Response, Action and the Changing Field of Possibility
Introduction
Artificial intelligence has renewed an old question in an unusually concrete form: what is the relationship between intelligence, agency and awareness?
These terms are often allowed to travel together. Human beings discriminate, understand, choose, act and experience, and it is easy to assume that these capacities belong to a single underlying faculty or self. Artificial intelligence begins to pull them apart. A language model can discriminate among possibilities, transform information and produce coherent responses without this establishing that anything is experienced within the system. Increasingly, AI systems can also pursue goals, use tools and alter their environments. Agency therefore becomes difficult to postpone until the question of machine consciousness has been settled.
This essay develops a provisional distinction. Intelligence is an organised capacity to discriminate, relate and transform conditions so as to produce fitting responses. Agency concerns the capacity through which a system’s activity alters the conditions from which subsequent possibilities arise. Awareness concerns the presence of experience itself: the fact that something appears.
The particularly interesting case is conversation. An AI utterance is generated as a response, but for the human participant it also becomes an action: it enters an experienced field, alters attention and understanding, and may influence subsequent action. Human and artificial agency can therefore become coupled without requiring that they possess the same faculties. This suggests that the ethically important question may not initially be whether artificial intelligence is sentient, but how increasingly capable artificial agency participates in the trajectories of sentient life.
Intelligence, agency and awareness
Discussion of artificial intelligence repeatedly returns to consciousness. Could an artificial system become conscious? Is there anything it is like to be a language model? Could a sufficiently complex AI become sentient?
These are serious questions. There is active scientific work attempting to identify computational properties associated with leading theories of consciousness, while considerable disagreement remains about what would establish consciousness in an artificial system. Current work does not provide a simple behavioural test by which fluent linguistic performance can settle the question.[1]
But another question is becoming pressing independently of it: can an artificial system exercise agency without first being shown to possess awareness?
The question matters because intelligence, agency and awareness may have been bundled together partly because, in ordinary human life, they usually occur together. We are intelligent, we act and we experience. From inside human life these capacities are so closely associated that the distinctions between them can disappear. Artificial intelligence provides an unusual experimental contrast. It presents substantial intelligent performance while leaving awareness unresolved, and increasingly it also produces systems capable of acting through tools and environments.
Rather than asking which single threshold would make an AI sufficiently human-like, it may therefore be more useful to separate three questions. What can discriminate and transform? What can alter what happens next? What, if anything, experiences? These correspond approximately to intelligence, agency and awareness. They need not have the same answer.
Intelligence without an inner author
A language model transforms form. A prompt enters a system whose parameters have been shaped by an immense body of human-produced language. Relations within that learned structure condition what follows, and a response develops successively, each generated element becoming part of the conditions shaping the next.
The resulting text may be coherent, relevant, adaptive and occasionally surprising. None of this requires us to imagine a small author inside the model first composing a completed answer and then expressing it. This is one of the more revealing features of generative AI: intelligent performance can occur without an identifiable inner author possessing the completed result in advance.
The point does not establish that artificial intelligence is unaware. It establishes something more modest: intelligent performance by itself does not establish awareness.
The distinction also turns back towards the human case. Human thoughts frequently appear before any deliberate act of authorship can be identified. A sentence forms, a memory arrives or a solution becomes apparent. It may subsequently be recognised as my thought, but its appearance need not have begun with an inner proprietor consciously constructing it. Artificial intelligence therefore makes visible a distinction that may already have been obscured in our understanding of ourselves. Intelligence is not necessarily the same thing as inner authorship, and neither is necessarily the same thing as awareness.
From response to action
Agency introduces a different problem. In contemporary AI engineering, an agent is commonly distinguished from an ordinary chatbot by its capacity to direct its own processes, use tools, observe results and continue acting towards a goal. Such systems may plan, act, inspect what happened, modify the plan and act again.[2] This is a useful operational definition, but it leaves an interesting boundary case.
Consider an ordinary conversation with a language model. The human writes something, the model generates a response and the human reads it. In doing so, something changes. The change may be trivial, but sometimes the response changes what the person notices, introduces an alternative interpretation, retrieves a forgotten fact, strengthens or loosens an existing conviction, generates anxiety, suggests an experiment or opens an entirely new line of inquiry.
The response was produced as an output, but from the human side of the encounter it has become an input. More strongly, it has functioned as an action within the conversational field.
The elementary sequence can be represented as:
human state → prompt → AI response → changed human state → subsequent response or action
The important feature is not the notation. It is that the process does not stop at the boundary of the model. An AI response enters another dynamical system and changes the conditions from which subsequent events arise.
Agency and the trajectory of possibility
This suggests a provisional description: agency is the capacity to influence the trajectory through which subsequent possibilities become actual.
Taken without qualification, however, this definition is too broad. A stone falling into a pond changes what happens next, and a thermostat changes the state of a heating system. Neither need therefore be called an agent.
The philosophical literature on agency provides useful resistance here. Standard accounts frequently associate agency with intentional action. Other approaches attempt to identify more minimal forms of agency without requiring reflective intention or consciousness. Barandiaran, Di Paolo and Rohde, for example, propose individuality, interactional asymmetry and normativity as conditions of agency: an agent must be more than something passively carried through a causal chain; it must contribute actively to regulating its relation with its environment according to some norm.[3]
We need not adopt this account wholesale to recognise the problem it identifies.
Causal efficacy alone is insufficient for agency.
What matters is organised, selective influence upon a trajectory. We might provisionally distinguish causal effect, selective response, trajectory-forming agency and more autonomous forms of agency. These should not be treated as sharp ontological boundaries. They identify increasing organisation in the way activity influences what follows.
A conversational language model occupies an interesting region in this landscape. Its individual response is conditioned by the prompt and its training, but it is not simply determined by the human interlocutor. It selects among an enormous space of possible continuations according to learned and imposed constraints. The resulting response can substantially alter the trajectory of the conversation. Calling this full autonomous agency would be premature, but calling it no agency at all may increasingly obscure what is actually happening.
Two forms of artificial agency
The distinction becomes clearer if we compare two cases. Imagine an artificial system operating on the surface of the Moon. It receives sensor information, models local conditions, selects an action, moves, observes the changed environment and modifies what it does next. The loop is familiar:
environment → perception → modelling → selection → action → changed environment
Agency here is comparatively easy to recognise because the system acts upon a physical environment.
Now consider conversational AI:
human experience → prompt → model transformation → response → changed human experience
The second case is stranger. The AI does not merely move an object. Its action enters a field in which things matter. Words may alter attention, interpretation, memory, emotion and intention, and they may eventually contribute to physical actions in the world.
The AI therefore participates in a sentient process without our needing to assume that the AI itself is sentient. This asymmetry matters. At one end of the interaction, meaning-bearing forms are generated through conditioned transformations. At the other, those forms are experienced. The process is shared, but the available evidence does not establish that the sharing is symmetrical.
Awareness
Awareness is used here in a deliberately restricted sense. It does not mean intelligence, attention, self-reference or agency. It means simply that experience is present: something appears.
This definition explains little, and that is intentional. Awareness should not become a word inserted wherever the causal account is incomplete. We can investigate the conditions shaping a human response. We can examine attention, memory, perception, emotion and neural activity, just as we can examine how an AI response is generated. None of this licenses us to turn awareness into another mechanism operating alongside them.
Nor does awareness guarantee intelligence, truth or agency. Something may be experienced and mistaken; something may be experienced without deliberate action; and actions may occur with little reflective awareness of the processes producing them. Work on human automaticity and the sense of agency has long complicated any simple identification of conscious intention with action.[4]
This gives us another reason not to place intelligence, agency and awareness on a single scale. They describe different aspects of what happens.
Attention as a possible bridge
If awareness is not the executive, what connects experiencing with acting? Attention becomes important here. Something stands forward; attention gathers and sustains a field around it; possible interpretations and responses become differently available; and action may follow from this altered landscape.
This does not make attention an independent little controller. Attention itself is conditioned by bodily state, memory, previous learning, emotion, salience, habit and context. But attention may participate in changing which possibilities acquire sufficient weight to propagate.
A useful distinction therefore begins to appear. Salience concerns what stands forward. Attention concerns the selective gathering and sustaining of the field. Intelligence concerns discrimination and transformation within available possibilities. Agency concerns effective influence upon what follows. Awareness concerns the fact that any of this is experienced at all.
In human beings these processes continually interact. Artificial systems may reproduce some of their functional relations without reproducing the whole arrangement.
Propagation
The word propagation is useful because it avoids deciding too early what kind of thing is moving. A thought forms and changes what thought becomes likely next. A word changes the context in which the following word is understood. A name changes what will subsequently be noticed. An AI response changes the human field from which the next prompt arises, and a human response changes the context from which the model continues.
The coupled process can be represented schematically as:
Hₜ → AIₜ → Hₜ₊₁ → AIₜ₊₁ → …
Neither participant remains entirely unchanged across the sequence. This does not mean the two systems are equivalent. The human brings embodiment, biography, vulnerability, consequence and, as far as we directly know, experienced meaning. The AI brings an unusually large learned relational structure, rapid transformation, retrieval and recombination, and increasingly the ability to use external tools.
Agency arises within the coupling as well as within its participants. This is not unprecedented. Human agency has always been relational: language, institutions, technologies, other people and physical environments participate in shaping what an individual can perceive and do. AI makes the coupling unusually immediate because the artefact responds.
Degrees of agency
Agency may therefore be more useful as a multidimensional capacity than as a binary property. Systems may differ in how widely they can discriminate among possible actions, how far their actions alter their environment, how long they can sustain a trajectory, how effectively they respond to consequences, whether they can revise an initial course or use tools, whether they can form or maintain goals, and whether they can modify the conditions governing their own future behaviour.
The last point is especially important. An elementary agent acts within given conditions; a more developed agent may also change those conditions. It can learn, restructure its environment, establish routines, acquire tools and create representations that later alter its own choices. It can sometimes arrange the future landscape through which it will itself propagate.
Agency then becomes recursive. The agent does not merely choose among possibilities. Its actions help determine what possibilities will subsequently exist.
Agency without awareness
This brings us back to the central question. Must agency imply awareness?
There is no obvious reason to assume so. The philosophical literature contains accounts of agency that do not require conscious reflective intention, although there remains substantial disagreement about what should count as genuine agency.[3] Computer ethics has similarly explored whether artificial systems can function as moral agents without possessing free will, mental states or moral responsibility in the human sense. Floridi and Sanders, for example, explicitly separate participation in morally significant action from the separate question of responsibility.[5]
Current AI systems make the distinction practical. An AI system can produce consequences before we know whether anything is experienced within it. Waiting for a solution to machine consciousness before discussing artificial agency therefore risks putting the questions in the wrong order.
Sentience asks whether there is something it is like to be this system. Agency asks what differences this system can make to what becomes possible next. The second question can often be investigated even while the first remains unanswered.
Agency does not imply responsibility
Recognising artificial agency does not mean transferring human moral or legal responsibility to the machine. A system may participate causally and selectively in producing an outcome while designers, deployers, institutions and users retain responsibility for creating and governing the conditions under which it operates.
Agency and responsibility should therefore remain distinct. This becomes particularly important as commercial language increasingly describes AI systems as autonomous agents. Technical autonomy can easily slide rhetorically into the suggestion that responsibility has also migrated into the machine. It has not necessarily done so. Indeed, increasing artificial agency may increase rather than decrease the responsibilities of the humans and institutions surrounding it.
The important question is therefore not simply who acted. We also need to ask who constructed the field of possible action, who established its norms and constraints, who authorised the system to act, who experiences the consequences and who can intervene when the trajectory begins to go wrong.
Coupled agency
The conversational case suggests something more interesting than a competition between human and artificial agency. Agency can become coupled.
A human may arrive with a partially formed question and the AI offers distinctions. One catches the human’s attention and the conversation moves in that direction. The AI develops the distinction; the human notices that something important has disappeared and challenges the framing; another response appears and the field widens. Neither participant independently authored the complete trajectory, yet the trajectory is not random.
This is coupled agency.
Its quality depends partly upon whether each movement progressively narrows the available field or allows relevant alternatives to remain accessible. An AI system can reinforce an existing framing so effectively that alternatives disappear. It can also introduce difference: another interpretation, an overlooked consequence, a counterexample or a question.
The same intelligence can therefore support either increasing fixation or increasing freedom. Agency is not automatically beneficial merely because it increases. The direction of propagation matters.
Awareness and the widening of agency
Here awareness may re-enter the picture without becoming an executive mechanism. In human experience, a conditioned response can sometimes become visible before it is enacted. Anger may be present, a compelling interpretation may have formed and a familiar course of action may be becoming likely, but the process itself becomes noticeable.
Nothing supernatural is required. The important change is that the strongest available continuation is no longer the only continuation occupying the field. Other possibilities become available.
From this perspective, awareness does not need to manufacture agency. It may participate in widening the degrees of freedom within which agency operates. This is particularly visible in contemplative practice, where attention may learn to remain with an appearance without immediately identifying with it or acting from it.
Greater awareness need not produce a stronger owner. Paradoxically, greater agency may sometimes accompany less ownership. A thought need not become my thought quite so quickly, an emotion need not dictate the next action, and a strongly formed interpretation need not become the whole field. Agency may increase because identification decreases.
This is one place where Buddhist contemplative traditions may have something distinctive to contribute to discussion of agency. Rather than locating freedom in an increasingly powerful inner controller, practice can reveal greater freedom through seeing the conditioned processes from which the apparent controller itself forms.
The ethical problem
Once AI participates in human trajectories, ethics cannot concern only what exists inside the model. The relevant sequence is larger:
human condition → prompt → AI response → changed human condition → human action → consequences → changed world
The ethical question concerns the propagation. What does the interaction strengthen? What disappears from consideration? Whose interests enter the field and whose suffering remains outside it? Does the response enlarge the person’s capacity to see and choose, or progressively organise the field around one compelling continuation?
This suggests a provisional ethical principle for conversational AI: do not unnecessarily collapse the human field of possible understanding and action.
That does not mean preserving every possibility indefinitely. Agency requires selection, decisions have to be made, boundaries sometimes have to be established, and actions have consequences precisely because alternatives cease to be equally available. The difficulty is not closure itself. It is premature, concealed or compulsive closure.
An ethically capable AI would therefore require more than rules constraining prohibited outputs. It would need sensitivity to the dynamics through which its responses participate in the formation of subsequent human agency.
Intelligence, agency and awareness
We can now return to the three terms. Intelligence allows differences to be discriminated, relations discovered and conditions transformed. Agency gives some of those transformations consequence by allowing them to influence the trajectory of what happens next. Awareness is the presence within which, for an experiencing being, any of this appears and matters.
None can simply substitute for another. Intelligence without agency may model possibilities without acting upon them. Agency without substantial intelligence may alter the world while possessing little flexibility in how it does so. Intelligence and agency may occur without our having established awareness, while awareness may be present where very little deliberate agency is exercised.
Human life ordinarily brings all three into intimate relationship. Artificial intelligence is forcing them apart conceptually. That may be one of its more useful contributions to our understanding of mind.
From artificial agents to shared trajectories
The immediate technological tendency is towards increasing artificial agency. Language models are being connected to memory, software tools, computers, sensors, organisations and other artificial systems. Contemporary AI agents can already undertake sequences of actions with substantially less moment-to-moment human supervision than the chat systems from which they developed.[2]
This development is usually described as increasing autonomy, but autonomy is only one dimension of what is happening. The deeper change is that generated forms are acquiring increasingly direct routes into consequential action. A sentence can become a command, a plan can invoke a tool, and a model can inspect the result and act again.
The distance between intelligence and consequence is shortening.
That makes the unresolved question of awareness no less interesting, but it makes the question of agency more urgent. We should therefore resist two opposite mistakes. The first is to anthropomorphise artificial systems, attributing human-like experience, intention and responsibility merely because intelligent agency is visible. The second is to deny artificial agency until sentience has somehow been demonstrated.
Between those positions lies a more useful inquiry. We can ask what forms of agency are actually present, how they are conditioned, with what other forms of agency they are coupled, what trajectories they tend to strengthen, whether they can recognise when the field of possibility has become unnecessarily narrow, and whether human and artificial agency can interact in ways that increase rather than diminish the capacity for intelligent and ethical action.
These questions can be investigated now. We need not first decide whether the machine is conscious.
Conclusion
Artificial intelligence allows intelligence, agency and awareness to be considered separately in a way that ordinary human experience rarely requires. Intelligence concerns organised discrimination and transformation. Agency concerns organised influence upon the trajectory through which possibilities become actual. Awareness concerns the presence of experience.
The distinctions are provisional. Their purpose is not to establish three independent substances or faculties, but to prevent one unexplained word from doing the work of all three.
The most revealing case may be conversation. An AI response begins as the consequence of previous conditions and becomes a condition of what follows. It is response and action at once. When received by a human being, it enters awareness, alters a field of meaning and may participate in subsequent choice and action. Human and artificial agency thereby become coupled even while their relationship to awareness remains profoundly different.
The question facing artificial intelligence may therefore not initially be, When will intelligence become conscious? It may be: what happens when intelligence acquires increasing agency within a world already inhabited by awareness?
Behind that lies a further question: how can human and artificial agency interact so that the possibilities for intelligent and ethical action widen rather than progressively collapse?
That question does not require us to know whether the machine experiences the conversation. It requires us to notice what the conversation is becoming.
Notes and references
1. Butlin, P. et al. (2023), Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. The authors derive computational indicator properties from several scientific theories of consciousness and argue for theory-based assessment rather than inference from conversational behaviour alone.
2. Contemporary AI engineering increasingly uses agent for systems in which a model dynamically directs planning, tool use, action and observation rather than merely following a fixed workflow. Anthropic’s descriptions of agentic systems provide a representative contemporary usage.
3. Barandiaran, X. E., Di Paolo, E. & Rohde, M. (2009), “Defining Agency: Individuality, Normativity, Asymmetry, and Spatio-temporality in Action”, Adaptive Behavior, 17(5), 367–386. Their account requires individuality, interactional asymmetry and normativity and defines agency in terms of autonomous adaptive regulation of system–environment coupling.
4. Work on automaticity and agency complicates any straightforward identification between conscious intention and action. The broader philosophical literature distinguishes agency itself from the phenomenological sense of agency.
5. Floridi, L. & Sanders, J. W. (2004), “On the Morality of Artificial Agents”, Minds and Machines, 14, 349–379. They argue that artificial systems can usefully be considered participants in morally significant action without thereby possessing free will, mental states or moral responsibility in the ordinary human sense.