2. IntERMeDIUM: The 8 Pillars of Ethical AI

To build a system capable of "Moral Reasoning," the authors propose the IntERMeDIUM strategy. The most critical dimensions include:

• Intentional: Explanations should be in terms of Goals and Beliefs, matching how humans interpret rational behavior.
• Declarative: Knowledge must be inspectable. Every ethical bias must be written in a format humans can read and verify.
• Reward-Based (Inductive): Using Reinforcement Learning (RL) to fine-tune behavior. Legal norms are treated as rewards or constraints within the RL framework.

Deep Insight: Why "Intentionality" Matters

The true innovation here is the shift toward the Intentional Stance. Most XAI focuses on "Feature Attribution" (e.g., "this word triggered the classification"). In a legal context, this is insufficient. A lawyer needs to know the "Goal" of the agent. By using Declarative Decision-Theoretic Ethical Programs (DDTEPs), the AI can explain its actions based on its programmed "beliefs" about the world and its "desire" to follow specific legal norms.

Critical Analysis & Future Outlook

While the paper provides a robust conceptual framework, the implementation of "Common Sense" in these declarative systems remains a significant hurdle.

Key Takeaways for the Future:

1. Legal Design Patterns: There is a desperate need for a library of "Explanation Patterns" that can be reused across different legal applications.
2. Human-Centricity: Systems must move toward interactive dialogue. An explanation is not a static text; it is an exchange where the human can ask "What if?"
3. Verification: Using formal methods (Executability) to prove that an AI will never violate a specific legal constraint is the next frontier for high-stakes legal deployment.

In conclusion, the paper moves the conversation from how AI works to how it should be built to serve justice.