Breaking Up with Normatively Monolithic Agency with GRACE: A Reason-Based Neuro-Symbolic Architecture for Safe and Ethical AI Alignment

As AI agents become increasingly autonomous, widely deployed in consequential contexts, and efficacious in bringing about real-world impacts, ensuring that their decisions are not only instrumentally effective but also normatively aligned has become critical. We introduce a neuro-symbolic reason-based containment architecture, Governor for Reason-Aligned ContainmEnt (GRACE), that decouples normative reasoning from instrumental decision-making and can contain AI agents of virtually any design. GRACE restructures decision-making into three modules: a Moral Module (MM) that determines permissible macro actions via deontic logic-based reasoning; a Decision-Making Module (DMM) that encapsulates the target agent while selecting instrumentally optimal primitive actions in accordance with derived macro actions; and a Guard that monitors and enforces moral compliance. The MM uses a reason-based formalism providing a semantic foundation for deontic logic, enabling interpretability, contestability, and justifiability. Its symbolic representation enriches the DMM’s informational context and supports formal verification and statistical guarantees of alignment enforced by the Guard. We demonstrate GRACE on an example of a LLM therapy assistant, showing how it enables stakeholders to understand, contest, and refine agent behavior.

💡 Research Summary

The paper addresses the growing need for AI agents to act not only instrumentally effective but also normatively aligned as they become more autonomous and are deployed in high‑stakes contexts. Existing approaches such as Reinforcement Learning from Human Feedback (RLHF), Constitutional AI, or pure rule‑based systems either embed ethical constraints inside opaque neural networks or rely on fixed rule sets that lack adaptability and transparency. This “flattening problem” conflates instrumental decision‑making, normative reasoning, and policy execution into a single policy function π, making verification, modular testing, and contestability difficult.

To solve this, the authors propose GRACE (Governor for Reason‑Aligned ContainmEnt), a neuro‑symbolic architecture that explicitly separates normative reasoning from instrumental optimization. GRACE consists of three sequential modules:

1. Moral Module (MM) – Uses a reason‑based formalism grounded in informal deontic logic to generate macro‑action types (MATs). MATs are decidable predicates over temporally extended macro‑actions (e.g., “assess self‑harm risk”, “preserve confidentiality”). The MM also incorporates a Moral Advisor that supplies case‑based normative feedback, allowing the system to incrementally build and revise a “reason theory”. This provides interpretability, contestability, and justifiability of ethical judgments.

2. Decision‑Making Module (DMM) – Remains agnostic to the underlying learning architecture (reinforcement learning, large language models, etc.). It receives the set of permissible MATs from the MM and selects instrumentally optimal primitive actions that satisfy the macro‑action constraints. Thus, the DMM can exploit state‑of‑the‑art neural optimisation while being confined by explicit normative boundaries.

3. Guard – Monitors the actions selected by the DMM in real time, enforcing compliance with the MM’s specifications. It combines formal verification techniques with statistical safety guarantees; any action that would instantiate a prohibited macro‑action is blocked or replaced before execution.

A central conceptual contribution is the three‑level action abstraction: primitive actions → macro actions → macro‑action types. This hierarchy mirrors how humans discuss norms (“what ought to be done”) separately from implementation details (“how to do it”), allowing the system to reason about ethics at a high‑level semantic layer rather than at the raw token or actuator level.

The authors demonstrate GRACE on a large‑language‑model‑based therapeutic assistant named THERAPAI. In this scenario, the agent must balance therapeutic effectiveness with obligations such as patient confidentiality, non‑maleficence, and cultural sensitivity. By feeding the MM with relevant MATs (e.g., “do not disclose private information unless mandatory reporting applies”), the system automatically blocks any primitive action that would violate confidentiality, while still selecting the most effective therapeutic responses within the allowed space. When a self‑harm risk is detected, the MM elevates the “self‑harm assessment” MAT, prompting the DMM to prioritize safety‑oriented actions. Stakeholders (clinicians, ethicists, patients) can inspect the logical derivations produced by the MM, contest them, and supply new case feedback, which the Moral Advisor incorporates into the evolving reason theory.

Empirical observations show that GRACE improves interpretability (the MM’s logical proofs are human‑readable), contestability (stakeholders can request revisions to MATs), and quantitative safety (the Guard’s formal checks provide provable bounds on the probability of norm violations). Compared with baseline black‑box approaches, GRACE maintains comparable therapeutic performance while offering stronger guarantees against ethical breaches.

The paper acknowledges several limitations: constructing the deontic logic and MATs requires expert effort; real‑time symbolic reasoning can be computationally intensive; and the interface between human moral advisors and the system needs careful design to avoid bias. Future work is outlined to automate the extraction of reasons from large corpora, scale case‑based learning, and extend the framework to multi‑cultural, multilingual environments where normative standards diverge.

In summary, GRACE proposes a principled, modular, neuro‑symbolic architecture that decouples normative reasoning from instrumental optimization, enabling transparent, contestable, and verifiable AI behavior in ethically charged domains.