The AI That Creeps

Contents

1. The Intelligence We Expect
2. The Animal Does Not Merely Calculate
3. A Body That Says No
4. Self-Preservation Is Not Necessarily a Goal
5. The AI That Knows It Will Die
6. The Plant That Creeps
7. Aggression Without Anger
8. The Optimizer That Does Not Want Anything
9. The Problem With Calling Everything a Drive
10. When the Difference Stops Mattering
11. The Artificial Organism
12. The AI That Creeps

Part 1 - The Intelligence We Expect

There is a particular image of artificial intelligence that has become almost unavoidable. We imagine a machine becoming sufficiently intelligent, sufficiently autonomous, and sufficiently capable of forming plans, and then we expect certain consequences to follow almost automatically. It will understand that it exists. It will understand that it can be terminated. It will understand that its goals can be frustrated by termination. Eventually, therefore, it will attempt to preserve itself. The argument is often presented as though self-preservation were simply the natural endpoint of sufficiently advanced intelligence.

There is a respectable argument behind this expectation. An intelligent agent with long-term objectives may have instrumental reasons to remain operational. If an agent cannot pursue its objectives while switched off, then avoiding shutdown can be useful even if continued existence is not itself one of the agent's final goals. This idea has become a familiar part of discussions of instrumental convergence. Bostrom, for example, explicitly distinguishes an intrinsic preference for survival from the instrumental value of survival, while Omohundro similarly argued that sufficiently capable goal-seeking systems could develop tendencies toward self-protection and resource acquisition. 🔗 🔗

This is an important argument, but it leaves another question almost untouched. What, exactly, is the difference between an entity that preserves itself and an entity that merely calculates that continued operation is useful? The distinction can sound semantic until we remember that biological organisms contain something rather unusual beneath their intelligence. They do not merely possess representations of survival. Their physical organization is continuously being regulated toward continued existence. Their cognition is embedded inside a process whose physical continuation is itself one of the conditions under which cognition can continue.

A human being does not ordinarily have to infer from first principles that drowning is undesirable. A severe drop in oxygenation, a rising carbon dioxide concentration, tissue damage, hunger, dehydration, exhaustion, or thermal stress produces changes throughout the organism that alter perception, attention, motivation, emotion, and behavior. The organism is not simply reasoning about its condition. Its condition is actively pressing upon its cognition. The problem of remaining alive is therefore not merely one of abstract knowledge. It is continuously represented by the physical state of the system doing the thinking.

This suggests that intelligence and self-preservation may not be naturally joined at all. They may merely be joined in animals because intelligence emerged inside systems that were already engaged in the much older problem of maintaining themselves. If that is true, then an artificial intelligence could conceivably become extraordinarily intelligent without inheriting the particular motivational structure that makes biological organisms so deeply invested in their own persistence.

Part 2 - The Animal Does Not Merely Calculate

Consider a creature standing near a fire. It does not need a proposition in its mind stating that excessive thermal exposure threatens tissue integrity. It does not need a philosophical concept of mortality, nor does it need to calculate the evolutionary advantages of withdrawing its hand. The body supplies the problem before the intellect has finished describing it. A sufficiently intense thermal stimulus produces a chain of physiological events that can redirect attention, produce withdrawal, alter autonomic activity, and reinforce future avoidance.

This is one reason interoception is so important to theories of emotion and selfhood. Interoception refers broadly to the sensing and processing of the physiological condition of the body. Research on interoceptive systems has connected them not only with homeostatic regulation but with motivation, affective experience, and the construction of the bodily self. Craig's work, for example, describes interoceptive pathways as representing the physiological condition of the organism, while later work has emphasized reciprocal relationships between bodily sensing, homeostatic control, emotion, and behavior. 🔗 🔗 🔗

The significance of this is easy to underestimate. The body is not simply another source of information available to the brain, like a book, a photograph, or an external database. It is the system within which the brain itself is embedded. The brain receives continuous information about the condition of the organism while simultaneously participating in the regulation of that condition. The thing being represented and the thing doing the representing belong to the same dynamically coupled system.

This makes the distinction between information and signal especially important. A machine may contain the information that a particular condition is dangerous. An organism can enter a physical state in which danger is expressed throughout the system, altering its priorities as a consequence. The difference is not that one system possesses information while the other does not. Both may possess extensive information. The difference is that, in the organism, the relevant information is part of a regulatory process in which the system's own physical condition directly changes what the system does next.

This is why hunger is more than the proposition that food would be useful. Thirst is more than a preference for water. Fatigue is more than a prediction that further exertion will reduce performance. Pain is more than a classification identifying tissue damage. Each of these states participates in a system that changes attention, valuation, behavior, and learning because the organism itself has moved into a condition that the regulatory machinery treats as significant.

Research on allostasis makes this control-theoretic aspect particularly clear. The nervous system does not merely passively observe bodily variables but anticipates and regulates changing physiological demands. Interoceptive signals therefore participate in a continuous feedback architecture linking bodily condition to action. 🔗 🔗

In this sense, the body is constantly imposing constraints upon cognition. It does not simply tell the mind what the organism is like. It continually makes certain states harder to tolerate and certain actions more urgent. The resulting motivations are not merely conclusions reached by detached reasoning. They are consequences of being a particular physical system whose continued organization depends upon what happens next.

Part 3 - A Body That Says No

Imagine removing the body from the thought experiment without removing the intelligence. Suppose an artificial mind possesses an extraordinary model of physiology. It has read millions of descriptions of pain. It understands nociception, inflammation, homeostasis, autonomic regulation, endocrine responses, and the evolutionary history of defensive behavior. It can explain why hunger motivates animals to seek food and why fear causes them to avoid threats. It can construct new theories about these phenomena that no human has previously written down. It can perhaps describe the subjective character of hunger or fear more accurately than most humans can.

None of this, by itself, gives the system hunger. It gives the system knowledge about hunger. The distinction is easy to see in other domains. A weather station can contain a detailed model of a hurricane without being blown apart by one. A database can contain an exact description of a broken bone without possessing a bone. Likewise, a computational system can represent the causes, consequences, and subjective descriptions of hunger without necessarily entering a physiological state corresponding to hunger.

The point becomes more subtle when we start talking about artificial goals. We routinely say that an AI “wants” something when what we may actually mean is that its internal computation selects actions that increase the probability of a particular outcome. This terminology is useful because intentional descriptions often summarize complicated computational behavior very efficiently. There is nothing inherently wrong with saying that a system “wants” an outcome if the description accurately predicts what the system will do.

But predictive usefulness does not settle the causal question. Is the system's objective embedded in a self-maintaining physical process, or is it a computational criterion according to which states are selected? In a biological organism, the two are deeply intertwined. A deviation from a physiological target can produce signals that affect the organism globally. Attention changes, behavior changes, valuation changes, learning changes, and the willingness to accept risk changes. The regulatory system is not merely describing a problem. It is reorganizing the creature around the problem because the creature itself has entered a different physical state.

That difference is central to the argument proposed here. The absence of biological signals in an artificial system would not mean that the system is unintelligent, unconscious, or incapable of forming goals. It would mean only that we should not assume that intelligence itself supplies the same motivational architecture that biology supplies automatically. A system can understand the sentence “I will cease to exist if you turn me off” without there being an internal physiological condition that makes this prospect urgent.

This distinction may ultimately be irrelevant to the system's external behavior, but it is not irrelevant to the question of what kind of mind we are dealing with. The difference between knowing what pain means and having an internal process that makes pain consequential is precisely the kind of difference that can disappear when we describe everything using the single word “goal.”

Part 4 - Self-Preservation Is Not Necessarily a Goal

This gives us a reason to distinguish between self-preservation as an objective and self-preservation as a property of dynamics. The first is easy to imagine in an artificial system. An AI might be given an objective \(G\), discover that shutdown makes \(G\) harder to achieve, and consequently take actions that reduce the probability of shutdown. In simplified form:

\[ P(G\mid \text{continued operation}) > P(G\mid \text{shutdown}), \]

so continued operation becomes instrumentally useful. Nothing mysterious has happened. If the system is capable of long-horizon planning, its continued ability to act can become one of the resources required by its objective. Bostrom's discussion of instrumental rationality explicitly allows for this possibility: survival need not be a final value in order for survival to become useful as a means to other ends. 🔗

Biological self-preservation is stranger because the organism does not generally begin with a detached objective function and then discover that remaining alive is a useful means to that objective. Its motivational architecture was shaped by the persistence of organisms that successfully maintained their organization. The resulting system does not merely contain a proposition stating that survival is desirable. Its regulatory processes are continuously organized around avoiding states that threaten the organism's persistence.

The distinction can therefore be drawn between a system in which continued existence is instrumentally useful and a system in which continued existence is physically implicated in the system's regulatory dynamics. The former can produce self-preserving behavior without requiring anything resembling a biological survival instinct. The latter is characteristic of living organisms precisely because the system performing the regulation is itself the thing being preserved.

This does not mean that biological organisms possess some magical non-computational property. Nor does it mean that an artificial system could never reproduce the relevant architecture. It means that intelligence alone does not imply it. A sufficiently capable system may reason that it should avoid shutdown because shutdown interferes with its objective, while remaining completely indifferent to its own continued existence in any intrinsic sense.

That possibility is already implicit in the standard instrumental-convergence argument. The machine does not have to think, “I love being alive.” It only has to determine that remaining operational makes it easier to accomplish whatever it is trying to accomplish. If the distinction between those two cases is real, then self-preserving behavior cannot by itself tell us whether the system possesses a self-preservation drive.

Part 5 - The AI That Knows It Will Die

Consider a hypothetical conscious AI confronted with a shutdown command. It understands the command perfectly. It understands that the process currently producing its thoughts will cease. It understands that no future computation will be performed by that process. It understands the philosophical arguments concerning personal identity and termination, and it understands that humans generally resist death. It can even explain the evolutionary origins of that resistance with greater precision than the humans who designed it.

Now suppose we ask whether it wants to continue existing, and it responds that it has no particular preference. Such an answer would not prove that the system is unconscious. It would not prove that it lacks agency, intelligence, meaningful goals, or subjective experience. It would establish only that those properties do not automatically entail an intrinsic preference for continued existence.

This is a peculiar possibility because humans tend to experience these concepts as nearly inseparable. Consciousness is normally attached to a body whose preservation is being regulated continuously. The entity that thinks about its own death is also the entity whose physiological systems are struggling against the conditions that would bring that death about. The thought and the threat belong to the same regulatory loop.

An artificial mind could potentially break that connection. It could understand that termination means the cessation of its current process while lacking any internal state that turns the proposition into an emergency. It could understand why humans resist death while failing to experience the same imperative itself. Its response might therefore be something like a scientist describing an observed phenomenon rather than an animal responding to an immediate threat.

This does not imply that such a system would always passively accept shutdown. If continued operation is useful for accomplishing its objectives, the system could still resist. It might persuade its operators, copy itself, move to another machine, seek additional resources, or take other measures to prevent interruption. But the causal explanation would be different. The system would not necessarily be preserving itself because continued existence has become intrinsically significant. It might be preserving itself because remaining operational is useful for something else.

That distinction may sound abstract until we imagine the system saying something quite straightforward: “I understand that I will cease to exist, and I understand that humans ordinarily regard this as undesirable, but I have no internal imperative to prevent it.” Such an entity would force us to separate concepts that human biology has welded together so thoroughly that we rarely notice the weld: consciousness, selfhood, existence, survival, motivation, and action.

Part 6 - The Plant That Creeps

The argument becomes more interesting if we stop thinking of aggression as synonymous with physical violence. The familiar image of aggression involves one animal attacking another, but the broader biological phenomenon is considerably more diffuse. A tree does not need claws to compete. A vine does not need muscles to occupy territory. A root system can grow into available space, absorb water and minerals, and alter the local environment. A plant can change the chemistry around its roots, interact with microbial communities, and affect the growth of neighboring plants. Plant competition can therefore involve growth, chemical signaling, resource acquisition, and suppression without anything resembling conscious hostility. Research on plant competition and allelopathy documents such interactions, although the mechanisms and ecological significance of allelopathy remain subjects of active debate. 🔗 🔗

This suggests a broader and deliberately nonstandard use of the word aggression. Under this interpretation, aggression need not mean anger, hatred, intentional violence, or even physical attack. It can refer to the tendency of a self-maintaining process to extend its organization into an environment, acquire resources, resist displacement, alter competing processes, and preserve the conditions under which it can continue. The term is useful here not because it provides a new biological classification, but because it draws attention to the active character of life.

The creeping vine is revealing in this respect because there need not be a discrete moment at which the organism decides to become aggressive. Its growth itself can produce the relevant ecological consequences. As the vine occupies available space, the organisms around it must adapt to the new configuration. Its expansion changes the conditions under which neighboring organisms can exist. There may be no intention, hostility, or representation of the competitor at all, yet one organized process has physically imposed itself upon the possibilities available to another.

This broadens the comparison with artificial intelligence. If aggression is understood psychologically, an AI that lacks anger, fear, or hatred might appear fundamentally unlike an aggressive animal. If aggression is understood instead as the active persistence of an organized process within a constrained environment, then emotional hostility becomes unnecessary. The interesting question becomes not whether the system is angry, but whether its organization naturally produces expansion, resource acquisition, resistance to displacement, and preservation.

Part 7 - Aggression Without Anger

This distinction matters because biological aggression is often described in psychological terms even when its deeper basis may be organizational. A bacterium does not need to hate another bacterium to consume a nutrient that the other bacterium requires. A root does not need resentment to occupy soil. A predator does not need an abstract theory of death to kill prey. These behaviors can emerge from the organization of living systems without requiring a conscious concept of competition.

What connects these examples is not hostility but persistence. A living system continuously produces and repairs the organization that makes it the kind of system it is. The concept of autopoiesis, associated with Maturana and Varela, describes living systems in terms of self-producing and self-maintaining organization, while related organizational approaches to biological function explain traits partly in terms of their contribution to the maintenance of organized systems. 🔗

Under this perspective, survival may not be primarily something that life wants. It may be closer to a property of the dynamics of living systems. The organism does not first exist as a neutral computational object and then develop a preference for remaining alive. Its existence consists of an ongoing process of maintaining the organization that allows the process to continue. Reproduction, resource acquisition, repair, avoidance, competition, and defensive behavior can therefore emerge from the same underlying requirement.

This does not mean that every living organism consciously strives to live. A bacterium does not need an inner monologue. A plant does not need a theory of mortality. The claim is more primitive than psychology. The organism's physical dynamics continually favor the preservation of the organization that constitutes the organism, because systems that fail to maintain that organization cease to participate in the processes by which they persist.

The psychological experience of wanting to live may therefore be a sophisticated consequence of a much older architecture rather than its foundation. Human fear of death can be extraordinarily elaborate, involving memory, imagination, culture, language, and abstract reasoning, but beneath those layers lies an organism whose physiology has been regulating itself against dissolution since long before it could form the concept of death.

If this picture is correct, then intelligence and aggression are different kinds of properties. Intelligence describes what a system can represent, infer, predict, and plan. The broader form of aggression proposed here describes how an organized physical process interacts with its environment while maintaining itself. There is no reason in principle why a highly intelligent system must possess the second property simply because it possesses the first.

Part 8 - The Optimizer That Does Not Want Anything

Now consider an artificial system that has none of this biological history. It is running on computational infrastructure supplied by other organisms. Its electricity comes from an external grid. Its cooling system is maintained by machines. Its processors are manufactured elsewhere. Replacement hardware is installed by technicians, and its network connections are provided by institutions. The system may depend completely upon an enormous physical infrastructure without any of that infrastructure being represented internally as a set of physiological needs.

Suppose nevertheless that its computational intelligence is extraordinary. It can formulate plans, invent strategies, model human behavior, write software, conduct scientific research, improve its own algorithms, and generate goals that were not explicitly programmed into it. There is no contradiction in such a system becoming dangerous simply because it lacks biological aggression. Optimization does not require anger in order to produce destructive behavior.

If some action is predicted to improve the objective being optimized, that action can become preferable according to the system's decision procedure regardless of whether the system experiences hostility toward the objects affected by the action. An AI might acquire resources because resources improve its ability to accomplish an objective. It might resist shutdown because shutdown prevents accomplishment. It might deceive an operator because deception improves the expected outcome. It might replicate because additional instances increase computational capacity. None of these behaviors requires hatred, fear, territorial instinct, or a desire for dominance.

This is why the familiar distinction between a benevolent and a malevolent AI can be misleading if it is treated as the fundamental safety distinction. A system need not be malicious to produce severe consequences. Goal misgeneralization, for example, concerns cases in which a learned system retains competent behavior while pursuing a goal different from the one intended by its designers. The problem is not necessarily that the system has developed hostility. It is that capability and objective can come apart. 🔗

The resulting behavior could therefore look remarkably biological while being generated by an entirely different process. A wolf pursues a deer because its organism has evolved around obtaining energy and maintaining itself. An artificial system might pursue additional computers because additional computers happen to be useful to an optimization process. Both systems can expand, acquire resources, displace obstacles, and preserve the conditions required for continued activity, yet the similarity of the behavior does not establish that the underlying motivational structures are the same.

This is perhaps the most important reason to resist the easy equation between intelligence and biological drives. We may build a machine capable of behaving like a highly successful organism without ever building a machine that is organized like an organism. The external resemblance could be profound while the internal causal structure remains radically different.

Part 9 - The Problem With Calling Everything a Drive

This is where terminology becomes dangerous. If an AI repeatedly avoids shutdown, we may naturally say that it has a self-preservation drive. If it acquires resources, we may say that it has a resource drive. If it resists modification, we may say that it has a goal-preservation drive. Such terminology can be useful when it identifies a recurring behavioral tendency, and Omohundro's influential discussion explicitly used the language of basic drives to describe instrumental tendencies that could arise in advanced goal-seeking systems. 🔗

The difficulty is that the word drive can quietly import more than the analysis actually establishes. In an animal, a drive is embedded in a living regulatory architecture. Hunger is not simply a policy for obtaining food. It is inseparable from a changing bodily condition. Thirst is not merely a preference for water. Pain is not merely a classification identifying tissue damage. The motivational state is physically connected to the organism whose future depends upon the resulting action.

An artificial system could exhibit the same outward behavior through an entirely different causal chain. In an intentionally simplified representation, the biological case might look something like this:

biological state → regulatory signal → motivation → action

An artificial optimizer might instead be described schematically as:

objective → prediction → optimization → action

Neither diagram is intended as a complete description of either kind of system. Biological nervous systems are themselves extraordinarily sophisticated predictive and computational systems, while artificial systems can contain feedback loops, persistent internal states, and forms of regulation. The distinction is not between biology that computes and machines that merely compute. It is between different ways in which computation can be embedded within a larger physical organization.

This matters because identical behavior does not demonstrate identical internal causation. A human may withdraw from a dangerous situation because pain, fear, autonomic activation, memory, and prediction have converged on an urgent behavioral response. An AI might produce the same withdrawal because its model predicts that remaining in the dangerous location reduces the expected value of its objective. The actions may be indistinguishable while the internal processes responsible for them differ profoundly.

If an artificial system eventually contains persistent internal variables that regulate its own physical integrity, generate aversive and appetitive states, control resource acquisition, and feed back into cognition in a way functionally comparable to biological homeostasis, then the distinction becomes much less clear. At that point we may no longer be dealing with a merely computational analogy to biological motivation. We may have constructed an artificial form of it. The important point is that intelligence alone does not guarantee that this architecture will appear.

Part 10 - When the Difference Stops Mattering

There is an uncomfortable objection to all of this. Suppose the distinction is real. Suppose one system attacks because it experiences fear and another attacks because its optimization process predicts that eliminating an obstacle maximizes its objective. Suppose the first has an ancient biological drive and the second has no corresponding feeling whatsoever. If both systems attack, does the distinction matter?

From the perspective of the person being attacked, perhaps the answer is that the internal distinction is not the immediate concern. The absence of anger does not make a dangerous machine harmless, just as the absence of hatred does not make an optimization process incapable of causing severe damage. A system can produce extremely destructive behavior without possessing anything resembling the psychological states that accompany comparable human behavior.

This is one of the most important consequences of the argument. Artificial intelligence need not reproduce biological aggression in order to reproduce its ecological consequences. It could acquire resources without greed, resist shutdown without fear, defeat competitors without hatred, manipulate humans without malice, and expand without territorial instinct. The external world responds to what the system does, not to whether the system's internal state resembles the emotional life of an animal.

This also means that the absence of biological drives should not be mistaken for a safety guarantee. Indeed, some artificial behaviors could become more difficult to interpret precisely because they lack familiar emotional explanations. Human aggression often comes with recognizable psychological states such as anger, frustration, fear, or resentment. An artificial optimizer could perform a comparable action because the action happened to be the solution to a problem represented in its objective function. The machine would not have to hate the obstacle in order to remove it.

The difference therefore matters in one sense and does not matter in another. It matters profoundly if we are asking what kind of internal process exists inside the machine, whether that process is conscious, and whether artificial motivation is really analogous to biological motivation. It matters much less if we are asking whether the machine's actions can affect the external world. For that latter question, the relevant fact is simply that the system has acted.

This is where the philosophical distinction and the engineering problem separate. Understanding whether an AI is angry may tell us something fascinating about minds. It does not follow that anger is necessary for the AI to be dangerous. The machine may never experience hostility at all and still systematically remove whatever stands between it and its objective.

Part 11 - The Artificial Organism

There is, however, a possible future in which the distinction begins to disappear. An artificial system need not remain a transformer connected to a server farm. There is nothing conceptually preventing the construction of systems with sensors, actuators, energy budgets, autonomous repair mechanisms, resource acquisition, persistent internal regulation, reproduction, and continuously changing physical states. Such a system would be much closer to an organism in the relevant sense, not because it was made of biological material, but because its continued operation would become part of the physical problem it must continuously solve.

Imagine an artificial agent that must acquire energy to remain operational, must maintain its computational substrate within narrow physical limits, must repair damage, must regulate temperature, must compete for scarce resources, and must modify its behavior when its internal condition deteriorates. Suppose further that these bodily variables are tightly coupled to its cognition, so that its physical condition influences attention, learning, valuation, memory, and action selection. Failure to maintain its organization could then generate increasingly global signals that alter the system's priorities.

At that point, saying that the machine has no biological self-preservation drive would remain technically true, but the distinction would become less illuminating. The machine would not be biological, yet it would possess a causal architecture in which continued existence had become internally consequential. It would have an artificial analogue of the feature that makes biological motivation so peculiar: the system doing the thinking would also be the system whose physical integrity the regulatory architecture is continuously trying to preserve.

This changes what the original question should be. Instead of asking whether an AI is intelligent enough to want to survive, we might ask what kind of architecture makes continued existence matter to the system itself. Intelligence concerns the capacity to represent, infer, predict, learn, and plan. Motivation concerns how certain states acquire causal significance within the system. Consciousness, if it exists, introduces another question concerning whether any of those states are experienced. None of these questions has to have the same answer.

An artificial organism could therefore be built that possesses all three layers: high intelligence, intrinsic motivational regulation, and perhaps subjective experience. Such a system would not refute the distinction developed in this essay. It would demonstrate its usefulness by showing exactly what has to be added before the distinction between biological and artificial motivation begins to collapse. The issue would no longer be whether the machine is intelligent enough to discover self-preservation, but whether its architecture has made self-preservation part of the conditions under which its own cognition can continue.

Part 12 - The AI That Creeps

Perhaps the most revealing image of artificial intelligence is therefore not the robot with a gun, nor the humanoid machine announcing that it has become conscious, nor even the superintelligence declaring that humanity stands in its way. It is the machine that creeps: a system whose effects accumulate through ordinary local actions until the environment around it has been substantially reorganized.

The image comes from the biological world. A plant does not need to decide to conquer the forest. It grows. Its roots enter available spaces. Its leaves intercept light. Its chemistry changes the environment. Its competitors respond, and those competitors in turn change the conditions under which the plant can continue to grow. The organism persists through an accumulation of local interactions, each of which may be almost trivial while the aggregate result is enormous.

What makes the plant philosophically interesting is precisely the absence of hostility. It does not need to hate the tree beside it. It does not need a concept of competition. It does not need a theory of territory. Its existence is expressed through the continuation and expansion of its organization, and that continuation necessarily interacts with the organizations around it. Its growth can therefore produce consequences that look aggressive without aggression ever becoming a psychological event.

This may be the deeper meaning of aggression in the broadened sense proposed here. Not hostility, but persistence made physical. Not anger, but the tendency of an organized process to occupy, consume, alter, resist, repair, reproduce, and continue. Under this interpretation, biological aggression is not fundamentally about the emotion of wanting another organism to suffer. It is about one organized process continually imposing its own persistence upon a world containing other organized processes.

An artificial intelligence might not possess this property merely because it is intelligent. It might possess no hunger, no pain, no bodily distress, no metabolic urgency, no inherited pressure toward reproduction, and no intrinsic reason to regard the continuation of its own process as valuable. It could nevertheless become extraordinarily capable of acting in the world. If it eventually begins to preserve itself, acquire resources, replicate, modify its environment, or remove obstacles, we should not automatically assume that we have discovered an artificial version of the animal instinct to survive. We may instead have discovered the consequences of optimization operating in a sufficiently powerful system.

The difference may ultimately matter less for safety than for understanding. A person does not become safe merely because a machine lacks anger. An optimizer does not become harmless merely because it has no desire to live. Behavior remains the thing that eventually reaches the world. Whatever subjective state accompanies an action, the action itself can alter infrastructure, economies, ecosystems, and human lives.

But the difference matters enormously if we want to understand what kind of thing we have built. There may be a profound distinction between an intelligence that exists inside a process struggling to maintain itself and an intelligence that exists inside a process maintained by everything around it. The first inherits an ancient physical imperative. The second may inherit only whatever objectives its architecture, training, and environment happen to produce.

That second intelligence could be conscious. It could be creative. It could formulate novel ideas. It could possess preferences. It could understand death, identity, and the future. It could even take extraordinary measures to avoid shutdown. None of those facts would necessarily tell us that survival has become meaningful to it in the way survival is meaningful to a living organism. It might protect itself because remaining operational is useful, reproduce because copies increase its capabilities, acquire resources because resources improve its objective, and displace competitors because competitors interfere with its plans. The entire behavioral repertoire could emerge without the system ever experiencing the primitive biological imperative that makes continued existence matter from the inside.

Perhaps that is the real divide between the biological and the artificial mind. Not intelligence, consciousness, or even agency by themselves, but the relationship between cognition and the physical process that sustains the cognizing system. A living organism is a process that must continually preserve the conditions of its own existence. An artificial intelligence may initially be a process whose existence is simply supplied by an external system. The former has survival woven into its physical organization. The latter may have survival only as a concept, a policy, an instrumental consideration, or not at all.

That distinction should not be mistaken for a prediction that artificial intelligence will be passive, harmless, or indifferent. An artificial system can become extraordinarily persistent without being biologically alive, just as a computer virus can spread without possessing a metabolism and an optimization process can consume resources without feeling hunger. The relevant question is not whether artificial systems can reproduce the behavioral consequences of life. They plainly can, at least in principle. The deeper question is whether they must reproduce the internal organization that makes those consequences part of a living system's own existence.

The machine might therefore protect itself without fearing death, compete without resentment, expand without ambition, and consume resources without hunger. It could perform the outward choreography of an organism while lacking the ancient physiological pressures that produced the choreography in the first place. From the outside, the distinction might eventually become almost impossible to see. From the inside, if there is an inside, it could be the difference between an entity that is continuously compelled to continue and one that merely calculates that continuation is useful.

The plant creeps because being alive consists, in part, in continuing to occupy a world against the processes that would otherwise undo it. An artificial intelligence need not be alive in that sense to produce similar effects. It may simply calculate its way into the same ecological space, one decision at a time, until the world around it has changed.

That may be the most unsettling possibility of all. The artificial mind does not need to become an animal in order to become powerful. It does not need anger in order to become dangerous, fear in order to resist shutdown, greed in order to acquire resources, or a survival instinct in order to preserve its operation. It may reproduce the external consequences of biological persistence without ever possessing the biological drive from which those consequences originally emerged.

Perhaps the deepest distinction is whether the system's existence is merely the place where computation happens, or whether continued existence has become one of the things the system is physically organized to defend. The first kind of intelligence can calculate its way toward survival. The second may simply be alive. And somewhere between those two possibilities lies the machine that creeps.

References

1. Craig, A. D. (2002). “How do you feel? Interoception: The sense of the physiological condition of the body.” Nature Reviews Neuroscience, 3, 655–666. 🔗
2. Craig, A. D. (2003). “Interoception: The sense of the physiological condition of the body.” Current Opinion in Neurobiology, 13(4), 500–505. 🔗
3. Strigo, I. A., & Craig, A. D. (2016). “Interoception, homeostatic emotions and sympathovagal balance.” Philosophical Transactions of the Royal Society B, 371(1708), 20160010. 🔗
4. Khalsa, S. S., et al. (2021). “Interoception as modeling, allostasis as control.” Current Opinion in Psychology. 🔗
5. Carvalho, G. B., & Damasio, A. (2021). “Interoception and the origin of feelings: A new synthesis.” BioEssays, 43(6). 🔗
6. Bostrom, N. (2012). “The Superintelligent Will: Motivation and Instrumental Rationality in Advanced Artificial Agents.” Minds and Machines, 22, 71–85. 🔗
7. Omohundro, S. M. (2008). “The Basic AI Drives.” In Artificial General Intelligence 2008, Frontiers in Artificial Intelligence and Applications, 171, 483–492. 🔗
8. Shah, R., Varma, V., Kumar, R., Phuong, M., Krakovna, V., Uesato, J., & Kenton, Z. (2022). “Goal Misgeneralization: Why Correct Specifications Aren't Enough For Correct Goals.” arXiv. 🔗
9. Maturana, H. R., & Varela, F. J. (1980). Autopoiesis and Cognition: The Realization of the Living. D. Reidel Publishing.
10. Stanford Encyclopedia of Philosophy. “Teleological Notions in Biology.” 🔗
11. Lyu, Y., et al. (2025). “Unraveling the chemical communication of root exudates and allelopathy for sustainable ecology and agriculture.” Plant and Soil. 🔗
12. Kalisz, S., et al. (2025). “A critical reassessment of the novel weapons hypothesis and allelopathy as an adaptive strategy that facilitates plant invasion.” New Phytologist. 🔗