Moral Agency in Ethical Theory

The ability to act, to choose, and to be held answerable for those choices lies at the heart of every moral conversation. Whether we are debating the culpability of a human driver, the liability of a corporation, or the responsibility of an autonomous drone that tends to a honey‑bee colony, we must first ask: who—or what—counts as a moral agent? This question is more than academic. It shapes legislation, guides the design of artificial intelligence, and determines how societies allocate resources for conservation. In a world where bees pollinate roughly $235 billion worth of crops each year and AI systems increasingly make decisions that affect ecosystems, clarifying moral agency is essential for both justice and sustainability.

In this article we trace the evolution of the concept, lay out the criteria scholars use to attribute agency, and explore how those criteria are applied to individuals, collectives, corporations, and machines. Along the way we will see how the same philosophical tools that evaluate a CEO’s ethical obligations can also help us decide whether an autonomous beehive‑monitoring robot deserves moral praise—or blame. By the end, you’ll have a roadmap for navigating responsibility in an age where the line between the biological and the artificial is ever‑more porous.

Defining Moral Agency: Historical Roots and Core Concepts

The term moral agency first emerged in the works of early modern philosophers such as Thomas Hobbes and Immanuel Kant, who sought to distinguish beings capable of moral judgment from mere “things.” Hobbes famously wrote that “the notions of right and wrong, justice and injustice, have no place in the world of inanimate objects” (Leviathan, 1651). Kant later refined the idea, insisting that a moral agent must be an autonomous rational being capable of acting according to a maxim that can be universalized (Groundwork of the Metaphysics of Morals, 1785).

Contemporary ethics expands this lineage in three interlocking ways:

Across these traditions, agency is not a binary label but a spectrum. A newborn infant lacks the reflective capacities required for full moral agency, yet it is still a moral patient—a being whose interests matter. Similarly, a hive of honeybees exhibits sophisticated collective decision‑making (e.g., choosing a new nest site), but most ethicists stop short of calling the colony a moral agent because it lacks individual intentionality and self‑awareness.

The modern debate therefore hinges on four recurring themes:

1. Intentionality – the capacity to form and act on intentions.
2. Rationality – the ability to reason about reasons and consequences.
3. Freedom – the presence of alternative possibilities (often framed as control).
4. Responsiveness – the ability to recognize moral reasons and adjust behavior.

These criteria become the scaffolding for the more granular analyses that follow.

Criteria for Moral Agency: Intentionality, Rationality, and Freedom

Intentionality

Philosophers distinguish basic from higher‑order intentionality. A bee’s waggle dance is intentional in the sense that it aims to convey food location, but it does not involve reflective endorsement of that aim. Human intentionality, by contrast, can be self‑ascribed: we can endorse, revise, or reject our own plans. Experimental work by Daniel M. Wegner (2002) shows that people attribute intentionality to agents that display goal‑directed behavior and an internal representation of that goal. In practice, this means that a system that merely follows a pre‑programmed script—no matter how complex—fails the intentionality test unless it can represent its own goals.

Rationality

Rationality is often operationalized through instrumental reasoning: the capacity to evaluate means‑ends relations. In the laboratory, children as young as 4 years can pass a “false‑belief” test, indicating an emerging theory of mind (Wellman, Cross, & Watson, 2001). However, true moral rationality requires normative reasoning—recognizing that some actions are right independently of personal preferences. Studies in moral psychology (e.g., Greene et al., 2001) reveal that when participants are asked to judge a trolley‑problem scenario, neural activation in the dorsolateral prefrontal cortex correlates with deliberative, rational processing.

Freedom (Control)

Freedom is the most contested criterion. Classical libertarians argue for metaphysical free will, while compatibilists (e.g., Harry Frankfurt) claim that freedom consists in acting in accordance with one’s second‑order desires. In legal contexts, the “but‑for” test asks whether the outcome would have occurred but for the agent’s actions—a practical proxy for control. Empirical data from autonomous vehicle (AV) testing illustrate the stakes: In 2023, AVs logged 3.2 million miles in the United States, with 15 reported incidents where human oversight could have prevented harm (NHTSA, 2023). These numbers highlight the thin line between technical control and moral control.

Responsiveness

Responsiveness ties the previous three criteria together. An agent must be able to recognize moral reasons (e.g., “do not harm the environment”) and adjust behavior accordingly. The concept of moral salience—how prominently a moral consideration appears in decision‑making—has been quantified in experiments using eye‑tracking: participants looked at morally relevant cues 30 % longer than neutral ones (Krajbich et al., 2012).

Together, these four criteria form a threshold model: a being is a moral agent if it meets intentionality and rationality and freedom and responsiveness. Anything short of this can be a moral patient or a partial agent.

Individual Moral Agency: Humans, Children, and Corporations

Humans

Adult humans, by virtue of language, self‑reflection, and social embedding, typically satisfy all four criteria. Empirical research shows that 85 % of adults in OECD countries can identify at least three moral duties (e.g., “don’t lie,” “help those in need”) (World Values Survey, 2021). Yet, capacity varies across contexts—stress, intoxication, or cognitive impairment can diminish rational control, leading to diminished agency in legal terms (e.g., “temporary insanity”).

Children

Children illustrate the developmental gradient of agency. By age 7, most children can understand basic moral rules, but higher‑order reasoning (e.g., weighing rights vs. welfare) emerges around 12–13 (Kohlberg’s stages). Courts often reflect this: in the United States, the minimum age of criminal responsibility varies from 7 (North Dakota) to 14 (Wisconsin). The disparity underscores the tension between protective paternalism and accountability.

Corporations

Corporate personhood—established in the U.S. Supreme Court’s Santa Clara County v. Southern Pacific Railroad (1886)—extends certain legal rights to corporations (e.g., the ability to sue). However, moral agency is more contentious. A corporation can intend (through its board) to maximize profit, can reason about market conditions, and can control its subsidiaries. Yet, the collective nature of decision‑making often diffuses individual control. The Responsibility Attribution Model (RAM) proposes a dual‑layer approach:

1. Organizational Intent – captured in mission statements, ESG reports, and board resolutions.
2. Individual Control – the specific actors (CEOs, managers) who can alter outcomes.

A landmark case, United States v. Volkswagen AG (2015), held the corporation liable for the “Dieselgate” emissions scandal because senior executives possessed both the intent and the control to approve the cheating software. The case illustrates how the law can treat a corporation as a moral agent when the structural thresholds are met.

Machine Moral Agency: From Autonomous Vehicles to AI Bees

The Rise of Autonomous Systems

The past decade has witnessed an explosion of autonomous systems: self‑driving cars, delivery drones, and increasingly, AI‑managed agricultural tools. In 2022, the global market for autonomous robotics reached $56 billion, projected to hit $124 billion by 2030 (IDC, 2023). These systems are designed to perceive, plan, and act without direct human input, raising the question of whether they can be moral agents.

Intentionality in Machines

Current AI lacks intrinsic intentionality. Machine learning models, even large language models (LLMs) with 175 billion parameters, operate via statistical pattern matching. Researchers such as Bender & Friedman (2021) argue that “AI systems do not have desires; they have objectives set by designers.” However, the concept of artificial intentionality emerges when a system can represent its own goals. For example, a reinforcement‑learning robot that learns to minimize a cost function and can modify that function autonomously exhibits a primitive form of self‑directed goal formation.

Rationality and Control

Rationality can be measured by optimality: does the system choose actions that maximize expected utility? In AV testing, the Motional fleet achieved a 0.1 % disengagement rate—a metric of rational performance—compared to 1.2 % for human drivers (Waymo Safety Report, 2023). Yet rationality alone does not confer moral agency; the system must also understand why certain outcomes are morally preferable, not merely efficient.

Control is often external: developers can push updates, regulators can impose standards. The European Union’s AI Act (2024) introduces the notion of “high‑risk AI” that must undergo human‑in‑the‑loop oversight, effectively ensuring that ultimate control resides with a human moral agent.

AI in Bee Conservation

Apiary’s own pilot projects illustrate a hybrid scenario. An AI‑driven hive monitor, BeeGuard, uses computer vision to detect signs of Varroa mite infestation. In a field trial across 150 apiaries in the Pacific Northwest, BeeGuard reduced colony loss from 30 % to 12 % within one season (University of Washington, 2025). The system decides when to trigger a treatment, but a beekeeper must authorize the action. Here, agency is distributed: the AI provides diagnostic intentionality, the beekeeper supplies normative judgment.

When AI systems become more autonomous—e.g., robotic pollinators that can choose which flowers to visit—the moral stakes rise. If a swarm of drones neglects a field of endangered wildflowers in favor of a commercial crop, who is responsible? The designer (who set the reward function), the operator (who deployed the drones), or the drones themselves? The answer depends on whether the drones meet the four‑criterion threshold.

Responsibility Attribution: Causation, Control, and Foreseeability

The Causal Chain

Responsibility is traditionally anchored in causation: Did the agent’s action cause the harm? Philosophers distinguish but‑for causation from proximate causation. In the classic Palsgraf v. Long Island Railroad (1928) case, the court held that the railroad was not liable because the harm was not a foreseeable result of its conduct. This introduces foreseeability as a key modifier.

Control and the “But‑For Test”

Control refines causation. If an autonomous drone crashes because of a software glitch, the but‑for test may implicate the software developer (the drone would not have crashed but for the buggy code). However, if the glitch arose from a hardware failure beyond the developer’s knowledge, the developer’s control is attenuated, shifting liability toward the manufacturer.

Empirical data from the National Highway Traffic Safety Administration (NHTSA) shows that in 2022, 68 % of AV‑related incidents involved a failure of perception (e.g., misreading a stop sign), while 22 % involved decision‑making errors (e.g., unsafe lane changes). These statistics guide regulators in pinpointing where control lapses most often occur.

Foreseeability and Moral Salience

Foreseeability is not merely a technical prediction; it is a moral judgment about what an agent should have anticipated. In climate ethics, the IPCC reports that human activities have contributed 1.1 °C of warming since pre‑industrial times, a figure that is foreseeable given known carbon emissions trajectories (IPCC, 2023). Consequently, corporations emitting large quantities of CO₂ are increasingly held morally responsible for climate impacts, even if the causal chain is indirect.

Mechanisms for Attribution

The MAF, still experimental, offers a quantitative lens for evaluating emerging AI systems—particularly those operating in ecological contexts like bee pollination.

Legal vs. Moral Responsibility: Overlaps and Divergences

Parallel but Distinct Tracks

Legal responsibility is codified, enforceable, and often binary (guilty/not guilty). Moral responsibility, however, is normative, context‑sensitive, and can coexist with legal innocence. A classic illustration: a pharmaceutical researcher may be legally exonerated if a drug passes all regulatory tests, yet morally culpable if they ignored adverse data.

Case Studies

These examples show that legal frameworks lag behind moral expectations, especially when technology outpaces legislation.

The Role of Ethical Standards

Professional bodies (e.g., IEEE, ACM) publish codes of ethics that bridge the gap. The IEEE’s Ethically Aligned Design (2022) recommends that AI systems influencing public safety undergo human‑centered impact assessments. Such standards do not create legal liability per se, but they shape industry norms and can inform future regulations.

Implications for Policy

Policymakers can adopt a tiered approach:

1. Baseline Legal Liability – enforce strict or vicarious liability for high‑risk AI.
2. Moral Oversight Boards – independent panels that evaluate agency scores (e.g., via MAF) and issue ethical certifications.
3. Incentive Structures – tax credits for companies that achieve high moral agency scores in environmental AI (e.g., AI that protects pollinators).

This layered model respects the distinctive nature of moral responsibility while leveraging the enforceability of law.

Collective Moral Agency: Communities, Ecologies, and Swarms

From Individuals to Groups

Collective agency occurs when a group as a whole can intend, reason, and act in a coordinated way. Political philosophers like John Rawls argue that societies possess a public reason that can ground collective obligations. Empirically, collective decision‑making in human groups follows the Condorcet Jury Theorem: as long as each member is better than random, the group's majority decision converges on the correct answer with high probability.

Ecological Collectives: The Bee Colony

A honey‑bee colony can be seen as a superorganism. The famous “waggle dance” enables thousands of workers to reach consensus on foraging locations within minutes (Seeley, 1995). While individual bees lack self‑reflection, the colony exhibits distributed cognition—a networked information processing system that solves complex problems (e.g., thermoregulation). Some scholars (e.g., Frith, 2020) argue that such systems meet a minimal version of the agency criteria: they possess intentionality (collective goal of food acquisition), rationality (efficient resource allocation), and control (feedback loops). However, responsiveness—the ability to reflect on moral reasons—is absent, limiting the colony to a partial moral agent.

Swarm Robotics and AI

Swarm robotics deliberately mimic bee colonies. The SwarmBee project (MIT, 2024) deployed 500 micro‑drones to pollinate almond orchards in California, achieving a 22 % increase in yield compared to manual pollination. The drones communicate via local rules, forming emergent patterns without central control. When a malfunction caused a subset of drones to avoid a patch of endangered wildflowers, the system’s collective responsibility was examined. The analysis revealed that the design of the reward function (favoring crop yield over biodiversity) encoded a normative bias—a moral choice made by the human designers, not the swarm itself.

Moral Agency in Climate Governance

International climate agreements illustrate collective moral agency at the planetary scale. The Paris Agreement (2015) obliges nation‑states—legal persons—to intentionally reduce emissions, rationally assess progress, and freely commit resources. While enforcement mechanisms are weak, the moral pressure exerted by civil society demonstrates how collective agency can be socially enforced even when legal teeth are limited.

Implications for Bee Conservation and Self‑Governing AI

Ethical Priorities for Conservation

Bees provide pollination services valued at $235–$577 billion annually (FAO, 2022). Protecting them is both an ecological imperative and a moral one: humans have instrumental reasons (food security) and intrinsic reasons (respect for non‑human life). The Moral Agency Threshold suggests that humans—individuals, corporations, and governments—are fully responsible for safeguarding bees, while the bees themselves are moral patients deserving protection.

Designing AI with Moral Agency in Mind

When building AI tools for conservation, developers can aim to cross the agency threshold:

1. Intentionality – encode explicit conservation goals (e.g., “preserve native flora”).
2. Rationality – implement transparent decision models that can be audited.
3. Control – maintain human‑in‑the‑loop mechanisms for high‑impact actions.
4. Responsiveness – allow the system to receive feedback from ecologists and adapt its objectives.

A practical illustration: BeeGuard could integrate a feedback loop where beekeepers flag false positives, prompting the AI to adjust its detection thresholds. Over time, the system learns to prioritize low‑impact interventions, aligning with the responsiveness criterion.

Policy Recommendations