Adaptive moral compass framework for ethical reasoning and self-oversight in autonomous AI agents

Abstract

The growing use of autonomous artificial intelligence agents in high-risk areas such as -driving cars, medical triage tools, military robots, and disaster-response drones, has made AI ethics an important concern for researchers. Current alignment methods, especially Reinforcement Learning from Human Feedback, have shown good results but have several known weaknesses, including reward hacking and poor generalization to new situations. This study proposes the Adaptive Moral Compass framework, a new architecture that places ethical reasoning directly inside the decision-making process of an autonomous agent. The framework is built on four main concepts: moral pluralism, case-based reasoning, self-monitoring, and balancing multiple objectives. The agent evaluates each candidate action using three evaluators that score consequences, duties, and moral character, and combines their scores with a weighted aggregation rule. An oversight mechanism reviews the decision before it is carried out, and a learning module reviews the outcome afterwards so that the agent can improve over time. The framework was evaluated through an agent-based disaster response simulation written in Python and run in Google Colaboratory. The environment was a 50-by-50 grid containing eight survivors and two relief camps, and each run lasted 80-time steps. An autonomous rescue agent faced triage decisions, conflicts between duty and compassion, and resource allocation problems. Three agents were compared: the AMC agent, a utilitarian agent, and a rule-based agent. In the scenario reported here the AMC agent recorded the lowest cumulative suffering index at 52 units, compared with 74 for the rule-based agent and 101 for the utilitarian agent, but it saved three survivors while each baseline agent saved seven. Across thirty perturbed replications this ordering held, with mean suffering of 48.80, 72.77 and 95.60 units and non-overlapping confidence intervals. An ablation study showed that this behavior is attributable to the multi-perspective scoring function rather than to the oversight or reflective learning components, whose trigger conditions were not met. The results are therefore best understood as a multi-objective ethical trade-off between preserving life and reducing the burden borne by those awaiting relief, rather than as evidence that one agent is ethically superior. The study provides proof of concept that several ethical perspectives can be combined and measured computationally and indicates that a single optimization objective renders such trade-offs invisible.