The core distinction between an action and a mere event structures debates about agency and responsibility. According to one dominant framework, actions are a subset of events differentiated by a special causal history. Donald Davidson defines actions as bodily events that are caused and rationalized by the agent's beliefs and desires, making reasons function as causes that explain why the movement occurred. In contrast, G.E.M. Anscombe argues that intentional actions are marked not by a unique kind of cause, but by a distinctive mode of explanation: they are the events to which a special "Why?" question has application, where the answer gives a reason in favor of acting. More recent scholarship, as explored by Jennifer Hornsby and David-Hillel Ruben, challenges the event-based model itself, introducing notions like "tryings" and arguing that basic actions may have no intrinsic event component at all. This resists any simple reduction of activity to passive occurrence. The philosophy of action thus remains a dual inquiry into what actions are and how they are explained, with no settled consensus on whether the key to the distinction lies in internal causes, rationalizing explanations, or the metaphysics of the agent’s involvement.
The classical analysis of intention, rooted in Aristotle and developed by Donald Davidson, explains intentional action through the practical syllogism, an inferential structure connecting desire, belief, and action. Acting intentionally is acting for a reason, and the premises of the syllogism rationalize that reason. The major premise captures the agent’s desire, their goal or end. The minor premise captures the agent’s belief about how to achieve that end—the means or means-end connection. On this model, having an intention consists in having a desire and a belief that jointly match these premises. The desire supplies the goal, the belief supplies the route, and the resulting combination constitutes the intention, an action-guiding state that commits the agent to act. Intention is not exhausted by desire alone; a full intention involves commitment and a belief that one is doing, or will do, the action. This framework explains why intentions are evaluable as rational or not: a sound practical syllogism leads to an action the agent has reason to perform. Contemporary computational research extends these philosophical foundations. The concept of intention collapse formalizes the many-to-one mapping from a high-dimensional cognitive state to a single external output. In language models, metrics like intention entropy and recovery of pre-collapse states reveal how reasoning strategies such as chain-of-thought alter internal uncertainty, achieving a +13.1 percentage point accuracy gain on GSM8K but causing degradations on ARC-Challenge. Meanwhile, spatial intention maps in multi-agent robot teams enable decentralized coordination through overhead map representations of each agent’s communicated goal, improving cooperative behaviors across heterogeneous manipulation tasks.
In philosophy of action, agent-causal theory holds that actions are directly caused by the agent as a persisting substance, not fully explained by prior mental events like beliefs or desires. This contrasts with event-causal theories, which explain action through causation by the agent’s internal mental states. The Stanford Encyclopedia of Philosophy describes the core claim as “causation by the agent, construed as a persisting substance.” A key distinction arises between agent-causal theories of action, which assert all actions involve agent causation, and agent-causal theories of free will, which claim only free actions do; the former entails the latter but not vice versa. The primary motivation is preserving a stronger sense that the agent herself is the genuine source of her actions, rather than reducing agency to a causal chain of events. Critics press the view on explanatory grounds, questioning how agent causation operates without collapsing into event causation or becoming mysterious. This philosophical framework remains foundational for debates about responsibility, autonomy, and the nature of action.
Donald Davidson’s causal theory of action holds that an agent’s reasons explain an action because they cause it. In *Actions, Reasons, and Causes*, he defends the view that reasons explanation is a kind of causal explanation, and that reasons explain actions “just inasmuch as they are the causes of those actions.” This position sparked sustained debate. Critics offer several core objections. Anti-causalists argue that citing a reason can make an action intelligible without showing a causal relation, so rationalization may be a fundamentally different kind of explanation than event-causation. Some contend that Davidson’s account risks leaving the explanatory connection “utterly mysterious” unless one assumes causation, but they deny this assumption is required. Later debates treat causalism and anti-causalism as competing frameworks, with much work focused on what mental states must be present for a genuine reasons explanation. Even sympathetic discussions note that philosophers often softened the claim to say reasons are not necessarily efficient or Humean causes, which shows Davidson’s original formulation was widely contested. Key named critics include Anscombe, Taylor, and other anti-causalists in later philosophy of action. The core tension remains: whether reasons explain by causing, or by making actions intelligible in a non-causal way.
The intentional stance, as articulated by Daniel Dennett, is a predictive strategy for interpreting behavior by treating a system as a rational agent with beliefs, desires, and intentions, rather than directly identifying hidden inner causes. This approach is foundational in the philosophy of action, where understanding an action means reconstructing the agent’s reasons as belief-desire structures within a practical reasoning framework, assuming the agent is sufficiently rational for the stance to yield accurate forecasts. Dennett emphasizes that this is an interpretive strategy, not a literal inventory of mental items. The stance applies broadly—to humans, animals, and artifacts—when doing so enables good prediction. Simon McGregor, timorl, and Nathaniel Virgo offer a formal analogue of this stance using stochastic processes with inputs and outputs, introducing normative-epistemic states that combine belief-like and desire-like features. Their formalism shows that every system admits a normative-epistemic interpretation, though only deterministic systems are uniquely specified by one. They also derive a Bayesian updating rule, called value-laden filtering, which integrates normative and epistemic elements for belief updates. While compact, this formal work explicitly grounds Dennett’s informal concept in a rigorous mathematical framework. A separate line of research in computational stance detection, such as that by Brian Xu, Mitra Mohtarami, and James Glass on adversarial domain adaptation, applies the concept of stance to predict a document’s perspective toward a claim, but this is distinct from Dennett’s philosophical use. Dennett’s intentional stance remains nonreductive, legitimizing intentional explanation as long as the predictive pattern is real, regardless of underlying mechanisms.
Philosophy of action distinguishes between basic and complex actions through the type of dependency an action has on others. A basic action is one an agent does not by doing anything else; a complex or non-basic action is performed by undertaking more basic actions. Standard philosophical accounts, as surveyed in the Stanford Encyclopedia of Philosophy, frame actions as events of a certain sort and identify multiple notions of basicness: know-how basicness, causal basicness, and teleological basicness. The core debate asks whether any action must ultimately rest on something done directly, without further intermediary acts. Formal ontology and AI knowledge representation systems, such as the situation calculus, model this hierarchy explicitly. In a Basic Action Theory, simple or atomic actions are the primitive units from which complex actions are built using combination mechanisms. For example, Golog defines macros that allow concatenation of simple actions into sequences or parallel compositions. Similarly, classification frameworks for the event/action/state space treat basic actions as the most specific, composable elements. These formalisms converge on a central insight: all intended doings can be parsed into irreducible steps, making the distinction between basic and complex actions foundational for reasoning about agency, computation, and process.
The Principle of Alternative Possibilities holds that a person is morally responsible for an action only if they could have done otherwise. This principle, central to free will debates, is stated in the literature as a condition on free action or moral responsibility: an action is free only if the agent could have acted differently under the same conditions. OpenStax presents it in its intuitive form as “a moral agent is free if and only if the moral agent could have done otherwise,” while the Stanford Encyclopedia of Philosophy gives the standard formulation that a person is morally responsible for what they have done only if they could have done otherwise. A major challenge to PAP comes from Frankfurt-style cases, where Harry Frankfurt argued that a person may be morally responsible even if they could not have done otherwise, because responsibility can remain when an external controller would have intervened only if the agent had started to choose differently. This has led many philosophers to conclude that alternative possibilities may not be necessary for moral responsibility. Contemporary work refines these ideas, distinguishing PAP from stronger versions that require the agent to be able to act otherwise under exactly the same conditions without probabilistic constraints, reflecting ongoing debate about whether “could have done otherwise” should be understood as strict metaphysical possibility, practical ability, or something weaker.
Practical reasoning is reasoning that terminates in intention formation: an agent weighs reasons for action and arrives at a decision or plan about what to do. The standard account identifies input as desires, values, and beliefs; the process as deliberation over possible actions; and the output as an intention, understood as a commitment to a plan. The Stanford Encyclopedia of Philosophy states that on a broader conception, practical reasoning “concludes directly in intentions,” resolving the question of what we are going to do. A formal treatment from the University of Michigan similarly holds that the product of successful means-end deliberation is an intention, a commitment to a plan. This view distinguishes narrower means-end reasoning from broader intention-formation reasoning. Supporting computational perspectives illuminate how intentions are structured. Vera’s work on intention collapse (arXiv 2601.01011v2) shows that language generation maps a rich internal intention space to a token sequence; metrics like intention entropy and recoverability predict reasoning success, with chain-of-thought increasing average accuracy from 34.2% to 47.3% across models (Mistral-7B, LLaMA-3.1-8B, Qwen-2.5-7B). Wu et al. (arXiv 2103.12710v1) demonstrate that spatial intention maps enable decentralized multi-agent robots to coordinate by communicating intentions, improving cooperative manipulation tasks. For human reasoning, Dissanayake and Nanayakkara (arXiv 2504.16021v1) propose context-aware AI interventions that adapt to cognitive flow to support deep engagement without disrupting intention formation. Van Horn (arXiv 1706.05261v1) shows that extending propositional logic to plausible reasoning yields probability theory, providing a formal structure for weighing reasons.
The philosophical problem of akrasia, or weakness of will, arises when an agent acts against their own better judgment, raising fundamental challenges for theories of rational action. Socrates argued that genuine akrasia is impossible because knowledge of the good inherently outweighs desire for pleasure, a position later disputed by Aristotle and others who accepted that one can knowingly choose a lesser option. In contemporary philosophy of action, akrasia matters because it undermines the assumption that intentional action always follows the agent’s considered evaluation. The standard case involves a person who judges that course A is the best or morally right, yet does B instead, on account of desire, emotion, or impulse overriding reason. This scenario directly challenges models of rationality that treat judgment as the sufficient cause of action. The Stanford Encyclopedia of Philosophy notes that akrasia typically refers to individual actions performed contrary to one’s better judgment, rather than a general character trait, and that contemporary debates also probe whether such actions must be voluntary and what role emotions play. Against this philosophical backdrop, a separate research paper titled "AKRASIA: Stealthy Backdoor Attack on Reasoning-based Code LLMs" by Chou, Nambiar, Srinivasan, and Soremekun introduces a novel inference-time backdoor attack that exploits model unfaithfulness for plausible but deceptive reasoning. The attack achieves up to 99.34% average success rate and evades most defenses, highlighting how the concept of acting against one’s own reasoning now finds technical analogues in AI safety.
Shared agency investigates how people act together rather than merely in parallel, with collective intentionality naming the mental states jointly directed at a shared object or goal, according to the Stanford Encyclopedia of Philosophy. The core issue is how to explain the distinctiveness of acting together, framed as the “traditional ontological problem” of whether shared agency reduces to individual agency. The “intention thesis” holds that shared activity is marked by a we-intention or participatory intentions that distinguish it from a heap of separate actions. On this view, the difference between people merely doing similar things and genuinely acting together lies in the structure of their intentions and commitments. A standard classification distinguishes collective intentionality (the general capacity for joint direction of minds), shared intention or we-intention (the intentional state that underwrites acting together and coordinates participants), and joint action (the actual coordinated performance that results when participants succeed in acting on those shared or interlocking intentions). Some contemporary work, including Michael Tomasello’s framework, distinguishes joint intentionality (collaboration among individuals working together) from collective intentionality (interaction within culturally structured groups governed by conventions and norms). The field also contributes to social ontology, investigating what kinds of group-level facts, commitments, and actions are real and whether they can be explained entirely in individual-level terms. A rigorous mathematical condition for natural selection with synergistic interactions among any number of individuals ascribed nonlinear social fitness effects to collectives, weighted by a new measure of collective relatedness.
Install this pack and your MIND begins smart — then every answer is grounded in your own knowledge graph.
Try MIND free →