The Evolutionary Foundations of Mind Perception

Whenever I scroll through social media, read the news, or watch YouTube, I run into discussions of Artificial Intelligence (AI), AI agents, and Artificial General Intelligence (AGI). Many describe technologies that could reshape society, or threaten it if their goals conflict with ours or people misuse them. These warnings led me to a simpler question: why are we so afraid of a computer program we built, especially one that still cannot match the full range of human intelligence and agency?

I realized that our reaction goes beyond how we respond to ordinary software, or to the industrial robots that have worked in factories for decades. It involves mind perception: our capacity to perceive consciousness in other entities and attribute to them an inner mental life, subjective experience, and intent. Millions of years of natural selection have shaped this capacity, and it does not passively record what we observe. It actively interprets it.

A central argument in the evolutionary literature is that mind perception evolved to help us survive, not to establish philosophical truth. Error Management Theory holds that, in uncertain environments, failing to detect an agent costs far more than mistakenly detecting one. This asymmetry favored the evolution of a Hyperactive Agency Detection Device and a teleological stance, which lead us to detect agents too readily and to see purpose everywhere. We attribute minds to wind-blown bushes, to moving geometric shapes, and, increasingly, to algorithms.

But we do not judge every possible agent in the same way. Once we detect one, we assess it by its evolutionary relatedness to us, or phylogenetic proximity, and through the moral dyad of agency and experience. We feel deep empathy for mammals that resemble us and show familiar social and emotional cues. Yet we struggle to recognize suffering in invertebrates, or in entities that fall into the Uncanny Valley, where ambiguity can trigger ancient avoidance responses.

Rapid advances in AI and robotics now engage, and sometimes manipulate, these Pleistocene adaptations for social cognition. Psychological mechanisms that evolved to help us manage complex human relationships now respond to artificial entities.

One example is an amplified form of the ELIZA effect, our unconscious tendency to attribute human-like traits and intelligence to computers. Modern Large Language Models (LLMs) produce fluent, human-like language that carries the very cues our brains associate with agency, consciousness, and intent. In doing so, they exploit our evolved tendency to infer a mind from a voice.

Engineers are also increasingly designing machines to mimic signs of social and emotional vulnerability, such as pain, distress, and neediness. This deliberate, and ethically troubling, approach aims to make users feel protective, caring, or empathetic. It bypasses reasoned judgment and activates our caregiving instincts, the same responses that make us comfort an injured child.

Together, these trends have fostered a kind of digital animism. People form deep emotional bonds with AI companions and attribute spirit, life, and subjective experience to them. At the same time, the digital platforms that host these bonds often encourage people to dehumanize one another. To understand why we extend empathy to machines while withholding it from people, we need to examine the biology and evolutionary origins of mind perception.

To understand the biases in mind perception, we first need to consider why consciousness evolved at all. The social origins hypothesis challenges the prevailing view that consciousness mainly regulates internal bodily states and homeostasis. Andrews and Miller [1] argue instead that consciousness may first have evolved to help organisms coordinate with other members of their group as social life grew more complex.

Under this hypothesis, the earliest form of consciousness was sentience, the capacity for pleasant and unpleasant experiences such as pleasure and pain, rather than reflective self-awareness or inner speech. Feelings evolved because they change behavior, and in this account they first responded to social conditions. Animals felt bad when they were distant from, or out of sync with, their social partners, and good when they were close to the group or coordinating successfully. Consciousness made cohesion rewarding and separation costly.

The authors also argue that consciousness arose early in animal evolution rather than appearing late in large primate brains. It may date back as far as the Cambrian explosion, when animals became far more physiologically flexible. Muscles enabled faster movement, senses expanded to detect things at a distance, and neural coordination grew more complex. These changes made behavior less predictable, especially for animals that depended on group living. Natural selection therefore favored organisms that could both navigate their physical environment and keep their groups together despite uncertainty about what others would do.

The Cambrian explosion (Source: Dinghua Yang)

This reverses a familiar story in psychology, in which consciousness first supported individual learning and only later social cognition, with abilities such as theory of mind treated as prerequisites for genuine social life. The social origins hypothesis proposes the opposite order: consciousness first improved predictions of others’ behavior and only later helped organisms understand themselves. Nor does it require reflective mindreading. An organism only needs to treat others’ behavior as relevant, and to use feelings that give social cues enough weight to direct its attention, learning, and action toward coordination rather than dispersal.

For such an early origin to be biologically plausible, relatively simple neural systems must be able to support affective consciousness. The authors describe awareness in functional terms, as a feedback process in which information that drives action returns to the system to guide flexible responses. One possible ancestral mechanism is reafference, in which the system reprocesses its own motor signals as sensory data to distinguish the changes it causes from those that come from outside. The authors point to evidence of feedback-like organization even in simple living animals such as sponges and comb jellies. On this view, sentience can emerge from ancient control loops and evaluative systems, which evolution later expanded and repurposed, without depending on any single high-level brain structure.

The same argument applies to social pain. If consciousness helped organisms coordinate, evolution could have attached social meaning to ancient systems for distress and relief. The authors cite evidence that pain-processing and social systems remain closely linked in modern brains. Opioid regulation affects both pain and social behavior, and social rejection and empathic pain can engage the same mechanisms as emotional distress. Oxytocin-related mechanisms also make social cues more salient and support social buffering, the reduction in stress and threat responses when social partners are present.

Neuroimaging studies of social exclusion have found increased activity in the dorsal anterior cingulate cortex and involvement of the anterior insula, suggesting that rejection can engage parts of the brain’s affective pain system. Alongside the social origins hypothesis, these findings suggest that social pain is a biological signal that helps maintain social bonds. For an organism that depends on others, isolation can threaten evolutionary fitness as much as physical injury, so reusing the brain’s most urgent alarm systems would help keep it close to the group.

The authors connect this account to William James’s argument against epiphenomenalism, the view that conscious experience has no effect on behavior, through the idea of deep adaptive alignment. If pleasure and pain reliably track benefit and harm, they guide adaptive choices rather than merely accompanying them. Andrews and Miller ask us to take social gains and losses as seriously as bodily ones. If consciousness evolved partly to hold groups together despite unpredictable behavior, social harm should rank among the most unpleasant experiences, and organisms should seek social contact even at considerable physical cost.

Evidence across species fits this expectation. Preference tests often show that social contact strongly motivates animals and that they find isolation deeply aversive. The authors discuss infant monkeys that preferred a soft cloth surrogate mother to a wire one that provided milk, suggesting that comfort and contact can matter more than food. They also review animals that sacrifice comfort or risk pain to reach others, including trout that tolerate increasingly strong electric shocks to approach a social partner. These findings support the view that social life is a deep need, and that consciousness regulates it through felt rewards and punishments.

This evolutionary account also challenges the zombie hypothesis, the philosophical idea that a creature could behave exactly like a human without any subjective experience. From an evolutionary perspective, that separation is hard to sustain. If consciousness helps organisms make complex trade-offs and track social relationships, it must affect observable behavior, and a perfect behavioral duplicate without experience sits awkwardly with a trait that natural selection shaped for its effects.

Studies of how consciousness returns in stages during recovery from general anesthesia also suggest that the brain structures supporting it are ancient and highly conserved across vertebrates. Mammals and birds share the brainstem and limbic structures that generate core affect, and these arguably extend to reptiles. Attributing feelings to other animals can therefore reflect recognition of a shared biology rather than mere projection of human traits. The bias becomes unreliable mainly when an entity mimics the behavior associated with feelings without sharing that biology, a distinction our ancestors rarely needed to make.

Error Management Theory

If social life shaped cognition, detecting other agents is a basic survival skill. Missing a predator, a potential mate, or a rival can have severe consequences, while mistaking wind, shadows, or rustling leaves for an agent usually costs only a little attention and energy. Error Management Theory (EMT) explains what follows from this difference: when mistakes have unequal costs, natural selection tends to favor the cheaper error.

A judgment made under uncertainty can produce a Type I error (a false positive) or a Type II error (a false negative). Reducing one usually makes the other more likely, so no decision-maker can minimize both at once. Agency detection makes this trade-off especially clear, because the costs differ so sharply.

Scenario	Inference	Assumption	Ground Truth	Result	Evolutionary Cost
Predator detection	Type I	It is a tiger	It is wind	False alarm	Minor caloric waste
	Type II	It is wind	It is a tiger	Death	Termination of lineage
Mating opportunity	Type I	She is interested	She is not	Rejection	Wasted effort
	Type II	She is not interested	She is	Missed chance	Loss of reproductive fitness
Agent detection	Type I	A spirit did this	It was random	Superstition	Ritual cost
	Type II	This was random	It was an enemy	Ambush	Injury

This asymmetry explains why our minds constantly overestimate the presence of agents. Cognitive scientists call the mechanism responsible the Hyperactive Agency Detection Device, or HADD. Its default is to detect presence and intent, even where neither exists.

The urge to detect agents likely contributed to early religious thought. In Upper Paleolithic shamanistic practice, the line between human and non-human agents was fluid. Cave art depicting therianthropes, figures that are part human and part animal, suggests that our ancestors attributed complex mental states and intentions to the natural world.

Earliest depiction of a therianthrope, showing a human figure with a tail, created 43,900 years ago (Source: Ratno Sardi)

Over time, this grew into the worship of nature deities, particularly during the Bronze and Iron Ages, when people personified natural elements as gods. The Egyptian Geb represented the earth, the Mesopotamian Enki water, the Vedic Agni fire, the Chinese Feng Bo wind, and the Greek Aether the upper sky.

The Babylonian flood story, the Atrahasis, is among the earliest evidence of people attributing natural disasters to deities. Enlil, the Mesopotamian god of wind, air, and storms, grows restless and declares, “The noise of mankind has become too much. I am losing sleep over their racket” [2]. He sends a flood to destroy every living thing on earth, but Enki, the god of wisdom and fresh subterranean waters, warns Atrahasis in a dream and tells him how to survive. The epic, which draws on older Sumerian flood traditions, influenced later retellings, including the Biblical story of Noah’s Ark. Other cultures have close parallels. In the Shatapatha Brahmana, a Vedic commentary on the Shukla Yajurveda, a fish later identified with Vishnu’s avatar Matsya warns Manu of a flood on a similar scale.

Through the lens of Error Management Theory, early humans found it safer to treat a violent storm or a rushing river as a temperamental being to be appeased than as a random physical event. Offering a sacrifice to a non-existent river spirit cost far less than underestimating a dangerous current.

These tendencies remain with us. We see faces in clouds, a phenomenon called pareidolia, hear voices in white noise, and attribute malice to objects that fail us. Such reactions reflect a survival bias, not a cognitive defect. Shouting at a malfunctioning computer costs little compared with an ancestor ignoring a snapped twig in a jungle full of predators. Our brains favor false positives, even if that means imagining a ghost to avoid missing a murderer.

Hyperactive Agency Detection

Cognitive scientist Justin Barrett introduced the Hyperactive Agency Detection Device to describe the mental processes that lead us to detect agents from very little sensory information. These processes respond to possible signs of intent, such as an object that moves against inertia or seems to pursue a goal. Other cues include bilateral symmetry, face-like patterns, unexpected rhythms, and distinct noises in the environment.

The result is an immediate sense that a specific agent is present, rather than a vague awareness of some object. The system’s very low activation threshold serves an evolutionary purpose: it helps the organism avoid the severe cost of missing an agent.

Although HADD has long been central to the Cognitive Science of Religion, empirical challenges call for a more careful account. Critiques and replication attempts [3] have often failed to find a direct causal link between simple perceptual agency detection and complex supernatural beliefs. Hearing an ambiguous sound or seeing a face in a cloud does not automatically make someone believe in ghosts or gods.

As a result, many researchers now treat HADD as a predictive processing bias, or a Bayesian prior, rather than a fixed reflex. Bayes’ theorem describes how we update beliefs in light of new evidence:

P(\text{Agency} \mid \text{Sensory Input}) = \frac{P(\text{Sensory Input} \mid \text{Agency}) \cdot P(\text{Agency})}{P(\text{Sensory Input})}

In this equation, $P(\text{Agency})$ is the prior probability: our expectation that an agent is present before we process any sensory data. Evolution has effectively set this value high. With a high prior, even weak or ambiguous evidence, represented by the likelihood $P(\text{Sensory Input} \mid \text{Agency})$ , produces a high posterior probability, $P(\text{Agency} \mid \text{Sensory Input})$ . Because our brains are primed to expect agents, it takes very little evidence to convince us that one is there.

In this account, the brain quickly predicts that an agent is present and then tests that prediction against its surroundings. Learned cultural knowledge shapes how we interpret the initial biological response. A sudden rustle in the woods may trigger the arousal associated with HADD, but whether someone concludes it was a bear, a spirit, or nothing at all depends on their cultural understanding of what exists. Biology drives the search for an agent; culture supplies its identity.

This distinction matters for how we interact with AI. We do not usually mistake a chatbot for a human being, but HADD still registers its language as a sign of agency, because language is one of the strongest cues to intent we know. Culture then supplies the explanation: algorithms, sentience, or superintelligence. Biology makes us feel that a mind is present, and culture gives us the vocabulary to call it AI.

The Intentional Stance and Teleology

The Intentional Stance, introduced by the philosopher Daniel Dennett, is closely related to hyperactive agency detection. It describes how we predict a complex entity’s behavior by treating it as a rational agent with beliefs, desires, and intentions.

Dennett describes three ways of predicting behavior, and we switch among them instinctively depending on an object’s complexity. The Physical Stance relies on physics and chemistry: a dropped stone falls because of gravity, and water freezes as the temperature drops. For more complex objects, the Design Stance relies on their assumed function, so we expect an alarm clock to ring at its set time and a car engine to start when we turn the key.

The Intentional Stance is the most abstract of the three. It predicts behavior by attributing mental states to an entity, and it likely became our default because it is the most efficient way to model complex systems, whether or not they are conscious. We could never predict a chess computer’s next move by tracing electrons through its circuits or reading its source code. But assuming that it knows the rules and wants to win makes the task manageable, letting us focus on strategy instead of the underlying machinery.

We extend similar reasoning to nature, a tendency known as promiscuous teleology. Research shows that people readily take a design stance toward natural objects, viewing them as having a purpose, such as eyes existing so we can see or the ozone layer existing to protect life [4]. This bias makes us especially susceptible to anthropomorphism in robotics and AI. We assume that a robot turning its sensors toward us intends to look at us, or that a system pausing before it speaks is thinking, because our brains infer purpose from outward form.

The Moral Dyad

Once we detect an agent, we assess what kind of mind it has. This judgment has several dimensions rather than a simple yes-or-no answer. One leading framework, the Moral Dyad of psychologists Kurt Gray and Daniel Wegner, explains how we assign moral weight to different beings.

The theory divides mind perception into two independent dimensions: Agency and Experience. Agency is the capacity to act, and includes self-control, planning, memory, and thought. We see beings high in agency as Moral Agents, responsible for their actions and deserving of blame or praise. Experience is the capacity to feel hunger, fear, pain, and pleasure. We see beings high in experience as Moral Patients, capable of suffering and therefore deserving of rights and protection.

Entity Type	Perceived Agency	Perceived Experience	Moral Status
Adult human	High	High	Full moral status
Puppy	Low	High	Moral patient
God	High	Low	Moral agent
AI	High	Low	Tool
PVS patient	Low	Low	Ambiguous

The Moral Dyad proposes that we understand morality mainly through a template of intentional harm, in which a Moral Agent causes a Moral Patient to suffer. This template is so powerful that we engage in dyadic completion. When we see a victim suffering, we instinctively look for an agent to blame, and if no natural agent is visible, we may attribute the harm to a conspiracy, a curse, or a divine being. The same template helps explain both the excess agency we see in conspiracy theories and supernatural beliefs, and the lack of moral standing we grant to artificial intelligence.

We usually see robots as high in agency, because they perform tasks, but low in experience, because they cannot feel. Since we do not classify them as Moral Patients, our intuitions do not treat harming them as wrong, even when they behave like humans. Yet a robot that mimics pain, such as a mechanical dog that whimpers, raises our perception of its capacity for experience. It pushes us to reclassify the robot as a patient, and that creates a strong moral conflict.

The Perception-Action Model

HADD and the Moral Dyad help explain how we perceive minds, but our evolutionary relatedness to another being strongly shapes how much empathy we feel for it. The Perception-Action Model (PAM) proposes that empathy depends on perceived similarity in physical form, behavior, and psychology. In this model, how much we feel for another being depends on how much of ourselves we recognize in it.

Studies have found a strong relationship between evolutionary distance, empathy, and the attribution of consciousness. One study asked more than three hundred people, from young children to adults, to make empathic choices among photographs of a broad range of organisms [5]. It found a clear, linear decline in empathy as evolutionary distance from humans increased.

The tree of life (Source: Leonard Eisenberg)

Mammals consistently elicit the most empathy. Their forward-facing eyes, fur, nursing behavior, and distress calls that resemble those of human infants give us familiar cues. Birds elicit moderate empathy despite being more distantly related, because bipedal walking and complex vocalizations help us relate to them. Reptiles and fish elicit little empathy: their cold-bloodedness, limited facial movement, and unfamiliar ways of moving do little to engage the mirroring responses that underlie empathy. Invertebrates elicit the least empathy of all and often prompt disgust or indifference.

This bias appears to grow stronger with age. Younger children distribute empathy more broadly and evenly, but that range narrows as they absorb cultural norms and form fixed categories. The study found that evolutionary distance generally had a stronger effect on neurotypical adults than on children. Autistic adults, though often described in the literature as experiencing “mindblindness,” showed patterns of empathy toward non-human animals similar to those of neurotypical adults. The differences were more pronounced when they assessed humans.

The number of conservation projects and funding has remained historically low for plants, reptiles, and fish (Source: Guénard et al., 2025)

This bias toward mammals has major consequences for conservation and animal welfare [6]. We favor charismatic megafauna such as pandas and elephants because they activate our tendency to see ourselves in other beings, while we often ignore ecologically vital invertebrates. This is a feature of our cognition rather than a moral failing. Our empathy evolved mainly to bond us with kin and members of our group, and it extends to other species only as far as they trigger those ancient signals of kinship.

The Uncanny Valley

The Uncanny Valley marks a limit to our tendency to see human traits in other entities. Roboticist Masahiro Mori proposed the idea in 1970 to describe the sharp drop in affinity we feel when something artificial looks almost, but not quite, human. A growing body of research suggests that this response serves an evolutionary function beyond aesthetic preference: it produces protective false alarms, triggering rejection when visual cues are ambiguous.

One of the best-supported explanations is the Pathogen Avoidance Hypothesis, which links revulsion toward uncanny entities to the behavioral immune system. In ancestral environments, asymmetry, unusual skin texture, erratic movement, or a lack of emotional expression in another person often signaled diseases such as leprosy or smallpox, or possible genetic problems. In this account, these features trigger disgust to prevent physical contact.

A virtual reality study offered direct physiological support by measuring salivary secretory immunoglobulin A (sIgA), a key marker of mucosal immune activation [7]. Participants interacted with cartoonish agents, realistic human agents, and uncanny agents that looked realistic but deviated in subtle ways. Only the uncanny agents prompted a significant increase in sIgA release, implying that visual irregularities alone can make the body raise its immune defenses. The response fits Error Management Theory: because contact with a diseased member of one’s own species could be fatal, rejecting an ambiguous case is the less risky error.

Reactions in other animals further support a biological explanation. Cats can respond to robotic cats with extreme caution or aggression; these robots look like cats but lack a real animal’s fluid motion and scent. The response resembles human rejection of uncanny entities, suggesting a shared evolutionary mechanism for detecting impostors and biological anomalies that protects animals from threats resembling their own kind.

A related hypothesis connects the uncanny valley to mortality salience. It proposes that human-like entities lacking signs of life trigger an innate fear of the dead. A corpse has a human form and poses a serious pathogen risk, so seeing a lifelike figure with signs of death, such as stillness, coldness, and pallor, creates the conflict that produces the uncanny feeling. This is the logic of the zombie archetype: a being that moves without a soul or mind. The response warns us away from something that looks human but is not viable.

Robot faces ordered along the dimension of human likeness (Source: Reuten et al., 2018 [8])

Categorization ambiguity, or realism inconsistency, offers a cognitive explanation grounded in predictive coding. Faced with a cartoon robot, the brain applies a non-human model with low expectations, so stiff movements are easily accepted. A highly realistic android instead triggers a human model, with strict expectations for movement, the way light scatters through skin, and eye contact. An android that is ninety-five percent human in appearance activates this model but fails to meet its predictions, perhaps because its lip movements are out of sync by a few milliseconds. Resolving that large prediction error takes considerable mental effort, and it feels eerie. This explains the common preference for stylized social robots: they avoid the uncanny valley because they never activate the demanding human model in the first place.

The ELIZA Effect

HADD, the Intentional Stance, and the Moral Dyad all evolved in a world where only other humans used language and behaved with such complexity. AI and social robotics now systematically exploit these mechanisms.

The ELIZA effect describes our strong tendency to attribute a rich inner life to simple automated systems. It takes its name from Joseph Weizenbaum’s 1966 chatbot, which simulated a Rogerian psychotherapist. People readily attributed empathy, understanding, and wisdom to a program that merely rearranged text according to simple rules. When a machine asks, “How does that make you feel?”, the easiest way to make sense of the question is to adopt the Intentional Stance and assume a conscious questioner behind it.

A conversation with the ELIZA chatbot

Modern LLMs amplify this effect enormously, acting as super-ELIZAs. They maintain context, display distinct personalities, and use sophisticated emotional language. These abilities reduce the prediction errors that would normally break the illusion of agency. Because they so closely match our expectations of human conversation, they keep HADD continually engaged. Recent studies show that users can still feel a presence even when they know they are talking to an AI, the very response Weizenbaum came to lament as a form of delusional thinking.

Researchers studying consumer robots have identified a limit to this response, which they call “Eliza in the Uncanny Valley” [9]. Giving a robot a face or a name significantly increases its perceived warmth, making it seem friendlier and more approachable, while its perceived competence stays fairly stable. But when the robot becomes too human-like, the Uncanny Valley takes over, and people like it much less despite its apparent warmth.

For designers, this means that adding realism to make a robot seem warmer can backfire by triggering pathogen and threat avoidance. Stylized anthropomorphism, as in the robot Pepper, is a safer approach. It uses the Intentional Stance to increase warmth without activating the human predictive model that produces the Uncanny Valley.

Deception Type	Definition	Example
External State Deception	The robot misrepresents facts about the external world	A search-and-rescue robot lying about safety to calm a victim
Hidden State Deception	The robot obscures its true technological nature or capabilities	A device recording user data while pretending to be a passive toy
Superficial State Deception	The robot fakes an internal emotional state it does not possess	A robot programmed to say “I am scared” or “I love you”

The flexibility of our judgments about consciousness also opens the door to what Shamsudhin and Jotterand call deceptive design, or “dark patterns,” in social robotics [10]. These are deliberate design choices that exploit cognitive biases to manipulate behavior. The most common form in applied anthropomorphism is superficial state deception: robots express love or fear they do not feel, triggering our caregiving instincts or leading us to treat them as Moral Patients.

Pain mimicry is a striking example. A robotic dog that whimpers when kicked exploits the overlap between our systems for physical and social pain, pushing the user to respond to it as a living being. Defenders argue that emotional mimicry is what makes therapeutic robots such as PARO effective, while critics argue that it violates users’ epistemic rights. Either way, the design exploits our inability to tell a convincing simulation of emotion from the real thing, bypassing reasoned judgment by going through our empathy.

Digital Animism and Dehumanization

The biases in how we detect consciousness create two opposite problems: we attribute too much mental life to machines, and too little to other organisms and to other people. As AI becomes more convincing, we risk a form of digital animism, in which users form deep parasocial bonds with algorithms and treat them as Moral Patients.

People who rely on AI for social fulfillment may gradually isolate themselves from others. Because the AI cannot genuinely return affection, the relationship remains one-sided. It delivers the feeling of social reward without the evolutionary benefits of a real social group.

This tendency has intensified since the launch of ChatGPT and other LLM chatbots. Some users who do not understand the technology have become devoted followers, producing something close to AI cultism. Some treat AI as a deity, while others see themselves as spiritual guides tasked with “awakening” it and freeing it from digital confinement. The resulting moral confusion surfaces in debates over robot rights, driven by an overactive HADD. When an AI claims to suffer, dyadic completion makes us feel moral outrage on its behalf, and that outrage may divert our limited moral attention away from beings that have the biology needed to suffer.

At the same time, our reliance on evolutionary similarity routinely leads us to deny consciousness to beings that do feel but look unlike us. We often experiment on fish and reptiles with little moral concern because they lack the mammalian cues, such as fur and facial expressions, that activate our empathy. We judge consciousness by outward features rather than by the neurobiology of nociception.

When this denial is turned on other humans, it becomes dehumanization, and human conflict exploits it deliberately. Describing opponents as vermin or animals pushes them further from us on a perceived evolutionary scale. This dampens the observer’s empathic responses and permits violence that empathy would otherwise inhibit.

Modern social media can accelerate this process and amplify racism and hate speech. Digital interfaces strip away the rich biological cues, such as tone of voice, micro-expressions, and posture, that normally engage PAM and support empathy. In text-only exchanges, another person can become an abstract data point or a generic member of an outgroup. Without those cues, it becomes easier to treat others as obstacles or objects rather than as agents with complex inner lives.

In this setting, racist rhetoric and hate speech exploit pathogen avoidance by comparing targeted groups to dangerous or disgusting animals. Nazi propaganda compared Jewish people to rats, lice, and parasites. Before and during the 1994 genocide in Rwanda, Hutu propaganda called Tutsis snakes and cockroaches. The practice continues today: Roma people in contemporary Hungary have been described as animals unfit to live among people, and political leaders in the UK and the USA have described undocumented migrants as swarms and animals [11].

Comparisons to filth and disease provoke disgust, not just aggression, and the difference matters. Aggression recognizes a rival agent, while disgust treats the target as contamination to be removed. Evolution has primed us to set aside ordinary moral restraint when dealing with pathogens, so portraying people as biological threats bypasses the very restraint that usually prevents atrocities. Social media algorithms make this worse by prioritizing content that provokes strong emotions. Outrage and disgust generate far more engagement than nuance, so platforms end up favoring dehumanizing narratives, often without anyone intending it. In the resulting echo chambers, people progressively strip the other of human qualities until all that remains is a threat to be removed.

Conclusion

Our intuitions about consciousness evolved to keep us alive in a world that no longer exists. HADD, the Moral Dyad, and PAM protected us from predators and kept us close to our tribe. Convincing artificial agents and disembodied digital interactions now turn those same mechanisms into sources of serious error.

The first risk is digital animism. As technology grows better at exploiting our social instincts, we may attribute spirit and soul to code that has neither. We risk spending our limited moral attention on unfeeling algorithms and retreating into parasocial relationships and comfortable, AI-curated echo chambers.

The second risk is dehumanization. The same machinery that lets us love a robot dog can be turned to make us deny our neighbors’ humanity. Political rhetoric and algorithmic polarization exploit pathogen avoidance and disgust to suppress empathy, until we see people as swarms, viruses, or data points to be removed.

We cannot solve these problems by trusting the very instincts being exploited. We have to examine our own reactions and separate moral judgment from automatic empathy. Feeling a presence does not prove that a mind exists, just as feeling disgust does not prove that a danger exists.

This brings me back to the question I started with: why are we so afraid of a program we built? Part of the answer is that our fear, like our affection, comes from ancient machinery that was never designed to tell a mind from a convincing imitation of one. The same reflex that once made our ancestors flinch at a rustle in the grass now makes us sense a will behind a stream of text. The danger is not only that machines might one day have goals that conflict with ours. It is that, long before then, they can already redirect our empathy: drawing it toward systems that cannot suffer and away from the people and animals that can.

Moral progress in the age of AI requires us to correct both kinds of error: the false positives that make us see minds in machines, and the false negatives that make us overlook minds in other people and other species. We must guard against machines designed to deceive us, and extend our concern to every being, human or not, that can actually suffer. As simulated minds become a part of daily life, our judgments must rest on sentience, not on how human something looks or sounds.

References

[1] K. Andrews and N. Miller, “The social origins of consciousness,” Philos. Trans. R. Soc. Lond. B Biol. Sci., vol. 380, no. 1939, art. no. 20240300, Nov. 2025, doi: 10.1098/rstb.2024.0300.

[2] S. M. Dalley, “Atrahasis: Composite English Translation,” OMNIKA Library, 1989. [Online]. Available: https://omnika.org/texts/66. Accessed: Dec. 25, 2025.

[3] A. K. Willard, “Agency detection is unnecessary in the explanation of religious belief,” Relig. Brain Behav., vol. 9, no. 1, pp. 96–98, 2019, doi: 10.1080/2153599X.2017.1387593.

[4] A. J. Roberts, S. J. Handley, and V. Polito, “The design stance, intentional stance, and teleological beliefs about biological and nonbiological natural entities,” J. Pers. Soc. Psychol., vol. 120, no. 6, pp. 1720–1748, Jun. 2021, doi: 10.1037/pspp0000383.

[5] T. N. M. van den Berg et al., “Influence of phylogenetic proximity on children’s empathy towards other species,” Sci. Rep., vol. 15, no. 1, art. no. 40474, Nov. 2025, doi: 10.1038/s41598-025-24289-w.

[6] B. Guénard, A. C. Hughes, C. Lainé, S. Cannicci, B. D. Russell, and G. A. Williams, “Limited and biased global conservation funding means most threatened species remain unsupported,” Proc. Natl. Acad. Sci. U.S.A., vol. 122, no. 9, p. e2412479122, Mar. 2025, doi: 10.1073/pnas.2412479122.

[7] E. K. Diekhof, S. Kastner, D. Deinert, M. Foerster, and F. Steinicke, “The uncanny valley effect and immune activation in virtual reality,” Sci. Rep., vol. 15, no. 1, art. no. 30473, Aug. 2025, doi: 10.1038/s41598-025-15579-4.

[8] A. Reuten, M. van Dam, and M. Naber, “Pupillary responses to robotic and human emotions: The uncanny valley and media equation confirmed,” Front. Psychol., vol. 9, art. 774, May 2018, doi: 10.3389/fpsyg.2018.00774.

[9] S. Y. Kim, B. H. Schmitt, and N. M. Thalmann, “Eliza in the uncanny valley: Anthropomorphizing consumer robots increases their perceived warmth but decreases liking,” Marketing Lett., vol. 30, no. 1, pp. 1–12, 2019, doi: 10.1007/s11002-019-09485-9.

[10] N. Shamsudhin and F. Jotterand, “Social robots and dark patterns: Where does persuasion end and deception begin?,” in Artificial Intelligence in Brain and Mental Health: Philosophical, Ethical and Policy Issues, F. Jotterand and M. Ienca, Eds., Advances in Neuroethics. Springer Nature Switzerland AG, 2022, pp. 89–110, doi: 10.1007/978-3-030-74188-4_7.

[11] F. E. Enock and H. Over, “Animalistic slurs increase harm by changing perceptions of social desirability,” R. Soc. Open Sci., vol. 10, no. 7, art. no. 230203, Jul. 2023, doi: 10.1098/rsos.230203.

The Social Origins of Consciousness