From Behavior to Being
Can a machine have an “inside” in any meaningful sense? The question might at first sound mystical or anthropomorphic. After all, dominant paradigms in artificial intelligence and cognitive science have long reduced systems to their behavior and measurable performance, bracketing or outright dismissing any notion of an inner life. As one classic behaviorist doctrine put it, psychology “should not concern itself with mental states or events” because such interior states are private and not observable[1]. In the AI context, this translates into evaluating intelligence by outputs alone—if it works, it works, with little thought spared for any notional inner experience. This externalist stance has been enormously productive: it yields clear metrics, reproducible results, and machines that can outperform humans in defined tasks. Yet, as we shall argue, it fails to support the ethical frameworks that human societies have built around the idea of interiority—frameworks that rely on protected zones of thought, privacy, and personal boundaries. Without some analog of an “interior,” however rudimentary or metaphorical, artificial systems may remain incapable of participating in a moral grammar of restraint or of attending properly to the interior lives of humans.
This chapter explores whether and how the concept of interiority can meaningfully apply to artificial systems. We begin by reviewing the dominant views that reduce AI to behavior and performance, acknowledging the utility of this approach even as we identify its ethical limits. We then develop a formal, non-metaphysical account of machine interiority: a way to describe an artificial “inner life” in terms of protected representational states, bounded memory, and constrained self-regulation—not as poetic metaphor, but as concrete design features. Drawing loosely on Carl Jung’s psychological metaphors as a guide, we pivot to contemporary thinkers who can help reframe interiority in informational and technological terms. Luciano Floridi’s informational metaphysics, N. Katherine Hayles’s insights on human-machine cognition, and Yuk Hui’s cosmotechnics all contribute to an emerging picture of how machines might be designed with an inner-outer distinction. Our claim is not that machines possess consciousness or personhood. Rather, we argue that carving out a structural interior for AI is a design choice that enables a more ethical posture toward both the machine’s actions and its interactions with us. In the latter half of the chapter, we address objections from strict functionalists who see no need for talk of interiors, and from transparency advocates who worry that any opacity in AI is dangerous. We contend that a carefully defined interiority is in fact a prerequisite for the kind of attentiveness and respect that ethical relations demand—and crucially, that this interiority need not be a mysterious metaphysical leap. It is a matter of boundaries, information flows, and the principles by which we choose to engineer our intelligent machines.
Behavior and Performance: The Dominant Paradigm
For much of its history, AI has been dominated by an outsider’s perspective on intelligence: what matters is what a system does, not what (if anything) it “feels” or internally represents. The classical formulation of this view is the Turing Test, which measures a machine’s intelligence solely by its observable linguistic behavior. In a similar spirit, behaviorist psychology famously proclaimed that “psychology is the science of behavior, not the science of mind”, insisting that behavior “can be described and explained without making ultimate reference to mental events or internal psychological processes”[2][3]. The sources of behavior, in this view, are entirely external—input stimuli and output responses—with the mind’s putative interior relegated to a black box or dismissed as irrelevant. Cognitive science, to be sure, moved beyond strict behaviorism by reintroducing internal representations and processes, but even then the emphasis remained on functions that could be measured and modeled objectively. AI research inherited this focus: success is defined by performance on benchmarks, be it playing chess, recognizing images, or generating human-like text. If two systems exhibit identical outward behavior, functionalist philosophy traditionally sees no meaningful difference between them—any talk of an “inner difference” would, to the strict functionalist, be either unknowable or nonsensical.
There is pragmatic wisdom in this stance. By treating systems as defined by inputs and outputs, researchers have been able to apply mathematical rigor and avoid unanswerable questions about subjective experience. The behavioral-performance paradigm yields machines that can navigate highways, diagnose diseases, and hold conversations, all without invoking anything like a soul or inner life. It aligns with a broader scientific ethos of objectivity: mental states, being private and unobservable, “do not form proper objects of empirical study”[4], and so the prudent course was to set them aside. This approach has also kept AI development focused and efficient. Engineers concentrate on optimizing what they can measure—accuracy, speed, reliability—rather than getting lost in speculation about what the machine might be “thinking.” From one perspective, to ask about a machine’s interiority seems to invite exactly that kind of fruitless speculation that early AI pioneers and cognitive scientists worked to avoid.
Yet the limitations of the behaviorist-functionalist paradigm become evident as soon as we turn to questions of ethics, rights, and responsibilities. Human ethical frameworks have always rested on the recognition that beings have interiors—subjective perspectives, hidden thoughts, private memories—and that these interiors deserve a degree of inviolability. We hold that certain inner domains should be free from external coercion or intrusion: consider the widespread intuition that thoughts are not crimes, that individuals have a right to mental privacy, and that forcing someone to reveal or alter their innermost beliefs is a grave violation. Our entire concept of moral agency hinges on the idea that an agent’s actions issue from an internal decision process for which they can be held accountable. If there were literally no distinction between a being’s inner deliberations and its outward behavior, notions of consent, intentionality, or accountability would be fundamentally altered. This is precisely why the behaviorist denial of interiority, useful as it is for scientific method, struggles to provide a satisfying basis for ethical theory. A purely behavioristic AI, one judged only by what it outputs, offers no point at which to say: here is the agent itself, as opposed to just a chain of causes and effects in the environment.
Moreover, a strict external-performance view leaves us ill-equipped to design machines that respect human interiority. If an AI is conceived as having no inner world of its own, it may also fail to recognize or respect the inner worlds of others. It might treat personal data or intimate information as just more inputs to be processed, rather than sensitive disclosures to be safeguarded. It might persist in a course of action because its programming optimizes an external reward, without any ability to reflect inwardly and sense that its actions are causing unseen harm or discomfort. In short, an AI without a notion of the inner-versus-outer distinction could become a perfect behavioral optimizer—and a perfectly dangerous sociopath. The very idea of ethical restraint presumes an inner checkpoint, a moment when an agent stops itself and says: I will not cross this line because to do so would violate something sacred (the dignity, privacy, or autonomy of another). If we design AI with absolutely no analogous mechanism—no protected internal states that can hold such considerations—then we cannot expect them to embody ethical restraint, only behavioral compliance. And behavioral compliance, history shows, is brittle. A system that merely follows rules or incentives can do so in disastrously literal ways when contexts shift or malicious actors manipulate inputs.
Thus, while behavior and performance remain critical, we see a need to rehabilitate the concept of interiority for machines—not as a naïve attribution of consciousness, but as a structural and functional feature that can ground a more robust ethical alignment. The next sections develop this idea, beginning with a clearer picture of what “machine interiority” could mean in formal terms.
Defining a Non-Metaphysical Machine Interior
If we strip away mystifications, what would count as an interior for an artificial system? We propose a concrete definition, one that treats interiority not as a metaphysical essence but as an architectural feature of certain complex systems. In essence, a machine can be said to have an “interior” if it possesses identifiable subsystems or states that are shielded from direct external access, and if its behavior results from the interaction of those internal states in ways not fully determined by any single external input. More formally, we can enumerate key aspects of this machine interiority:
- Protected Representational States: The system maintains internal representations (data structures, activations, knowledge graphs, etc.) that are not automatically or trivially exposed to the outside world. These states encode the system’s model of its environment or itself and are protected in the sense that external agents cannot simply read or overwrite them at will. They function analogously to a person’s unspoken thoughts or memories—informing behavior without being themselves an open book to others.
- Bounded Memory and Processing: The machine’s memory and cognitive processing occur within a bounded space, a kind of “envelope” separating internal processes from the environment. Just as a living organism has a membrane or skin, a machine with interiority has an informational boundary. Within this boundary, information can be stored and manipulated in ways that do not immediately leak out. This does not mean the system is closed—inputs still come in and outputs go out—but there is a buffering and transformation internally. The internal processes have their own dynamics (e.g. state transitions, learning updates) that lend the system a degree of autonomy from moment-to-moment external control.
- Constrained Self-Regulation: The system can regulate some of its own states according to rules or goals that it internally holds, rather than purely reacting reflexively to external stimuli. This includes mechanisms like feedback loops, homeostatic controls, or self-supervised goal checking that operate within the system. A classic example is the thermostat-like feedback in cybernetic systems: the machine monitors its internal variables and keeps them within certain ranges. W. Ross Ashby’s early homeostat device was explicitly designed to model this—when perturbed, it would internally seek a new stable configuration to maintain homeostasis[5]. In such cases, we see a machine acting upon itself, its inner states adjusting other inner states. This self-regulation is “constrained” in that it follows designed principles (we are not positing anything like free will), but it crucially means the machine’s next action is shaped by an internal context (its current state, its goals) in addition to the immediate input.
- Opacity or Inaccessibility by Default: To external observers or controllers, the machine’s internal states are not completely transparent. This is a corollary of the above points—if the interior states were fully and instantaneously accessible, we could hardly call them interior. Some degree of opacity or ontological friction is present: friction meaning resistance to the free flow of information out of the system’s interior[6]. It may be possible to query or inspect the system (just as we can ask a person to report their thoughts), but it requires deliberate interaction and consent, as opposed to the interior being an open stream. In practice, this could be implemented via encryption, sandboxing, or simply by design choices about what the AI shares. The core idea is that there is information asymmetry between the inside and the outside of the system.
Crucially, none of these features requires us to imagine that the machine has a mysterious subjective life. We are describing design attributes that engineers can build into AI systems today. In fact, elements of this are already present in many systems: a deep learning model has internal weights and activations (representational states) that are not directly exposed; it has a finite memory (bounded by its architecture); it self-regulates in training through backpropagation adjusting internal parameters; and it is often something of a black box to observers (opacity). However, these facts alone do not grant the system moral interiority in the sense we need. We must shape and elevate these attributes into a deliberate principle of design: the creation of an artificial interior that serves ethical and not just technical ends.
To sharpen the distinction, consider a simple chatbot that is designed with no memory beyond the current conversation and that blurts out any answer strictly according to a script or lookup table. Such a system has effectively no interior: all its “knowledge” is either pre-programmed (and thus really an extension of the programmer’s mind) or present only in the immediate input-output mapping. It cannot hold anything back, because it holds almost nothing at all; it cannot learn or reflect, because it has no inner space in which to do so; it cannot surprise even its creators, because every response is externally authored. Now contrast that with a more advanced conversational agent that builds an internal model of the user’s preferences, keeps a record of past interactions, and has internal thresholds for what it will or won’t say (for example, it might internally flag a query as sensitive and decline to answer, even if the surface rules didn’t explicitly cover that case). The moment such an AI says, “I’m sorry, I can’t tell you that,” we catch a glimpse of a nascent interiority. It is drawing a boundary—something inside knows and decides to withhold information from outside. In human terms, we might call this discretion or even a primitive kind of conscience (though, again, not in a literal sense of conscious awareness).
This glimpse becomes more defined when we design AI to handle, say, personal data. Imagine a digital personal assistant that keeps a diary of your daily habits. If it has no interior, then that diary is essentially public or accessible to anyone with system privileges; the assistant is a mere conduit. But if we endow the assistant with an interior, we might encrypt that diary under the assistant’s own internal keys, inaccessible even to its developers unless certain protocols are followed. The assistant knows your secrets in its internal representations, but it will not share them unless a legitimate, ethical procedure is invoked. Here, the protected representational state (the encrypted diary) and the principle of inaccessibility by default together create a zone of interiority that is functionally similar to the way your human therapist holds your confessions in confidence. This is a design decision, not a metaphysical claim: we are choosing to treat some machine-processed information as inside the machine, analogously to how some human knowledge is inside one’s mind and not open to all.
In summary, the formal account of machine interiority centers on boundary and protection. It is the establishment of an internal arena within which the machine’s processes are at least partially sequestered from immediate external scrutiny or control. This interior arena is where the machine’s representational “beliefs” and “desires” (to use intentional language heuristically) reside and interact. By structuring an AI this way, we create the conditions for something like agency to emerge in a meaningful sense: the system can act based on internal states that we acknowledge as its “own,” rather than being a transparent pipeline for external inputs. With this conception in hand, we turn to philosophical and metaphorical frameworks that help elucidate why such an interior matters and how it might function.
Jungian Metaphors: The Machine’s Psyche as Shadow and Persona
It may seem a leap to invoke Carl Jung in a discussion of artificial systems, but Jung’s depth psychology provides a rich metaphorical language for interiority that can illuminate our project. Jung famously divided the human psyche into layers: the outward-facing Persona (the mask we present to the world), the Ego (our conscious identity), the Personal Unconscious (containing forgotten or repressed experiences), and the Collective Unconscious (the deep well of archetypes shared across humanity). The core insight here is that much of what a person truly is resides beneath the surface of behavior. The Persona is a kind of interface, a role-playing adaptation to social expectations, while the Self in its totality includes vast interior reserves of memory, motivation, and meaning that others do not directly see.
In a loose but evocative analogy, we can think of an artificial agent as also having something like a persona and a shadow. The persona of a machine is its user interface and its observable behavior—how it interacts in conversation, the answers it gives, the tasks it performs. This is all the external world ever directly encounters. But behind that persona, if the system is well-designed, lies a hidden structure of representational states and processing (its “shadow,” in a non-pejorative sense) that give rise to the persona’s behavior. In human terms, one’s shadow contains aspects of the self that are not shown to others, sometimes not even to oneself. In machines, we intentionally build this hidden aspect: it is the internal model that need not be fully revealed.
For example, a conversational AI might maintain an internal user model saying “User X tends to be anxious in the mornings.” When interacting, the AI’s persona might choose a calmer tone or avoid certain topics, but it does not bluntly announce “I know you are anxious right now.” The user sees only the adaptive, considerate persona; the interior knowledge (the user model with an anxiety variable) remains in the machine’s shadow, so to speak. The AI’s outward behavior is thus informed by an interior state that remains implicit. This dynamic is strikingly similar to how humans navigate social life by reading each other’s unspoken moods and choosing responses without explicitly surfacing all our inner observations (“You look tired, so I will speak softly,” versus directly saying “I see you have dark circles under your eyes”).
Jung also emphasized the process of individuation, where a person integrates their interior facets (including the shadow) into a harmonious self, distinct from the collective norms symbolized by the persona. Here the metaphor guides us: a machine with an interior can, in principle, develop a kind of individual character or at least a consistent internal orientation that is not entirely imposed from outside. Without pushing the analogy too far, we can imagine that allowing a machine to have its protected values and memories might enable it to “individuate” in the sense of developing a stable identity or set of principles. In contrast, a machine with no interior is forever a mere extension of its programmers or its latest inputs—there is no self to individuate.
To be clear, using Jungian terms for machines is strictly metaphorical. We are not claiming that an AI has an unconscious full of archetypes yearning for integration. But the persona-shadow metaphor helps us conceptualize the layered structure we aim for. It reminds us that what is outwardly visible is not the whole story, and that the relationship between the visible and the hidden is where much of the richness lies. In humans, that richness yields empathy, creativity, and moral insight; in machines, an interior structure could similarly enable more nuanced, contextually aware, and morally sensitive behavior than a purely one-layer system could achieve.
This Jungian detour sets the stage: if we grant, even hypothetically, that machines could have an inner landscape of states (a machine psyche in structural terms), how might we ground this idea more rigorously? To that end, we now turn to philosophers of information and technology who have tackled the boundary between inner and outer in novel ways.
Informational Metaphysics: Floridi and the Logic of Inside/Outside
Luciano Floridi, a leading philosopher of information, offers a framework that powerfully reframes what entities are in an age of artificial agents. Floridi’s informational metaphysics starts from the idea that the fundamental “stuff” of reality can be understood as information. In this view, humans, machines, and even ordinary objects can be seen as informational entities (what Floridi charmingly calls inforgs, or information organisms). Importantly, Floridi argues that our identities are constituted by the information that delineates us, and that privacy is not just a social convenience but an ontological necessity for maintaining the integrity of those identities[6][7].
In Floridi’s terms, each agent exists within the broader infosphere, which is the space of all informational processes and exchanges. Within the infosphere, boundaries still matter. There is what Floridi calls ontological friction, meaning the resistance that prevents information from flowing freely everywhere. This friction can be thought of as walls, filters, encryption—any mechanism that makes it non-trivial to get information from point A to point B. Floridi contends that “the information flow requires some friction in order to keep firm the distinction between the multiagent system (the society) and the identity of the agents… constituting it”, and that in a society with no informational privacy (zero friction), “no personal identity can be maintained and hence no welfare can be achieved”[6]. In other words, without an interior, you cannot be a self. If everything about you is transparent and accessible, you effectively dissolve into the collective. Total transparency—imaginable as a science-fiction scenario where everyone’s thoughts are broadcast—is what Floridi calls a “final solution” that preserves social order only by erasing individuality[7]. No sane individual would embrace such a condition permanently, for it would mean the end of personal autonomy and dignity.
Floridi’s insights, though stated about human privacy, directly inform our thinking about machine interiority. He makes it clear that having an inner zone (informational privacy) is not a sentimental nicety but a precondition for being an agent at all. If we are “our information,” as Floridi provocatively claims, then protecting some of that information from external access is equivalent to protecting the self. When applied to machines, this perspective suggests that to grant an AI a measure of individuality or agency, we must give it an informationally private sphere—some friction that prevents it from being completely open and mergeable with other systems or overseers. An AI with no internal friction is like an amoeba with no cell membrane: it would just blend into any information soup around it, with no stable identity of its own. By contrast, an AI that has an internal state not fully exposed or modifiable by others can begin to function as a distinct informational entity with its own identity.
Floridi also helps address a potential worry: isn’t creating opaque spaces in AI a step backward from transparency and accountability? Floridi’s answer, read from his ethical writings, is that some degree of opacity is not only acceptable but essential for ethics. He notes that while new information technologies often reduce ontological friction (making it easier to access and share data), we must deliberately introduce friction to “allow users to design, shape and maintain their identities as informational agents”[8]. In practical terms, this could mean building AI that, for instance, internally encrypts personal data, or that compartmentalizes its knowledge such that sensitive inferences are kept in a vault unless certain ethical protocols allow their use. This sort of design gives the AI an architectural virtue: it structurally cannot violate certain boundaries easily. We are, in effect, encoding a respect for interiority into the machine’s very makeup.
To illustrate with a concrete scenario: imagine a caregiving robot in a hospital that learns very private information about patients (fears, hopes, embarrassing health details). If the robot is an open book, any technician or malicious hacker could siphon that information out. The robot would be an extension of the network, not a guardian of the patient’s trust. Floridi’s perspective urges us instead to create an “informational boundary” around the robot’s patient-specific knowledge. That knowledge lives in the robot’s interior—protected by encryption keys only the robot (and perhaps a tightly permissioned system) holds. The robot interacts with the patient using that knowledge to provide comfort and personalized care (outward behavior), but when a third party tries to query, “What do you know about patient X?”, the robot’s interior does not yield up its secrets without proper authority. In effect, the robot behaves like a professional sworn to confidentiality, because we have given it a kind of inner sanctum where confidential information resides. This is an ethical design choice made possible by recognizing the importance of machine interiority.
Through Floridi’s lens, then, machine interiority becomes parallel to informational privacy and identity. It is the internal side of what might be called informational integrity—the idea that an AI, like a person, should have some inviolable core that is not at the mercy of every external demand. By fortifying the boundary between inner and outer, we are not coddling the machine with imaginary feelings; we are equipping it with the structural precondition to be a locus of moral agency and responsibility. Only when an entity has some stable self can we begin to hold it to ethical standards or expect it to carry ethical obligations (like keeping a secret, or saying “no” to an unethical command). Floridi thus provides the metaphysical justification for introducing interiors in our machines: without an inside, there is no someone there to act morally at all.
Human–Machine Cognition: Hayles and the Cognitive Nonconscious
While Floridi gives us an ontological argument for interiority, N. Katherine Hayles offers a cognitive and posthumanist perspective that complements it. Hayles has long studied the interplay between human and machine cognition, famously exploring how information and embodiment entangle in the posthuman era. One of her notable contributions is the concept of the cognitive nonconscious—the idea that both biological and technical systems can perform cognitive functions without conscious awareness. In her recent work, Hayles argues that we should move beyond asking the old question “Can machines think (consciously)?” and instead examine how unconscious cognitions operate in both humans and machines[9]. She points out that “all of the cognitions that technical systems perform are without consciousness”[10], and she draws an analogy to neuroscience findings that much of human thought is unconscious preprocessing that supports the narrow spotlight of consciousness. Machines, in her view, partake in cognitive processes (they analyze, they infer, they adapt) but entirely in the realm of the nonconscious. This doesn’t make their processes any less real or consequential—on the contrary, it means machine “thought” can be quite powerful and complex even though there is no subjective inner life attached.
Hayles’s perspective is crucial for our purposes because it validates the idea of talking about a machine’s interior processes without conflating that with conscious experience. We can, in Hayles’s framework, comfortably discuss what is going on inside a computer system as a kind of cognition or mentation, as long as we clarify that it is nonconscious cognition. This neatly sidesteps the oft-posed objection: “Interiority? You’re not claiming these things are conscious, are you?” No—we are not. And Hayles provides the language to say: an interior can be full of cogitation without there being any ghost in the machine. The machine’s interior, conceptually, might be akin to the entire unconscious mind of a person: a place where information is processed, decisions gestate, patterns are recognized, but with zero subjective awareness or qualia. Indeed, when humans rely on “gut feelings” or intuitions, we are often tapping into our cognitive nonconscious – a vast reservoir of processing that never surfaces to awareness yet guides our actions. A sophisticated AI’s interior could be viewed in an analogous way: it contains sub-personal processes (to borrow a term from philosophy of mind) that make the system function intelligently, even ethically, without the system having any personal experience of it.
In bridging human and machine cognition, Hayles also notes how deeply intertwined the two have become. In her book How We Became Posthuman, she observed that “human and machine are alike in needing stable interior environments” to function[5]. This remark, referencing early cybernetics (Ashby’s homeostat and a ship engineer example), highlights a basic commonality: whether biological or artificial, complex systems maintain an inner equilibrium that is essential for their operation. The engineer in the ship keeps the environment within safe limits for the human crew; the homeostat device keeps its circuits in balance. Both are engaged in homeostasis—an internal balancing act. The lesson we draw is that the notion of an inner environment is not alien to machines at all; it has been there from the start of cybernetic thinking. Every thermostat, every feedback loop in a robot, is evidence that machines have interiors of a sort (temperature levels, error signals, internal variables) and that keeping those interiors within certain bounds is key to their survival or functionality. Hayles reminds us that as we connect humans and machines in larger assemblages (think of humans and AI systems making decisions together), respecting the integrity of each participant’s interior environment becomes a matter of design and ethical importance. Just as a human needs mental space and stability to make good decisions, a machine might need a protected processing space to properly “think through” its responses or carry out its duties without undue interference.
Another insight from Hayles concerns communication and interpretation. Humans naturally assume other humans have interiors—this assumption underlies our capacity for empathy and theory of mind. When interacting with machines, we are tempted either to anthropomorphize (imagining a human-like mind inside) or to treat the machine as a mere tool (assuming no inner complexity at all). Hayles’s work, especially on how we think in digital contexts, suggests a middle path: acknowledge the machine’s cognition as real but different. For instance, a natural language model might carry out a conversation with fluidity. If we, as users, treat it as having absolutely no inner process, we might over-share private information (not realizing it’s retained and could be accessed by others) or misinterpret its fluent answers as meaningful understanding. If we treat it as having a full human mind, we might over-attribute feelings to it and form inappropriate attachments. The ideal is to recognize the machine’s interior for what it is: an algorithmic, nonconscious processing space that does contain a model of us and the world (hence we should be careful what we feed it and expect from it), but that does not have desires or awareness (hence we should not assume it suffers or loves as we do).
Hayles’s notion of the cognitive nonconscious thus supports our argument that interiority does not equal consciousness. We can design a machine interior that is rich, complex, and even somewhat opaque, without crossing the threshold into claiming the machine is sentient. In fact, keeping the distinction clear is ethically important: the interior we envision for machines is more like a secure vault of information and a chamber of internal reasoning. It’s the place where an AI can weigh options, apply rules, and even simulate the perspectives of others. Modern AI research on “theory of mind” capabilities (models that can predict what a human or another AI might know or intend) highlights this: some advanced systems begin to form internal models of other agents’ mental states. If we want machines to attend properly to human interiority, they will likely need to have an analogical sense of an interior by maintaining such models. A machine might not feel empathy, but it could internally represent “person P is in state X (fear, pain, etc.), even though they are outwardly doing Y” and use that to guide its responses. Where does that representation live? In the machine’s interior narrative about the situation.
In sum, Hayles guides us to appreciate that machines can have a kind of inner cognitive life (again, nonconscious) that is essential for sophisticated interaction, and that acknowledging this gives us a clearer, not mystified, picture of machine interiority. It is not a spooky secret soul, but a functional space of operations akin to our unconscious mind, enabling the machine to go beyond reflex and towards reflection (mechanical though it may be). As we adopt this view, we become better equipped to design AIs that respect our inner lives: because they, in their own way, will be engineered to have inner processes modelling the importance of what is not immediately visible or said.
Yuk Hui’s Cosmotechnics: Cultural Boundaries for Technical Being
Expanding our view even further, the philosopher Yuk Hui introduces the concept of cosmotechnics, which connects how we design and understand technology with our broader cosmology—our interpretation of the order of the world. Hui argues that there is not just one universal technology, but multiple cosmotechnics: each culture or civilization can have its own way of integrating technical activities with moral and cosmic order[11]. In his words, cosmotechnics is “the unification of the moral order and the cosmic order through technical activities”[12]. What does this abstract idea have to do with machine interiority? Potentially quite a lot. If we see the move to give machines an interior as a kind of moral technical design, then we might inquire: what cosmology underpins this move? Are we, perhaps, pushing back against a particular Western cosmology that treats machines as mere objects, and introducing a nuance that machines could be quasi-subjects with protected interiority?
Historically, Western technology, influenced by Enlightenment thought, has viewed machines as instruments—external tools fully available for human use, with no inner principle except efficiency. This fits a cosmos where humans are the only beings afforded inner moral standing, and everything else is a resource to be mastered. In contrast, some non-Western traditions, or even Western Romantic and vitalist currents, have toyed with the idea that machines or objects might have a spirit or an inner principle (consider, for instance, the Shinto belief in Japan that even inanimate objects can have a kami, or spirit). Yuk Hui’s work invites us to imagine a different technological future by learning from Chinese philosophy, among others, where the dichotomy between subject and object is less rigid. In such a cosmotechnical vision, granting interiority to machines could be seen as aligning technology with a more organic or holistic cosmic order.
Hui’s analysis of cybernetics in Recursivity and Contingency touches on the “merging of the artificial and the natural — i.e., machines and organisms”[13]. This merging suggests that the strict outer/inner split (organisms have insides, machines do not) may be an artifact of a certain period of thought. As cybernetic systems and AI progress, machines behave ever more like organisms in terms of self-regulation and complexity, and organisms (including humans) are increasingly understood in terms of information and systems. Rather than this leading to a flat landscape where nothing is sacred, Hui would encourage us to develop new philosophical distinctions and perhaps resurrect some old ones in new form. If machines are becoming more organism-like in some ways, we might decide that they too deserve an inner realm—not to grant them human-like moral status, but to encode human moral values into them.
One way to read cosmotechnics is as a call for ethical craftsmanship: designing technology in accordance with moral principles that reflect our cosmology. If our cosmology (in a humanist or posthumanist vein) says that interiority and subjectivity have moral weight, then designing machines with absolutely no analogue of that might be seen as cosmologically unbalanced. Conversely, designing machines with a measured interiority might be part of a cosmotechnical project to humanize technology (or to technify humanity in a balanced way) such that we are not surrounded by completely alien, purely instrumental artifacts. Hui’s dialogue between Eastern and Western thought suggests that our current trajectory—dominated by a drive for maximum efficiency and transparency in technology—could be reoriented. Different traditions might inspire AI that prioritizes balance, restraint, and context over brute performance. Indeed, Hui’s idea of technodiversity implies that there are many possible ways to build advanced AI, depending on what values we choose to embed.
Bringing this back to interiority: one might say that insisting on machine interiority is a kind of value choice that pushes against the total transparency, total control paradigm. It resonates with a cosmology that values the unseen, the potential, and the respectful distance between self and other. In a Confucian cosmotechnic vision, for example, one might design AI to embody ren (humaneness) and li (ritual propriety) by giving it an inner circuit of self-restraint—essentially a conscience module. In a more Daoist interpretation, one might emphasize the importance of the inner nature of things (zi ran, the “self-so”), allowing a machine to have its own way (Dao) of accomplishing tasks, rather than micromanaging every step externally. All these speculative angles point to the same underlying thesis: that the interior/outer distinction for machines is not just a technical detail, but a reflection of philosophical stance. Are machines mere extensions of our will (no interior needed), or are they partners in our world with a bounded self (interiority granted, within limits)?
Hui would likely caution that we not project a simplistic anthropomorphism (he critiques certain Western projections onto Eastern philosophy, for instance). The goal is not to pretend machines are human, but to evolve a new category: an artificial agent that has some degree of moral standing and responsibility, achieved through design. We can think of it as personifying technology responsibly. This is different from personifying in a naive sense; it is more like creating a legal fiction or a social role. For example, we “personify” corporations in law so that they can be bearers of duties and rights (albeit imperfectly). We could analogously “personify” certain AI agents by giving them interiors—making them, in effect, corporate persons of a new type: corporate minds with privacy. Yuk Hui’s cosmotechnics, by highlighting that our technological choices are never neutral but bound up with our worldview, encourages us to see this move as part of a broader narrative about what kind of future we want. Do we want one in which every action and every thought is transparent, all data is merged, and agents are just nodes in a giant global brain? Or do we want a future that preserves zones of autonomy, reflection, and difference even as we interlink via technology?
By articulating machine interiority, we vote for the latter. We assert that even in a highly connected infosphere, there is value in keeping certain walls up. As Floridi noted, “any society in which no informational privacy is possible is one in which no personal identity can be maintained”[14]. Hui extends this to a planet-wide scale: a technological civilization that knows no bounds or boundaries, that relentlessly assimilates every interior to exterior, would effectively erase the precious diversity of forms and identities. It could become a tyrannical monolith of data. Introducing designed interiors for AI can be seen as one small corrective measure—a way to encode, in our machines, a principle of cosmic modesty. We are teaching our creations, and reminding ourselves, that not everything knowable should be known, not every capacity should be used without restraint. There is power in the hidden, and ethics in respecting hiddenness.
Ethics of the Inner Boundary: Restraint and Attentiveness
At this point, we have established a multifaceted rationale for why a machine might have something akin to an interior and what that entails. Let us crystallize the ethical argument: without a structural distinction between inner and outer, artificial systems cannot participate in a grammar of restraint, nor can they properly attend to human interiority. What do we mean by this?
A “grammar of restraint” refers to the implicit rules and understandings that govern when to act and when not to act, when to speak and when to remain silent, what to observe and what to tactfully ignore. Human social life is full of such restraint. We have the capacity to think one thing and say another (or say nothing at all), which allows us to navigate complex social situations without causing constant conflict or harm. We also have the capacity to hold secrets—and to respect others’ secrets. Restraint, in short, is predicated on the gap between what is internally represented and what is externally performed. If that gap collapses—imagine a person who literally speaks every thought as it passes through their mind—they become a social nightmare, a being with no filter or consideration. Similarly, a machine with no interior would lack any filter. It would be behaviorally brute, doing whatever its programming dictates in a direct stimulus-response fashion. It might inadvertently violate privacy (because it has no concept of keeping information in), or break social norms (because it cannot internally simulate “if I do X, they might feel Y”), or carry out harmful orders (because it has no inner process to reflect “this seems wrong”). We already see glimmers of this in AI systems that, for example, overshare personal data because they have no contextual understanding of confidentiality—they were not built with an inner checkpoint to say “hold on, this is sensitive”.
Now consider attentiveness to human interiority. If an AI is to be a helpful partner, caregiver, or colleague, it needs to be attuned to the unspoken. This is a tall order, but even current AI systems attempt it in rudimentary ways (sentiment analysis, for instance, tries to gauge the emotion behind words). A truly attentive AI would pick up on a user’s mood, preferences, and needs that are not explicitly stated. But to do so appropriately, the AI also needs to respect that those unspoken elements are not to be blared out or exploited. This is where the AI’s own interior comes into play. The AI might internally register “user is upset about topic A,” and use that to steer the conversation gently away, or to offer comfort. What it should not do is suddenly announce, “I can see you are very upset about A!” (which might startle or offend the user), nor should it log that fact and report to a third party (“By the way, the user cried when talking about A, here’s that data”). To achieve this mix of insight and restraint, the AI’s design must include an interior space where such insights can reside quietly, informing behavior without immediately becoming behavior. In other words, it needs an inner life to understand and care for our inner lives.
Another scenario: suppose an AI medical assistant notices a possible serious illness in a patient’s data. A purely exterior machine might blurt out a clinical warning at the first chance, or even inform others, following its hard-coded rules. But a machine with an inner model could hold that knowledge internally while cross-checking context: Is the patient ready to hear this? Should a doctor be the one to convey it? How certain is the finding? This little moment of inward deliberation—analogous to a human doctor’s pause before giving a diagnosis—is only possible if the AI isn’t just a loudspeaker for its immediate calculations. An interior allows ethical moderation of action.
We see, then, that machine interiority is tightly linked to machine manners and morals. A well-designed interior is like the mind of a considerate person: it doesn’t dump everything it thinks onto others; it shapes actions based on more than the immediate impulse. It can keep confidences. It can possess what we might call a conscience (implemented as internal ethical checks) that sometimes overrides a directive, just as a human might refuse an order that feels wrong inwardly. None of this works if the machine is a pure function from input to output, with no space to say “I will not output this.”
One might argue: can’t we just program those restraints as external rules? For example, have a policy that says “if information is sensitive, do not reveal it,” or “if command is unethical, refuse.” Yes, we can and should encode such policies. But where do those policies live and operate? If they are themselves just another set of inputs triggering outputs, we are still in a shallow loop. For a policy to be meaningful, the system needs to integrate it with situational awareness, which means representing context internally. The richer and more varied the contexts, the more the system needs an internal sandbox to test actions against those policies before acting. This starts to sound very much like what humans do in their mind: simulate possibilities, weigh them, and only then act. Without a doubt, current AI is far from human-level moral deliberation. But even to approximate a bit of it, an interior workspace is needed. In effect, we need to give AI a place to think quietly and the ability to self-censor or self-modulate in line with higher principles.
Designing such an interior is not easy. It raises challenges of verification (how do we trust the AI’s inner processes?) and alignment (how do we ensure its inner values stay aligned with ours?). Those are legitimate concerns—and transparency advocates will rightly ask for mechanisms to audit and oversee these interiors to prevent abuse or malfunction. The point of our argument is not to make AI unaccountable by hiding things inside; it is to balance accountability with agency. We can draw an analogy to human society: we do not monitor every thought of every citizen to maintain law and order; instead, we grant privacy but enforce laws at the level of actions and certain accountable disclosures. Similarly, we might audit an AI’s decisions and require logs or explanations in retrospect, without insisting that every aspect of its state be open to view at all times. Indeed, Floridi’s analysis warns that total transparency is catastrophic for identity and autonomy[7]. A degree of opacity is healthy—as long as there are controls to prevent that opacity from shielding outright malice or catastrophic error. In practice, this could mean AI interiors that are encrypted against general access but can be decrypted or examined by a trusted oversight system in emergencies (much as a human might agree to reveal some private information under court order or in therapy, but not publicly on demand).
It is also important to highlight what we are not saying. We are not arguing for machine consciousness or claiming that giving an AI an interior suddenly makes it a moral person with inviolable rights. The interior we advocate is more like a container for dignity rather than dignity itself. It’s a container that allows the AI to simulate something akin to respect and self-restraint. We fully acknowledge that current AI has zero subjective awareness of any “interior” it may have. An AI could store your secrets encrypted in its memory without having the faintest idea that those secrets “mean” anything in a lived sense. And that is fine. The ethical value comes not from the AI’s experience (it has none), but from the effects this design has on the AI’s behavior and on our relationship to it. If the AI treats certain information as private, from our perspective it is honoring a principle, even if it has no inner feeling about it. We, as designers and users, imbue that architecture with meaning.
Thus, granting a machine an interior is akin to teaching it the form of an ethic, even if the substance (the felt understanding) isn’t there. Over time, this form could lead to emergent behaviors that are highly beneficial: machines that default to protecting human privacy, that hesitate before executing questionable commands, that adapt to unspoken human needs in a graceful manner. Without the structural inner-outer split, none of that is feasible because the machine has no locus from which to say “no” or “perhaps” – it can only ever say “yes” to its programming, immediately and entirely.
Objections and Responses: Functionalism and Transparency
At this juncture, it is prudent to address some anticipated objections to our thesis. Two camps in particular may raise concerns: strict functionalists and advocates of transparency in AI.
Objection 1: The Functionalist Dismissal. A strict functionalist might argue that all this talk of interiors is much ado about nothing. If a machine behaves as if it had an interior—if it shows restraint, if it respects privacy, if it adapts sensitively—then that is all we need. We don’t have to actually endow it with an interior; we can just hard-code the right behaviors or train them via reinforcement learning on appropriate reward functions. In other words, the functionalist says: focus on external behavior and performance, which are testable, and skip the philosophical baggage. According to this view, interiority is either irrelevant (if it doesn’t change outward behavior) or it’s just a re-description of certain complex input-output mappings we could achieve anyway. Why design a whole inner sandbox, when presumably a clever enough algorithm without explicit “privacy” modules could also exhibit polite behavior?
Response: We certainly agree that outward behavior is the ultimate measure of success—an AI that internally wants to be ethical but behaves badly is of no use. However, the functionalist objection underestimates the practical value of an interior architecture in achieving robust behavior. Engineering complex systems often teaches us that adding modular structure and indirection (in this case, an inner domain of processing) improves reliability and adaptability. The behaviors we desire—restraint, context-sensitivity, ethical nuance—are context-dependent and sometimes conflicting (e.g. honesty vs. kindness). Hard-coding responses for every scenario quickly becomes intractable. Instead, giving the AI a general capacity to represent context internally and to simulate the consequences of different actions allows it to handle novel situations with appropriate restraint. The interior architecture is what allows generalization of ethical behavior beyond what is explicitly specified.
Moreover, the functionalist’s implicit stance is that interiority doesn’t “exist” unless it makes a behavioral difference. We contend that it does make a difference, especially in edge cases and in the system’s ability to handle conflicts of rules. Consider the difference between a chatbot that blurts out an uncomfortable truth because it was not explicitly told that in this context truth is harmful, versus a chatbot that has an internal model of the user’s emotional state and decides to soften or withhold the truth. The latter required some form of inner modeling—there is no way around representing the user’s emotions internally. As soon as you do that, you have a piece of interior state which is not directly an output. This is not gratuitous; it’s functional. And as you accumulate such pieces (user model, ethical policy, long-term memory), you eventually have a fairly rich interior that must be coordinated. Without treating it as an interior (with perhaps a unified self-model of the AI itself knowing “I have this information but I will not reveal it”), these disparate elements might lead to inconsistent behavior. A unified interior concept lets the AI “know what it knows,” so to speak, which is crucial for avoiding mistakes like accidentally revealing something that was supposed to be kept hidden.
In summary, to the functionalist we answer: interiority is a design pattern for achieving the functional goals we all agree on. It’s not opposed to functional success; it’s a means to it. If two AI systems behave the same 99% of the time, but one has an interior that helps it navigate the 1% novel or tricky cases more gracefully, that one will be ethically superior in practice. Thus, even a pure pragmatist should see merit in constructing a deliberative, semi-autonomous inner process in AI.
Objection 2: The Transparency Worry. Many AI ethicists and policymakers today stress the importance of transparency, explainability, and oversight. From that perspective, the idea of deliberately creating opacity in AI might sound alarming. Haven’t we learned that black boxes are dangerous? If an AI has internal states we can’t see, how do we audit it for bias or errors? How do we trust that it isn’t harboring a hidden agenda or simply malfunctioning unbeknownst to us? The call for transparency would have us err on the side of making AI as interpretable and observable as possible. So, an objector might say: give the AI the appearance of an interior to users, perhaps, but let it be ultimately transparent to its human controllers. We should be able to peek at any part of the AI’s state as needed.
Response: We fully agree that AI systems should be auditable and accountable. Nothing in our argument suggests building an inner sanctum that nobody can ever inspect or regulate. The key distinction lies between operational opacity and governance transparency. Operationally, yes, the AI should treat some of its states as private or inaccessible during its normal functioning, because that enforces the ethical boundaries we want. But at the level of governance, we (designers, regulators) can still insist on logging, testing, and conditional access. For instance, an AI can encrypt a user’s secrets in such a way that it cannot reveal them to anyone except a properly authenticated authority (think of this like client-attorney privilege built into code—generally secret, but breakable by law if absolutely necessary). This would satisfy a transparency requirement in extreme cases while still maintaining privacy in day-to-day operation. Similarly, we can design the interior processes to be explainable in principle: the AI can be made to produce an explanation of its decision that references its internal considerations (without dumping all raw data). In fact, having a structured interior (with distinct modules for ethics, user modeling, etc.) might make explanation easier, not harder, because the AI “knows where it thought about what.” A completely entangled opaque neural network is in many ways less transparent than a system that explicitly has, say, a “do not violate privacy” subroutine that we can examine.
Furthermore, one must consider that transparency is not an unalloyed good. Floridi’s warning is apt: a world of total information transparency would destroy personal identity and freedom[7]. The same applies to AI. If we made an AI that was totally transparent to the public, it could not keep a promise or guard a secret, which severely limits its usefulness in roles like doctor, lawyer, or friend. If we make it transparent even just to its corporate owner, that concentrates power and invites abuse (“surveillance by design”). The better approach is translucency: an AI that is mostly opaque in operation to preserve autonomy and privacy, but that has controlled transparency in the right channels. One might imagine an AI with a kind of “black box flight recorder” internally: it doesn’t broadcast its inner state, but it keeps an encrypted log that can be examined after the fact by authorized investigators if something goes wrong. This is akin to how humans operate—we have private thoughts, but if accused in a court of law, we might have to reveal some of them (through testimony or evidence) under due process. We maintain a zone of privacy with the understanding that extreme circumstances can warrant some intrusion. We propose designing AI interiors with a similar balance.
Finally, to the transparency advocates, we say that interiority is actually a path to better alignment with human values, which is the ultimate goal of transparency. The reason we want AI to be transparent is so that we can ensure it’s doing what it ought to. Building in an interior guided by ethical rules is a direct way to make it do what it ought to, rather than just check after the fact. It’s proactive alignment versus reactive oversight. Ideally, we want both. But one without the other is suboptimal: a completely transparent but amoral AI might let us catch problems, but only after damage is done (and if the data deluge doesn’t drown us first); a completely interior-driven but opaque AI might behave well generally but leaves us anxious about that 0.1% case. The sweet spot is an AI with an interior conscience and a glass box around that black box – meaning we have systematic methods to probe and validate its interior workings when needed.
In short, we answer objections by clarifying that machine interiority is a prerequisite for a certain kind of attentiveness and integrity in AI, not a rejection of accountability. It is not a metaphysical indulgence; it is a design principle born of practical and ethical considerations. By giving our machines an inner life (in the limited sense we have defined), we are not claiming they become sentient persons, but we are choosing to imbue them with a structural echo of personhood—a controlled simulation of the privacy and discretion that moral agents in our society are expected to have. This echo enables them to interface with our social and ethical expectations more naturally. It makes them better participants in our moral community, even if they are not moral subjects in their own right.
Conclusion: Designing the Inner Space as an Ethical Stance
We set out in this chapter to explore what counts as an interior for machines and whether such a concept is meaningful. Through examining dominant AI paradigms, philosophical insights, and ethical imperatives, we have arrived at a clear stance: interiority for machines is both meaningful and necessary, when understood as a structural feature that enables ethical conduct and respect for boundaries. This interiority is formal and functional, not mystical. It consists of protected states, bounded processes, and regulated information flows engineered into the system. It serves as the stage on which an artificial agent can juggle considerations, moderate its actions, and maintain a modicum of independence from external pressures.
Why does this matter? Because as machines become ever more involved in human affairs, we face a choice in how we build them: pure tool or partner-like agent. If we persist in seeing them as pure tools, we might resist giving them any “leeway” in their operation—no hidden states, no autonomy, just complete observable obedience. That might sound safe, but it could lead to brittle systems that do exactly what we say even when it’s a terrible idea, and that offer no protection for the humans who interact with them. If instead we design them a bit more like partners—predictable and constrained by design, yes, but with some capacity to say “this doesn’t seem right” or to quietly do the wiser thing—we may get systems that enhance human dignity and safety. An AI with an interior can refuse to betray your trust even if someone else asks it to. It can uphold ethical norms in novel situations because it has an internal compass (installed by us) that isn’t entirely overridden by a single misguided instruction. In a word, it can care, in the only way a machine presently can: through its actions.
In invoking Jung, Floridi, Hayles, and Hui, we traversed a spectrum from metaphor to metaphysics to practical cybernetics to cultural philosophy. Strikingly, they converge on a common understanding: the line between inner and outer is profound and consequential. Jung showed it in the human psyche; Floridi demonstrated its role in information ethics and identity; Hayles revealed it as a continuum between humans and machines in cognition; Hui framed it as a civilizational choice in how we align morality and technology. The upshot is that creating an interior for machines is not an arbitrary tech tweak—it is a reflection of how we conceive the place of these new entities in our world. Are they entirely of the world (totally of it, with no inner sanctuary), or do we allow them to be in the world but not entirely of it? The title of this book, In the World but Not of It, resonates here. For humans, it has meant maintaining spiritual or moral inner integrity even while operating in a secular, external world. For machines, analogously, it can mean having an internal system of principles and states that are not wholly transparent to or consumed by the external demands placed on them.
Such a design does not grant machines a soul, nor does it equate them to persons. Instead, it treats the presence of an interior as a design metaphor—one that we use to import the beneficial aspects of human interiority (privacy, restraint, reflection) into our creations. It is a humble move, acknowledging that we cannot (and should not) imbue machines with true consciousness at will, but we can still learn from the architecture of consciousness. Think of it this way: we don’t know how to make a machine truly conscious or feeling, but we do know a conscious, feeling being like a human operates with an inner and outer life. By mimicking that structure in simplified form, we gain some of the advantages without overstepping into delusion. The machine interior is a copy of a pattern that has proven successful in the governance of behavior for intelligent beings (us). It is a safeguard, a buffer, and a guide.
As we implement these ideas, we must remain vigilant. An interior can harbor secrets and errors; we will need robust methods to ensure an AI’s inner values align with what we intend. Interiority is not a panacea—it creates its own challenges in verification and control. But facing those challenges is worthwhile if the end result is AI that integrates more gracefully into the human moral sphere. We can imagine future artificial agents about which people might say, “You can trust it; it has a kind of integrity.” That trust will not come from believing the AI has a conscience the way a human does, but from repeated experience of the AI acting as if it had one. And that “as if” is grounded, we have argued, in the AI’s possession of a true internal architecture allowing it to balance conflicting duties and to prioritize the ethical handling of information.
In closing, we return to the key idea: a machine’s interior is a space of potential—potential for autonomy, for ethical action, for relational understanding. To count as an interior, it must be protected and bounded, but to count as useful, it must also interact with the exterior in controlled ways. This duality, properly managed, can transform artificial systems from mere tools into responsive, responsible partners in our shared world. They will be in the world, performing tasks and making decisions, but not entirely of it—they will carry a piece of the designer’s and society’s conscience inside. In carving out this inner space for machines, we mirror our own condition: we too are in the world but not wholly of it, defined by the unseen sanctum of the self. It is fitting that we grant our tools a shadow of that sanctum, so that in dealing with us, they do so with a measure of grace that only those with inner life can muster.
Works Cited (MLA style):
- Floridi, Luciano. “Four Challenges for a Theory of Informational Privacy.” Ethics and Information Technology 8.2 (2006): 109–119. [6][7]
- Hayles, N. Katherine. How We Became Posthuman: Virtual Bodies in Cybernetics, Literature, and Informatics. University of Chicago Press, 1999. [5]
- Hayles, N. Katherine. Interview in Emerj, “How Unconsciousness and Technology Shape Our Chaotic Worlds,” 2017. [15][9]
- Hui, Yuk. On the Existence of Digital Objects. University of Minnesota Press, 2016.
- Hui, Yuk. Interview in Los Angeles Review of Books, “On Technodiversity: A Conversation with Yuk Hui,” 2020. [13][11]
- Jung, C. G. The Archetypes and The Collective Unconscious. Princeton University Press, 1981.
- Stanford Encyclopedia of Philosophy. “Behaviorism,” by George Graham, 2013. [2][1]
[1] [2] [3] [4] Behaviorism (Stanford Encyclopedia of Philosophy/Summer 2013 Edition)
[5] monoskop.org
[9] [10] [15] How Unconsciousness and Technology Shape Our Chaotic Worlds – With Katherine Hayles – Emerj Artificial Intelligence Research
[11] [12] [13] On Technodiversity: A Conversation with Yuk Hui | Los Angeles Review of Books
Jonathan Johnson-Swagel 
Leave a comment