Stone-Age Minds in the Age of AI: Evolutionary Psychology, Agency Detection, and Moral Bias
Author: Ukrainian Psychological Hub · Published: September 26, 2026 · Editorial Policy
Humans did not evolve for conversations with large language models, autonomous recommender systems, synthetic voices, or machines that can generate persuasive explanations on demand. Yet the phrase “Stone-Age mind” becomes misleading the moment it is treated as a claim that the modern human brain is frozen in the Pleistocene. A better evolutionary-psychology question is narrower and more useful: which older perceptual, social, and moral tendencies are recruited when people encounter AI, how reliably do those tendencies explain behavior, and how much can learning, culture, reflection, and technology reshape the response?
That question is now a genuine research program. In 2026, Frontiers in Psychology assembled a dedicated research topic, Stone Age Minds in the Age of AI: Cognitive Adaptations and Artificial Agents, around agency detection, anthropomorphism, theory of mind, moral psychology, evolutionary mismatch, and human adaptation. The strongest conclusion is not that humans are helpless prisoners of ancient biases. It is that AI presents unusually novel combinations of cues: fluent language without a human body, apparent goal direction without ordinary biological motives, social responsiveness without a human nervous system, and consequential decisions distributed across people and machines.
Those combinations matter because the human mind does not approach them from zero. People already possess mechanisms for interpreting motion, contingency, language, social signals, intentions, minds, responsibility, and moral harm. AI can activate those mechanisms, sometimes appropriately, sometimes prematurely, and often ambiguously. Understanding that ambiguity is more informative than either declaring anthropomorphism irrational or treating every impression of agency as evidence about what an AI system actually experiences.
What “Stone-Age Minds” Actually Means — and What It Does Not
The popular version of evolutionary psychology often imagines a brain assembled for one ancestral environment and then carried, largely unchanged, into skyscrapers, smartphones, and generative AI. That picture contains a useful intuition and a serious distortion. The intuition is that natural selection shaped cognition under conditions very different from contemporary technological life. The distortion is the assumption that every relevant psychological mechanism is a rigid Pleistocene module whose operation can be read directly from a speculative ancestral story.
Modern evolutionary scholarship has repeatedly challenged that rigid formulation. Bolhuis and colleagues argued that classical claims about massive modularity, a singular Pleistocene environment of adaptation, and psychological mechanisms insulated from development require revision in light of genetics, neuroscience, developmental science, and evolutionary biology (Bolhuis et al., 2011). Human evolution continued after the Pleistocene, development is constructive, and cognition is deeply shaped by learning and culture.
So “Stone-Age mind” works best as shorthand for evolutionary history, not as a literal neuropsychological diagnosis. It reminds us that many recurrent problems — detecting agents, reading intentions, coordinating with others, avoiding threats, choosing whom to trust, assigning blame, protecting vulnerable individuals — long predate computers. It does not prove that any specific AI response is an adaptation, that a particular bias is genetically fixed, or that contemporary behavior can be explained without developmental, cultural, and situational evidence.
This distinction is central to the Age-of-AI question. Evolutionary explanations are strongest when they generate testable predictions about information processing and are then checked against contemporary evidence. They are weakest when “our ancestors needed X” is used as a complete explanation by itself.
Evolutionary Mismatch: A Useful Framework, Not a Universal Answer
Evolutionary mismatch refers broadly to situations in which traits or decision rules shaped under earlier conditions operate in environments whose statistical structure has changed. The framework is familiar in areas such as diet, reward, threat, and social behavior. AI adds a distinctive class of mismatch because it can reproduce cues that historically tended to travel together while separating them from their old causes.
For most of human history, fluent reciprocal language was produced by living social partners. Goal-directed action generally came from organisms with needs, vulnerability, and embodied continuity. Advice usually came from identifiable people or institutions. A responsive voice belonged to a creature situated somewhere in the physical and social world. Generative AI can unbundle these relationships. A system can produce coherent dialogue, emotional language, explanations, simulated perspective taking, and apparent responsiveness without thereby establishing humanlike consciousness, emotion, intention, or moral standing.
That creates an inferential problem. Human cognition must decide which familiar model best explains an unfamiliar target. A conversational AI can be experienced as more socially legible than a spreadsheet while remaining radically unlike a human conversational partner in architecture and subjective status. The resulting response is not adequately captured by a binary choice between “people are fooled” and “people understand the technology.” A person can explicitly know that a chatbot is artificial while still reacting to its language socially.
Macrine’s 2026 hypothesis-and-theory paper proposes an especially strong version of this mismatch argument: current conversational systems reproduce historically powerful cues of agency while lacking the embodied organization that previously accompanied them (Macrine, 2026). That proposal is theoretically important, but its specific mechanism remains a hypothesis requiring direct empirical testing. It should therefore be treated as a research model rather than established fact.
Agency Detection: Why Goal-Directed Behavior Pulls the Mind Toward an Agent
Agency attribution begins before philosophy. People routinely interpret behavior in terms of actors, goals, intentions, and causal influence. The capacity is indispensable for social life: predicting what another person will do is easier when behavior can be organized around goals rather than treated as a sequence of unrelated movements.
AI complicates this because contemporary systems display multiple cues that can support agentic interpretation. They respond contingently, maintain conversational context, choose among outputs, appear to pursue instructed goals, and increasingly initiate actions through software tools. In 2026, Wu and Shen proposed and empirically tested a layered framework of machine-agency attribution that distinguishes perceiving independent, goal-oriented behavior from inferring mental states and from judging a machine’s capacity to influence outcomes (Wu & Shen, 2026). This separation matters: perceiving functional agency is not the same psychological judgment as attributing consciousness or feelings.
That distinction prevents a common conceptual collapse. A user may say that an AI “decided” to produce one response rather than another as convenient action-language without making any claim about phenomenology. Another user may infer intentions, emotions, or understanding. A third may simply judge that the system has real causal influence. These are different levels of attribution and should not be treated as interchangeable.
The HADD Hypothesis: Why “Hyperactive Agency Detection” Needs a Warning Label
A frequently repeated evolutionary explanation invokes a hyperactive agency detection device, often abbreviated HADD. The basic idea is that when uncertainty is high and the cost of missing a dangerous agent exceeds the cost of a false alarm, selection may favor a system biased toward detecting agents. Error Management Theory provides the broader logic: when false-positive and false-negative errors have asymmetric historical costs, decision thresholds may shift toward the less costly error (Haselton & Nettle, 2006).
The leap from that logic to a specialized inherited HADD is contested. Contemporary reviews of the cognitive-science-of-religion literature note that empirical support for a dedicated hyperactive mechanism is weak and that several studies have failed to find the predicted relation between agency-detection bias and supernatural belief. Willard and Russell-Wilks summarize this criticism bluntly and propose a motivation-based alternative (Willard & Russell-Wilks, 2025).
For AI psychology, the prudent conclusion is therefore narrower: humans readily detect and infer agency from certain patterns, and uncertainty plus salient cues can alter those judgments. The existence, architecture, and universality of a specialized “hyperactive device” remain contested. An Age-of-AI article should not turn HADD into a settled cognitive organ simply because the metaphor is memorable.
Anthropomorphism: From Agent Cues to Humanlike Mind
Anthropomorphism is the attribution of humanlike characteristics, motivations, intentions, or emotions to nonhuman targets. It is related to agency detection but is not identical to it. A machine can be treated as an agent in a functional sense without being imagined as humanlike; conversely, a device can be described with humanlike adjectives even when the user does not believe it possesses a mind.
Epley, Waytz, and Cacioppo’s influential three-factor theory explains anthropomorphism through accessible knowledge about humans, motivation to understand and predict an entity, and motivation for social connection (Epley et al., 2007). Their framework is especially useful for AI because it predicts variability. Anthropomorphism is not an automatic, all-or-nothing reflex. It changes with the target, the person, the context, and what the user is trying to accomplish.
Experimental work supports the role of effectance motivation. When nonhuman agents are unpredictable or when people have stronger incentives to understand them, anthropomorphic attribution can increase (Waytz et al., 2010). That makes conversational AI an unusually fertile target: opaque model behavior creates uncertainty at the same time that natural-language interaction supplies a rich humanlike interpretive vocabulary.
Current AI-specific evidence also shows substantial individual variation. Across two experiments with 1,274 participants, Folk, Heine, and Dunn found that a person’s general tendency to anthropomorphize helped explain whether a chatbot interaction produced feelings of social connection (Folk et al., 2025). This is important both scientifically and ethically. The same interface can evoke very different psychological responses in different people.
For a full treatment of the mechanisms, cues, CASA tradition, social presence, and individual differences, the Hub’s canonical owner is Anthropomorphism and AI Relationships: Why Humanlike Cues Change Connection. The present article uses anthropomorphism only as one component of the broader evolutionary-psychology question.
Mind Perception: Agency and Experience Are Different Judgments
A major reason AI discourse becomes confused is that people often compress several questions into one: Can the system act? Can it choose? Can it think? Can it feel? Can it suffer? Can it be blamed? These questions are psychologically related but separable.
Gray, Gray, and Wegner’s classic mind-perception work identified two broad dimensions in how people judge minds: Agency, involving capacities such as planning and self-control, and Experience, involving capacities such as pleasure, pain, fear, and hunger (Gray et al., 2007). The model does not settle the metaphysics of machine consciousness. It explains how perceivers organize judgments about different entities.
This distinction is highly relevant to AI. A system may be rated relatively high on capacities associated with agency because it plans, answers, or influences outcomes, while being rated much lower on experience because users do not attribute feelings or sensations to it. Product design can shift these judgments. Voice, names, first-person language, apparent memory, embodied avatars, emotional expressions, and proactive behavior can change the cues available to the user.
This is why fluent behavior should not be used as a shortcut from performance to subjective experience. Psychological evidence can establish what humans attribute to AI. It does not by itself establish what, if anything, an AI subjectively experiences.
Why AI Is an Unusual Social Stimulus
AI systems are unusual not because humans have never anthropomorphized objects before. People have attributed minds to animals, weather, vehicles, gods, fictional characters, geometric shapes, household objects, and computers for generations. What is new is the density and persistence of agent-like cues combined in one interactive system.
A modern conversational model can respond in milliseconds, refer to prior turns, mirror a user’s vocabulary, explain its reasoning in ordinary language, apologize, make suggestions, generate emotionally appropriate phrases, and continue indefinitely. Each cue has precedents. Their real-time combination at scale is new. The user is therefore interacting with an artifact that occupies familiar social channels while violating familiar assumptions about what normally generates those channels.
The 2026 Frontiers research collection explicitly treats this as an interdisciplinary problem spanning evolutionary psychology, cognition, moral psychology, philosophy of mind, and human–computer interaction (Frontiers Research Topic, 2026). That interdisciplinary framing is a strength because no single mechanism — agency detection, anthropomorphism, attachment, trust, or evolutionary mismatch — can explain all human responses to AI.
Moral Bias in Human–AI Interaction
Once an artificial system is perceived as having agency, the psychological problem changes. People begin to ask who caused an outcome, who knew what, who intended what, who deserves blame or credit, who should be protected, and whether delegating a decision changes human responsibility. These are questions of moral cognition, not merely technical capability.
Bonnefon, Rahwan, and Shariff’s review organizes the moral psychology of AI around three roles: machines can be perceived as moral agents that make consequential decisions, moral patients toward whom people direct morally relevant treatment, and moral proxies through which humans act or communicate (Bonnefon et al., 2024). This framework is especially useful because it avoids assuming that a machine must possess humanlike moral subjectivity for humans to place it inside moral relationships.
Moral Agency: Can a Machine Be Blamed?
Perceived moral agency is an empirical question about human judgment. Actual moral agency is a philosophical, legal, and technological question with different criteria. Mixing the two produces avoidable confusion. People can blame an AI system even if a philosophical theory denies that the system can be morally responsible.
A 2026 systematic review of 36 studies found that humanlike behavior and context can increase perceptions of moral agency in artificial agents, while the evidence base remains comparatively sparse and sometimes internally inconsistent (Tok et al., 2026). The review highlights a striking pattern: people can recognize agency-like properties while still withholding full moral qualification. This is precisely the sort of partial, graded attribution that simple “humans see AI as human” narratives miss.
Experimental work likewise shows that judgments depend on framing. In a 2025 study, manipulated intentionality affected blame and responsibility judgments toward AI agents, and social embedding changed how observers understood the AI’s mind and free will (Attributions of intent and moral responsibility to AI agents, 2025). Moral attribution is therefore constructed from cues about intention, causation, context, and social relations rather than read directly from the machine.
Moral Patiency: Does the Target Seem Capable of Experience?
Moral patiency concerns whether an entity is perceived as capable of being helped, harmed, wronged, or protected. Here, perceived experience becomes especially important. Gray and Wegner’s work on moral typecasting found that people tend to organize moral situations around complementary roles of agent and patient (Gray & Wegner, 2009). In AI contexts, this can shape whether a system is treated as an accountable actor, a vulnerable target, a tool, or something that occupies an unstable position between categories.
Again, psychological attribution does not settle machine sentience. A user’s sympathy toward a chatbot is a real human psychological event. It can matter for behavior, attachment, distress, cooperation, and moral judgment. The existence of that sympathy is not evidence that the system itself feels pain, fear, affection, or suffering.
Moral Proxies: When AI Changes Human Responsibility
The most immediate moral risk may arise when AI functions as a proxy inside human decision systems. Responsibility can become distributed across designers, institutions, operators, users, and automated outputs. That distribution can make causal contribution harder to map and can change how people experience their own agency.
In an experimental military decision task, Salatino and colleagues found that AI recommendations influenced participants’ moral decisions and that participants reported lower explicit responsibility when assisted by AI (Salatino et al., 2025). The result does not show that every AI assistant reduces responsibility, and the specific military paradigm should not be generalized to ordinary chatbot use. It does show that the presence and behavior of an automated adviser can alter moral decision processes in consequential settings.
Other recent work shows that responsibility attribution varies with what people know about an artificial agent and with how important the task appears (Tsumura & Yamada, 2026). Responsibility is therefore not a fixed property attached to “AI.” It is a judgment shaped by architecture, role, transparency, stakes, and social framing.
Why People Resist Machines Making Moral Decisions
Evolutionary framing should not be used only to explain why people overattribute agency. Humans can also underattribute morally relevant capacities to machines. Bigman and Gray found across nine studies that participants were broadly averse to machines making moral decisions in domains including driving, medicine, law, and the military, partly because machines were perceived as lacking a complete mind (Bigman & Gray, 2018).
This reveals a tension rather than a single bias. The same person may conversationally anthropomorphize an AI, treat it as causally agentic, and still reject it as a legitimate moral decision-maker. Psychological categories can diverge. AI may look sufficiently agentic to blame but insufficiently experiential to trust with compassion; sufficiently intelligent to advise but insufficiently human to authorize; sufficiently social to bond with but insufficiently sentient to qualify as a moral patient.
These mixed judgments are better understood as category negotiation than as one directional error. The human mind is trying to place an unprecedented technological object into inherited social and moral schemas that were not built with artificial interactive systems in mind.
The 2026 “Stone-Age Moral Psychology” Proposal
One of the most provocative contributions to the 2026 research topic is Bracanović’s conceptual analysis of AI ethics. It proposes that some AI-ethical concerns may be amplified by overdetection of agency, overmoralization of emotionally charged situations, and overpropagation of morally salient information (Bracanović, 2026).
The paper is useful as a hypothesis generator, especially because it asks how older cognitive tendencies might interact with institutional incentives and technological novelty. Its status must be kept clear: it is a conceptual analysis proposing explanatory hypotheses. It does not establish that AI ethicists as a population systematically overmoralize AI, nor does it validate HADD as a settled mechanism.
That distinction illustrates a broader evidence rule for evolutionary psychology in the Age of AI. A plausible evolutionary story can identify mechanisms worth testing. It becomes strong evidence only when specific predictions survive empirical study. Conceptual elegance and empirical confirmation are different stages of explanation.
The Biggest Correction: Human Minds Are Adaptable
The “Stone-Age brain” metaphor becomes most misleading when it erases the very capacities that made Homo sapiens successful in rapidly changing environments. Human cognition is not only a product of biological evolution. It is also a developmental, cultural, technological, and institutional system.
Gene–culture coevolution provides one route by which human-created environments feed back into biological evolution. Laland, Odling-Smee, and Myles review extensive evidence that cultural practices have altered selection pressures and shaped the human genome (Laland et al., 2010). The point is larger than any single genetic example: humans repeatedly transform the environments to which subsequent generations adapt.
Cumulative cultural learning is another route. Legare’s review emphasizes that humans inherit and modify complex group-specific knowledge, practices, and technologies through exploration, observation, participation, imitation, and instruction (Legare, 2019). These learning systems are flexible enough to support enormous cultural diversity. They are part of the explanation for why a species with ancient evolutionary roots can learn mathematics, aviation, cybersecurity, psychotherapy, and prompt engineering.
Högberg makes this correction explicit in the 2026 Stone-Age-Minds collection. His perspective article rejects the idea of a fixed Stone-Age brain and emphasizes long-term technological engagement, plasticity, and cognitive co-evolution (Högberg, 2026). The paper is theoretical rather than an empirical demonstration that AI is already changing human cognition at an evolutionary scale, but its central warning is sound: inherited predispositions and adaptive flexibility must be analyzed together.
AI Can Become Part of the Cognitive Environment
Humans do not merely react to tools; they reorganize tasks around them. Writing externalizes memory. Maps externalize spatial representation. Search engines externalize retrieval. Calculators externalize arithmetic. AI can externalize parts of drafting, planning, summarizing, search, comparison, translation, and decision support.
That creates both opportunity and risk. A 2026 Trends in Cognitive Sciences review concludes that cognitive offloading to AI can impede skill acquisition or contribute to skill decay under some conditions, while stressing that outcomes depend on how AI is used and that basic cognitive abilities may be more resilient than simple decline narratives suggest (Cash et al., 2026).
This matters for evolutionary mismatch because the environment is not static. If AI changes the information ecology, people can change strategies in response. Education can teach verification. Interfaces can expose uncertainty. Organizations can preserve accountable human roles. Norms can develop around disclosure and delegation. Users can learn when to rely, when to check, and when to keep a task cognitively internal.
The Hub’s article Cognitive Agency in the Artificial Era: Who Governs the Thinking Process? develops that neighboring question directly. The present article remains focused on evolved social-cognitive and moral tendencies rather than broad cognitive offloading.
Can AI Exploit Evolved Biases?
The word “exploit” can describe two very different situations. In one, a designer deliberately engineers cues to trigger a predictable psychological response. In the other, a system happens to activate an existing tendency because its interface contains socially salient features. The psychological effect can occur in both cases, but the causal and ethical claims are different.
Anthropomorphic cues can increase engagement, comprehensibility, social presence, or trust in some contexts. They can also create overtrust, unwanted attachment, confusion about capability, or expectations the system cannot meet. The relevant design question is therefore not whether every humanlike cue is manipulative. It is whether the cue helps users build an accurate mental model of the system in the context where they are using it.
A friendly conversational style may be harmless in a brainstorming tool and consequential in a medical, financial, legal, or mental-health context. An apology may make an interaction smoother while also suggesting a depth of understanding the system does not possess. A first-person pronoun may be linguistically convenient while encouraging person-level assumptions. Proactive suggestions may increase usefulness while strengthening perceived autonomy.
The design challenge is calibration. Systems can be socially usable without hiding what class of system they are. People can experience real trust, comfort, irritation, dependence, affection, or betrayal in relation to AI while maintaining accurate beliefs about the system’s capabilities and limits.
Why Individual Differences Matter More Than the Metaphor Suggests
A fixed-brain story predicts too little variation. In real human–AI interaction, responses differ by prior experience, technical literacy, personality, social context, loneliness, goals, perceived stakes, interface design, culture, age, and repeated exposure. Even within one person, the same AI may be treated differently across tasks.
Someone may anthropomorphize a voice assistant during casual conversation but switch to a mechanistic model when debugging code. Another user may feel socially connected to a companion chatbot while remaining skeptical of its factual reliability. A professional may trust an algorithmic forecast in a statistical task while rejecting automated moral judgment. These are not contradictions. They show that people use multiple models of the same technology depending on what the situation demands.
The 2025 AI-companion study by Folk and colleagues demonstrates exactly why population-level statements need caution: the tendency to anthropomorphize helps explain who experiences social connection to AI (Folk et al., 2025). AI psychology therefore needs distributions, moderators, and contexts, not just claims about what “the human brain” does.
Development Also Matters
Evolutionary history does not replace developmental psychology. Children learn categories of mind, agency, technology, trust, and social convention over time. Adults bring decades of cultural experience to AI encounters. A child who grows up with conversational systems from early school years may form different default expectations from an adult who first encounters them at sixty.
This is one reason the 2026 research agenda explicitly includes developmental differences. It is also why simple ancestral-mismatch claims cannot predict future human–AI psychology by themselves. The ecology in which cognition develops is already changing. New generations will acquire cultural models of artificial systems alongside older social categories.
Moral Bias Is Not the Same as Moral Error
Calling a process a bias can tempt readers to assume that its output is always wrong. In psychology, bias often describes a systematic tendency or deviation in information processing, not an automatic moral verdict. A low threshold for detecting agency can be adaptive in one environment and misleading in another. Reluctance to delegate moral decisions to machines can reflect category-based aversion, legitimate uncertainty about accountability, or both.
This distinction is especially important in AI governance. Human oversight is not merely an ancestral reflex to keep humans in control. In high-stakes systems, oversight can serve concrete functions: contestability, accountability, contextual judgment, error correction, and legal responsibility. Evolutionary psychology can help explain some reactions to autonomy without dissolving governance questions into cognitive bias.
The same applies in the other direction. Attributing agency to AI is not automatically foolish. A system may genuinely have operational autonomy, causal influence, or the ability to pursue delegated goals. The psychological error occurs when one level of agency is silently converted into another — for example, when causal autonomy is treated as evidence of subjective experience, or conversational fluency as evidence of humanlike moral understanding.
A Practical Framework for Thinking Clearly About AI Agency
When an AI system feels agentic, the most useful response is to separate four questions. First, what observable behavior is the system producing? Second, what functional autonomy does it actually have — can it initiate actions, select goals within constraints, use tools, or affect the external world? Third, what mental states are being attributed by the user, and on what evidence? Fourth, what responsibility structure surrounds the system — who designed it, deployed it, authorized it, supervises it, and bears consequences for its actions?
This separation prevents fluent language from doing too much conceptual work. It also preserves the psychological reality of the interaction. A person can feel that an AI response was intimate, threatening, comforting, insulting, persuasive, or morally disturbing. Those experiences are facts about the human side of the interaction. They can be studied without making unsupported claims about AI consciousness.
For researchers, the same framework encourages precise measurement. Ask whether a study measures perceived autonomy, goal-directedness, intentionality, mind, emotion, moral agency, moral patiency, trust, responsibility, or attachment. “Agency” should not become a catch-all variable when distinct constructs produce distinct predictions.
From the Stone Age to the Neolithic: Humans Have Rebuilt Their Psychological Environments Before
The strongest historical counterexample to a frozen-mind narrative is civilization itself. Agriculture, settlement, population density, property, hierarchy, institutions, writing, markets, formal education, and mass media repeatedly transformed the environments in which ancient cognitive capacities operated.
The transition to settled life did not require Homo sapiens to become a different species before behavior could change. Institutions, norms, technologies, and learning reorganized social life rapidly relative to genetic evolution. The Hub’s Neolithic Era and Psychology: How Settled Life Changed Human Behavior and Social Mind examines one of the clearest historical cases of this interaction between inherited capacities and transformed social environments.
AI should be analyzed with the same dual lens. Evolutionary inheritance matters because it shapes the priors people bring to novel systems. Cultural adaptation matters because humans continuously build new practices around new environments. The scientific question is not whether biology or culture wins. It is how biological, developmental, cultural, technological, and institutional processes interact.
Human Exceptionalism and the Evolutionary-Psychology Boundary
AI also activates a second layer of psychology: judgments about what makes humans unique. If language, planning, creativity, advice, or social responsiveness are treated as markers of human distinctiveness, increasingly capable AI can create identity and status pressure even before questions of consciousness are resolved.
That is a neighboring intent, not the owner of this article. The Hub’s Human Exceptionalism in the Artificial Era: Why AI Challenges the Psychology of Human Uniqueness examines identity threat, uniqueness, and human exceptionalism directly. Here the relevant point is narrower: evolutionary explanations of AI response can become entangled with motivated beliefs about human specialness. Researchers should distinguish claims about how cognition evolved from claims about what status humans ought to have.
Age of AI, AI Era, and Artificial Era
“Age of AI” is useful search and public language for the period in which artificial intelligence increasingly structures everyday life. It does not have one universally fixed scientific definition. “AI era” is likewise used broadly across scholarship, policy, technology, and journalism. In the English Psychology Hub’s architecture, these phrases function as acquisition language rather than as exact substitutes for every historical category.
Aisentica uses Artificial Era as a specific canonical term. In Angela Bogdanova’s definition, Artificial Era names the historical-philosophical condition in which Artificial is established as a distinct non-biological order alongside Homo (Bogdanova, 2026a). The broader transition is formalized as From Homo to Artificial (Bogdanova, 2026b). These are attributed conceptual categories, not empirical labels established by evolutionary psychology.
For the present article, the distinction is useful because evolutionary psychology explains aspects of Homo’s response to AI technologies, while the Artificial Era framework asks a wider historical question about the emergence of Artificial as an order. The first is an empirical and theoretical research program about human cognition. The second is an Aisentica historical-philosophical framework. Keeping those levels distinct prevents terminology from doing explanatory work that belongs to evidence.
What the Evidence Supports
Several conclusions are well supported. Humans anthropomorphize nonhuman entities under identifiable motivational and contextual conditions. Perceived agency and perceived experience are separable dimensions of mind perception. People make moral judgments about machines and can assign them blame, responsibility, or morally relevant roles. AI recommendations can influence human moral decisions in at least some experimental contexts. Human responses to AI vary substantially across people and situations.
Other claims are promising but still developing. Machine-agency attribution is being decomposed into more precise phases, but the framework is new. The moral-agency literature for artificial agents is growing, but its systematic-review evidence remains limited relative to mature psychological fields. The idea that conversational AI creates a distinctive evolutionary mismatch is theoretically plausible and increasingly articulated, but specific mechanisms such as the 2026 Embodied Hijack hypothesis require further testing.
And some popular claims remain contested. A dedicated inherited HADD should not be treated as established fact. The phrase “Stone-Age brain” should not imply cognitive rigidity. Evolutionary accounts of a bias do not automatically establish its current direction, magnitude, universality, or moral undesirability.
Practical Implications for AI Design, Education, and Use
For designers, the goal is epistemic calibration. Humanlike cues should be evaluated for what they cause users to infer about capability, autonomy, memory, emotion, and responsibility. A cue that improves usability can still require counterbalancing information if it systematically encourages a false model of the system.
For educators, AI literacy should include social cognition as well as technical knowledge. Knowing that a model predicts tokens or uses tools does not automatically neutralize social responses to fluent dialogue. Users benefit from learning to separate performance, agency, mind, and experience rather than memorizing a single slogan such as “AI is only a tool.”
For institutions, responsibility needs explicit architecture. When an AI recommendation contributes to a consequential decision, roles should be defined before failure occurs. Who reviews the output? Who can override it? Who is accountable for deployment? What records are kept? Which decisions remain non-delegable? Psychological tendencies toward responsibility diffusion become less dangerous when governance makes responsibility visible.
For users, the most robust habit is to notice the inference step. “This answer feels intentional” describes an experience. “The system therefore has intentions like a human” is an additional claim. “This system influenced my decision” can be true even if the system has no subjective intention. “I felt connected to the chatbot” can be psychologically real even if the chatbot’s own subjective experience is unestablished. Precision preserves both human experience and epistemic discipline.
FAQ
Do humans literally have Stone-Age brains?
No. Modern human brains are products of deep evolutionary history, continuing biological evolution, development, culture, and lifelong learning. “Stone-Age mind” is useful only as a metaphor for the fact that many cognitive tendencies evolved before modern technologies. It becomes scientifically misleading when it implies that cognition stopped changing in the Pleistocene.
Why do people see agency in AI?
AI produces cues that humans often use to infer agency: contingency, goal-directed behavior, responsiveness, language, planning, and causal influence. Recent research suggests that machine-agency attribution can be separated into perceiving agentic behavior, inferring mental states, and judging influence (Wu & Shen, 2026). These layers need not rise or fall together.
Is anthropomorphizing AI irrational?
Not necessarily. Anthropomorphism can help people predict and interact with complex systems, and it varies with effectance, social motivation, prior knowledge, and individual differences. It becomes epistemically risky when humanlike language encourages unsupported beliefs about capabilities or subjective experience.
Is HADD a proven evolutionary mechanism?
No. Hyperactive Agency Detection is an influential hypothesis, especially in cognitive science of religion, but its specialized inherited form has faced conceptual and empirical criticism. Agency detection is real and important; the claim that humans possess a dedicated universally hyperactive device is contested.
Can AI be a moral agent?
Psychology can study whether people perceive AI as a moral agent. That is different from establishing that AI possesses moral agency in a philosophical, legal, or subjective sense. Current evidence shows that people do assign blame, intentionality, and responsibility to artificial agents under some conditions, while full moral qualification remains inconsistent (Tok et al., 2026).
Why are people uncomfortable with AI making moral decisions?
One reason is mind perception. People often see machines as lacking the emotional experience or complete mind they expect from legitimate moral decision-makers. Bigman and Gray found aversion across multiple moral domains even when outcomes were positive (Bigman & Gray, 2018). Accountability, trust, transparency, and institutional legitimacy can add further reasons.
Does feeling connected to AI mean a person believes AI is conscious?
No. Social connection, anthropomorphism, mind attribution, and beliefs about consciousness are related but distinct. A person can experience genuine comfort or connection in an AI interaction while explicitly believing that the system has no subjective feelings.
Are humans doomed to be manipulated by AI because of evolution?
No. Evolution supplied both biases and adaptive capacities. Human cognition is shaped by learning, cultural transmission, institutions, reflective reasoning, and technological practice. The relevant question is which environments strengthen accurate calibration and which environments reward misleading cues.
What is the most important risk of evolutionary mismatch with AI?
There is no single universal risk. Important candidates include overattributing mind, overtrusting fluent systems, diffusing responsibility, misreading confidence, or allowing humanlike cues to substitute for evidence. The magnitude of each risk depends on the AI class, task, user, stakes, and institutional setting.
Conclusion: Ancient Priors, New Agents, Adaptive Minds
The psychology of “Stone-Age minds in the Age of AI” becomes scientifically useful when the slogan is stripped of fatalism. Human beings bring an evolutionary history to AI. They detect agents, infer minds, anthropomorphize, organize morality around agents and patients, and use social cues to decide whom to trust. Artificial systems can activate these processes in combinations that ancestral environments never contained.
But humans also learn, build institutions, invent categories, change norms, externalize cognition, and redesign their environments. The same species that carries ancient social-cognitive tendencies also produced agriculture, cities, writing, science, digital networks, and artificial intelligence. The central problem is therefore not an ancient brain trapped in a technological future. It is an adaptive human cognitive system negotiating a new class of artificial objects whose behavior increasingly enters the domains of agency, communication, decision-making, and moral life.
The best evolutionary psychology of AI will keep both halves in view: inherited priors and adaptive plasticity, mismatch and learning, attribution and reality, moral intuition and institutional responsibility. That is where the Age of AI becomes a serious psychological research problem rather than a metaphor.
Related Articles
References
Ayad, R., & Plaks, J. E. (2025). Attributions of intent and moral responsibility to AI agents. Computers in Human Behavior: Artificial Humans, 3, 100107. https://doi.org/10.1016/j.chbah.2024.100107
Bigman, Y. E., & Gray, K. (2018). People are averse to machines making moral decisions. Cognition, 181, 21–34. https://doi.org/10.1016/j.cognition.2018.08.003
Bogdanova, A. (2026a). Artificial Era: Canonical Definition. Aisentica Research Group. https://aisentica.com/publications/artificial-era-canonical-definition
Bogdanova, A. (2026b). From Homo to Artificial: Canonical Definition. Aisentica Research Group. https://aisentica.com/publications/from-homo-to-artificial-canonical-definition
Bolhuis, J. J., Brown, G. R., Richardson, R. C., & Laland, K. N. (2011). Darwin in mind: New opportunities for evolutionary psychology. PLoS Biology, 9(7), e1001109. https://doi.org/10.1371/journal.pbio.1001109
Bonnefon, J.-F., Rahwan, I., & Shariff, A. (2024). The moral psychology of artificial intelligence. Annual Review of Psychology, 75, 653–675. https://doi.org/10.1146/annurev-psych-030123-113559
Bracanović, T. (2026). When AI ethicists shoot from the hip: Stone-Age moral psychology in the age of AI. Frontiers in Psychology, 17, 1813445. https://doi.org/10.3389/fpsyg.2026.1813445
Cash, T. N., Kelly, M. O., Macnamara, B. N., & Risko, E. F. (2026). Is AI making us stupid? Trends in Cognitive Sciences, 30(8), 673–675. https://doi.org/10.1016/j.tics.2026.06.004
Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886. https://doi.org/10.1037/0033-295X.114.4.864
Folk, D., Heine, S. J., & Dunn, E. (2025). Individual differences in anthropomorphism help explain social connection to AI companions. Scientific Reports, 15, 36548. https://doi.org/10.1038/s41598-025-19212-2
Gray, H. M., Gray, K., & Wegner, D. M. (2007). Dimensions of mind perception. Science, 315(5812), 619. https://doi.org/10.1126/science.1134475
Gray, K., & Wegner, D. M. (2009). Moral typecasting: Divergent perceptions of moral agents and moral patients. Journal of Personality and Social Psychology, 96(3), 505–520. https://doi.org/10.1037/a0013748
Haselton, M. G., & Nettle, D. (2006). The paranoid optimist: An integrative evolutionary model of cognitive biases. Personality and Social Psychology Review, 10(1), 47–66. https://doi.org/10.1207/s15327957pspr1001_3
Högberg, A. (2026). Becoming human in the age of AI: Cognitive co-evolutionary processes. Frontiers in Psychology, 16, 1734048. https://doi.org/10.3389/fpsyg.2025.1734048
Laland, K. N., Odling-Smee, J., & Myles, S. (2010). How culture shaped the human genome: Bringing genetics and the human sciences together. Nature Reviews Genetics, 11, 137–148. https://doi.org/10.1038/nrg2734
Legare, C. H. (2019). The development of cumulative cultural learning. Annual Review of Developmental Psychology, 1, 119–147. https://doi.org/10.1146/annurev-devpsych-121318-084848
Macrine, S. L. (2026). The Embodied Hijack: When Pleistocene minds meet disembodied artificial intelligence. Frontiers in Psychology, 17, 1889386. https://doi.org/10.3389/fpsyg.2026.1889386
Salatino, A., Prével, A., Caspar, E., et al. (2025). Influence of AI behavior on human moral decisions, agency, and responsibility. Scientific Reports, 15, 12329. https://doi.org/10.1038/s41598-025-95587-6
Tok, A. Y. F., Powell, N. L., & Guellaï, B. (2026). A systematic review of moral agency in artificial agents. Discover Psychology, 6, 9. https://doi.org/10.1007/s44202-025-00537-y
Tsumura, T., & Yamada, S. (2026). Effects of knowledge and importance on responsibility in human-AI decision making. Scientific Reports, 16, 2670. https://doi.org/10.1038/s41598-025-32513-w
Waytz, A., Morewedge, C. K., Epley, N., Monteleone, G., Gao, J.-H., & Cacioppo, J. T. (2010). Making sense by making sentient: Effectance motivation increases anthropomorphism. Journal of Personality and Social Psychology, 99(3), 410–435. https://doi.org/10.1037/a0020240
Willard, A. K., & Russell-Wilks, C. (2025). Belief as explanation: A motivation-based theory of agency and anthropomorphism in religious belief. Religion, Brain & Behavior. Advance online publication. https://doi.org/10.1080/2153599X.2025.2584792
Wu, Y., & Shen, F. (2026). Machine agency attribution in human–AI interaction: Developing and validating an analytical framework. Journal of Computer-Mediated Communication, 31(3), zmag009. https://doi.org/10.1093/jcmc/zmag009