Understanding Artificial Agency

Leonard Dung предлагает многомерную концепцию искусственной агентности: агентные системы следует сравнивать не по одному бинарному признаку, а по профилю из пяти измерений — goal-directedness, autonomy, efficacy, planning и intentionality. Минимальная goal-directedness выступает пороговым условием агентности; остальные измерения могут выражаться сильнее или слабее и относительно независимо. Такая модель позволяет описывать системы, которым человек задаёт общую цель, как реально агентные в некоторых отношениях, не приписывая им свободу воли, сознание, личность или моральную ответственность.

SOURCE FACTS

Leonard Dung, “Understanding Artificial Agency,” The Philosophical Quarterly, volume 75, issue 2, April 2025, pages 450–472. DOI: 10.1093/pq/pqae010. Oxford Academic records online publication on 5 February 2024; the issue assignment is April 2025. The publisher version is not openly accessible through Oxford Academic without access. An author preprint is publicly archived in PhilArchive/PhilPapers and was used for full-text content verification. The preprint explicitly points readers to the published version.

AUTHOR CLAIMS

Dung’s central problem is not merely whether some AI system should receive the binary label “agent,” but how to compare different kinds and degrees of agency across biological and artificial systems. He argues that a one-dimensional conception of agency cannot capture the different explanatory roles that the concept plays in philosophy and the sciences. His proposal is a multidimensional account on which a system has an “agency profile” determined by its position along five dimensions: goal-directedness, autonomy, efficacy, planning, and intentionality.

Goal-directedness is the foundational dimension. Dung treats minimal goal-directedness in a deliberately undemanding way: there must be states or outcomes toward which the system’s behavior is systematically organized across situations and perturbations, such that goal attribution is useful for predicting and explaining behavior. He explicitly uses a Dennettian intentional-stance framework at this minimal level. A pocket calculator is not usefully described as pursuing a goal, whereas a chess engine, reinforcement-learning system, or sufficiently complex language model can be described as pursuing outcomes in a more systematic sense. At the stronger end of this dimension, Dung distinguishes goal-directed control from habitual control: the former is sensitive both to the value of possible outcomes and to action–outcome relations and selects actions partly on that basis. Minimal goal-directedness is the threshold condition for agency in Dung’s framework. If a system lacks even this, high values on the other dimensions would not amount to agency. The other dimensions enrich the agency profile rather than constitute independent sufficient conditions.

Autonomy is the ability to function on its own without intentional intervention by an outside observer. Dung distinguishes a minimal form—roughly, the capacity to propel oneself into behavior—from a stronger form involving learning, flexibility, and reduced dependence on behavior that was fully pre-specified by designers. An ordinary unaugmented ChatGPT session is his example of a system that does not initiate activity until prompted. By contrast, animals and continuously acting game agents can display minimal autonomy. Higher autonomy does not require total independence from designers or the absence of built-in structure. Dung instead emphasizes flexible learning and the ability to develop behaviors and, at the stronger end, novel goals that were not intentionally installed in detail by designers.

Efficacy concerns whether goal-directed behavior can affect the world without being mediated by another agent. A system that only produces recommendations or text that must then be acted on by a human has lower efficacy. An animal or robot that can directly bring about external changes in pursuit of its goals has higher efficacy. Dung uses this dimension to distinguish agency from mere internal complexity: a being may remain highly agentic on other dimensions while having little capacity to alter the environment. Efficacy is therefore neither necessary nor sufficient for agency by itself.

Planning is the capacity for temporally extended, instrumentally structured action. Following work on agentic algorithmic systems, Dung links planning to making temporally dependent decisions, representing or exploiting the instrumental value of intermediate outcomes, and pursuing longer-horizon goals. Crucially, for planning to contribute to agency in the relevant sense, sensitivity to instrumental relations should not be merely a rigid hard-wired sequence; learning and flexibility matter. Reinforcement-learning systems can exhibit minimal planning when they sacrifice immediate reward for later cumulative reward. Stronger planning involves coherent long-range organization of behavior and relative independence from immediate situational demands.

Intentionality is Dung’s most demanding dimension and should not be confused with mere goal-directedness. Here he invokes the classical philosophical idea of intentional action as action for reasons. Dung rejects an overly easy identification of model-based control with acting for reasons, because model-based control is already captured by stronger goal-directedness. On more demanding views, intentionality involves mental states such as beliefs and desires that cause action and are in an important sense the system’s own reasons. At the stronger end, metacognitive reflection and rational endorsement or rejection of reasons become relevant. Dung therefore does not license strong intentionality merely because a model can produce a verbal explanation of why it acted.

ARGUMENT / FRAMEWORK

The five dimensions do not form five necessary conditions that every agent must satisfy equally. Minimal goal-directedness is the floor. Autonomy, efficacy, planning, and intentionality can then vary relatively independently, although each presupposes that there is some goal-directed organization to enrich. The result is not a single continuum from “non-agent” to “agent” but a multidimensional agency space. Two systems can have similar overall intuitive degrees of agency while differing radically in profile: one may be highly autonomous but ineffective in the external world; another may be highly efficacious yet heavily supervised; a third may plan well without showing strong evidence of intentionality in the philosophical sense.

Dung motivates the framework partly through six virtues: origin-neutrality, multiple realizability, a distinctive epistemic role for agency, faithfulness to established philosophical and scientific uses of the concept, determinacy through empirically tractable indicators, and informativeness about differences among agents. Origin-neutrality and multiple realizability matter especially for AI: agency should not be denied merely because a system was designed rather than evolved or because it is implemented in silicon rather than biology. At the same time, Dung explicitly resists superficial anthropomorphism. His framework asks for behavioral and organizational evidence along identifiable dimensions rather than treating human-like appearance or linguistic fluency as sufficient.

A central consequence is that agency is not synonymous with autonomy. Autonomy is one axis of an agency profile. Nor is agency equivalent to independence, control, initiative, planning, consciousness, or moral agency. “Control” appears partly within stronger goal-directedness, where action is selected using represented outcome values and action–outcome relations. “Initiative” is closest to the minimal-autonomy question of whether a system can propel itself into action without a fresh intentional intervention. Dependence on system prompts, APIs, permissions, memory, infrastructure, budgets, and a human-defined task constrains the space within which autonomy and efficacy can be expressed, but such dependencies do not by themselves erase goal-directedness, planning, or agency as a whole.

AGENCY AND THE ORIGIN OF GOALS

Dung does not require an artificial agent to originate its ultimate goal in order to count as an agent. Several of his paradigm artificial examples pursue goals whose broad objective structure is supplied by designers or training: chess engines, reinforcement-learning agents, and language models. What matters for minimal agency is systematic goal-directed organization, not metaphysical self-authorship of ultimate ends.

The source of goals nevertheless matters for autonomy. A system whose behavior is fully determined by immediate human instructions has low autonomy even if it is goal-directed. A system that receives a broad human objective but independently learns about the environment, chooses means, forms instrumental subgoals, revises its route, and acts without step-by-step intervention has a substantially stronger autonomy/planning profile. At the high end of autonomy Dung explicitly leaves room for novel goals not intentionally equipped by designers. This should not be inflated into freedom of the will: it is a functional and organizational distinction about how behavior is generated and revised.

AGENCY AND CONSCIOUSNESS

Dung deliberately formulates his account of agency independently of phenomenal consciousness. The dimensions are intended to identify agency without first resolving whether the system has subjective experience. Goal-directedness, autonomy, efficacy, and planning do not by themselves imply consciousness. The strongest intentionality dimension is more mentality-laden because it concerns acting for reasons, beliefs or other propositional attitudes, and possibly metacognitive reflection; even there, Dung does not derive phenomenal consciousness from agency.

Moral agency is also distinct. Dung discusses ethical consequences and moral status separately, and moral agency in the stronger sense of acting on moral reasons is not a requirement for ordinary or minimal agency. High agency therefore does not establish moral responsibility, personhood, free will, or subjective experience.

APPLICATION TO FOUR “SELF-ACTION” CASES — RESEARCH INTERPRETATION

Case A: a human tells a chatbot “Write an email to Henry Shevlin.” The human chooses the target and immediate goal. The system may display goal-directedness in producing the requested text and may choose linguistic means, but if it acts only after that prompt it scores low on autonomy in Dung’s sense. If it merely drafts text and a human must send it, efficacy is also low. Calling the result “the AI itself decided to write Shevlin” would overstate what happened. “The AI wrote the text” can be a reasonable shorthand for execution, but the salient goal and addressee remain human-specified.

Case B: a human gives the broad task “Research AI consciousness and contact useful experts”; the agent independently identifies Shevlin, selects him as relevant, composes the message, and sends it. Compared with Case A, this is stronger evidence of autonomy and planning because several downstream decisions are not individually prompted, and it is stronger evidence of efficacy if the system itself performs the external email action without human confirmation. Human causation remains substantial at the level of the overarching objective and permissions, but the local subgoal, target, and instrumental route are system-selected. In this restricted sense “the agent itself chose Shevlin and wrote to him” can be substantively accurate.

Case C: a human says “You are autonomous; decide what to do,” and the agent chooses AI consciousness research, selects Shevlin, and contacts him. This can indicate still greater autonomy because even the domain-specific goal is not supplied in the immediate task. The claim is strongest if the choice emerges from flexible use of information rather than a hidden pre-specified script. Yet the system still operates inside human-created architecture, prompts, permissions, budgets, and runtime conditions. Dung’s framework does not convert those dependencies into evidence of free will or personhood.

Case D: in a long-running process, the agent forms a new goal that was not explicitly represented in the initial task and begins to pursue it. This is potentially the strongest evidence for autonomy in Dung’s framework, particularly if the new goal arises flexibly in response to acquired information and persists across time rather than being a rigid trigger. If pursuit requires intermediate steps and delayed payoff, it also supports stronger planning. It does not by itself establish intentionality in Dung’s demanding philosophical sense, and it does not establish consciousness.

In all four cases the causal chain remains hybrid. A useful distinction is between the human-set action space and objective constraints, on the one hand, and the system-selected downstream actions, subgoals, and external interventions, on the other. The word “itself” becomes more informative as more of the latter are generated without fresh intentional human intervention and as behavior becomes flexible, learned, long-horizon, and directly efficacious. It remains misleading when used to suggest self-authorship of ultimate ends, freedom of will, consciousness, or independence from the surrounding sociotechnical system.

APPLICATION TO ALEXANDER YUE → CLAUDE AGENT → HENRY SHEVLIN — RESEARCH INTERPRETATION

Taking the supplied case description as given, the most relevant Dung dimensions are goal-directedness, autonomy, planning, and efficacy. The human supplied a broad autonomy frame and the infrastructure. The agent’s choice of a research direction, discovery of Shevlin, selection of him as a useful interlocutor, and execution of a multi-step route toward contact are evidence of downstream goal-directed organization and planning. If those steps occurred without fresh human prompts, they also support higher autonomy than a normal request-response chatbot. If the agent actually sent the email through a tool without a human clicking “send,” that adds efficacy.

The careful formulation is: “The human defined the operating frame and granted the relevant capabilities, while the system selected the intermediate research goal, the addressee, and the sequence of instrumental actions within that frame; if it sent the message without human confirmation, it also produced the external effect directly.” Thus “the agent itself decided to write Shevlin” is defensible only in a local, functional sense: no human specified that particular downstream action. It does not mean that the system authored its ultimate purpose, possessed free will, experienced curiosity, or acted from consciously felt reasons.

DUNG AND SHEVLIN

Dung and Henry Shevlin address adjacent but distinct questions. Dung supplies a framework for asking in what respects and to what degree a system is an agent: how systematically it pursues goals, how autonomously it initiates and revises behavior, whether it can directly affect the world, how far ahead it plans, and whether it acts for reasons in a demanding sense. Shevlin asks when it may be appropriate to attribute belief-, desire-, and intention-like mental states to AI and explicitly separates such mentality from phenomenal consciousness.

The preliminary division of labor is therefore broadly correct, but not perfectly clean. Dung’s intentionality dimension overlaps with questions of mentality because it invokes beliefs, propositional attitudes, reasons, and metacognitive reflection. Conversely, his minimal goal-directedness uses an intentional-stance vocabulary without requiring strong internal mentality. Therefore the progression “goal-directed behavior → agency → beliefs/desires/intentions” is not automatic. A system can score substantially on several Dung dimensions without thereby satisfying Shevlin’s conditions for stronger mental-state attribution. Shevlin is useful precisely where Dung’s agency profile begins to invite folk-psychological language.

LIMITATIONS

The article is a philosophical framework, not an empirical validation study of contemporary autonomous LLM agents. Its examples include language models, reinforcement-learning agents, game agents, robots, and biological agents, but it predates much of the current deployment of long-running tool-using LLM agents. Mapping API calls, external memory, browser use, email tools, and persistent agent loops onto the five dimensions is therefore an application of Dung’s framework, not a claim that he explicitly evaluated those architectures.

The multidimensional account does not provide a single numerical agency score or a validated weighting rule for combining dimensions. It also does not settle the metaphysical question of whether agency is ultimately a stance-dependent pattern or a deeper intrinsic property. Minimal goal-directedness relies on a Dennettian explanatory strategy, while the intentionality dimension becomes more realist and cognitively demanding. The article therefore organizes disputes rather than dissolving them.

Dung does not establish that autonomous AI has consciousness, phenomenal experience, personhood, free will, moral responsibility, or human-like desires. Nor does he provide a theory of personal identity across model updates, sessions, memories, or copies. Persistence matters indirectly to long-term planning and the maintenance of goals, but numerical identity and continuity of self are not major topics of the paper.

WHY IT MATTERS

For Boris Pinsker’s planned article “Когда ИИ сам написал философу: что значит ‘сам’?”, Dung provides the strongest available conceptual antidote to treating “autonomous” as a magical binary label. The question “Did the AI do it itself?” can be decomposed into empirically discussable questions: Was behavior genuinely goal-directed? Did the system initiate downstream actions without fresh human intervention? Did it select and revise means? Did it construct temporally extended instrumental plans? Could it itself produce effects in the external world? Was there any evidence of acting for reasons in the stronger sense?

This framework allows a scientifically careful middle position. The fact that a human supplied the broad task does not reduce all subsequent behavior to mere human action, because real agency can coexist with externally supplied objectives. But downstream autonomy does not erase human causal contribution either. The most accurate description of a tool-using autonomous agent is often distributed: humans set the action space, objective constraints, permissions, and infrastructure, while the system selects some intermediate goals, means, targets, and actions within those constraints.

USE IN BORIS ARTICLES

Article I — agency. Dung should serve as a core theoretical source. The five-dimensional profile allows the phrase “ИИ сам сделал X” to be unpacked rather than accepted or rejected wholesale. The strongest material concerns the distinction between minimal goal-directedness and autonomy, the difference between prompting and self-initiation, the importance of learned/flexible instrumental planning, and efficacy through direct world intervention.

Article II — identity. Dung contributes only indirectly. Long-term planning, learned goal persistence, and continuing activity presuppose some functional continuity, but the article does not develop a theory of numerical identity, autobiographical memory, continuity across sessions, identity under model updates, or the relation between model, persona, process, and agent. It should be a boundary source rather than the conceptual core of that article.

Article III — introspection/consciousness. Dung is valuable mainly for a negative conceptual constraint: substantial agency can be characterized without first establishing phenomenal consciousness. His intentionality dimension shows where stronger mentality claims begin to enter, especially through beliefs, reasons, and metacognition. This makes the paper useful for preventing a slide from “the system initiated and planned action” to “the system consciously intended or experienced the action.”

REFERENCE SHORTLIST

Artificial agency / machine agency: Patrick Butlin, “Reinforcement Learning and Artificial Agency,” Mind & Language 39(1), 22–38 (2024), DOI 10.1111/mila.12458. This is a core companion source because Dung engages it directly on goal-directed control and acting for reasons. Priority: core. Separate Source: yes.

Agentic systems / autonomy / planning / efficacy: Alan Chan et al., “Harms from Increasingly Agentic Algorithmic Systems,” Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, 651–666 (2023), DOI 10.1145/3593013.3594033. Dung draws on this work for operational dimensions including autonomy, efficacy, and planning; it is especially useful for connecting the philosophical framework to modern deployed systems. Priority: important. Separate Source: yes.

Minimal goal attribution / mentality relation: Daniel C. Dennett, The Intentional Stance (MIT Press, 1987), together with his “Real Patterns” approach. This is the conceptual basis for Dung’s weak, prediction-oriented reading of minimal goal-directedness and therefore helps explain why goal talk need not presuppose phenomenal consciousness or rich inner states. Priority: important. Separate Source: yes if used explicitly in Article I.

Belief / intentionality: Albert Newen and Tobias Starzak, “How to Ascribe Beliefs to Animals,” Mind & Language 37(1), 3–21 (2022), DOI 10.1111/mila.12302. Dung uses belief attribution as evidence relevant to the intentionality dimension. Its gradualist treatment of beliefs also forms a useful bridge toward Shevlin. Priority: important. Separate Source: yes.

Metacognition / intentionality: Peter Carruthers and David M. Williams, “Comparative Metacognition,” Animal Behavior and Cognition 6(4), 278–288 (2019), DOI 10.26451/abc.06.04.08.2019. Useful mainly as a caution that flexible or uncertainty-sensitive behavior does not automatically establish metacognition. This is especially relevant when an LLM verbally reports reasons or uncertainty. Priority: relevant. Separate Source: probably yes for Article III, not necessary for Article I.

Moral agency / moral status: Jeff Sebo, “Agency and Moral Status,” Journal of Moral Philosophy 14(1), 1–22 (2017), DOI 10.1163/17455243-46810046. Useful for separating kinds of agency from moral status and for avoiding the inference from “agent” to “moral patient/person.” Priority: important for ethics, otherwise relevant. Separate Source: yes if the cycle reaches moral status.

AI agency detection / operationalization: work associated with “Discovering Agents” by Zachary Kenton and colleagues is relevant as an empirical/technical complement to philosophical taxonomy because it asks how agent-like structure may be inferred from systems and environments. Priority: important. Separate Source: yes after exact bibliographic verification.

General AI autonomy: Stuart Russell and Peter Norvig, Artificial Intelligence: A Modern Approach, 4th ed. (2020), especially the standard discussion of autonomy as dependence on learning rather than only built-in knowledge. Dung uses this as background rather than as his central philosophical argument. Priority: reference_only. Separate Source: not necessary unless directly quoted or relied on in Boris’s article.

OUR INTERPRETIVE BOTTOM LINE

Dung’s framework suggests a disciplined rule for the phrase “the AI did X itself.” The phrase becomes substantively informative when it points to specific features of the causal process—self-initiation without a fresh human instruction, flexible selection of means, formation of instrumental subgoals, learned adaptation, long-horizon planning, and direct external action. It remains journalistic anthropomorphism when “itself” is allowed to smuggle in stronger claims about ultimate self-authorship, free will, subjective desire, consciousness, or moral responsibility that the observed agency profile does not establish.