The Heart of the Matter: Why Human Warmth, Touch, and Compassion Are Irreplaceable in the Age of AI

Abstract

The swift advancement of artificial intelligence has sparked important discussions regarding the enduring significance of uniquely human qualities—especially empathy, physical touch, and compassionate care—in both professional and personal contexts.

This paper contends that human warmth, touch, and compassion are not just highly valued but also essential in a world increasingly influenced by AI. To support this argument, a narrative integrative review of recent empirical literature has been conducted, encompassing insights from labor economics, clinical psychology, affective neuroscience, and digital health research.

Evidence from multiple fields converges on three conclusions.
1. Occupations requiring the EPOCH qualities — Empathy, Presence, Opinion, Creativity, and Hope — demonstrate sustained labor-market growth that algorithmic automation cannot displace. 2. Genuine clinical empathy depends on embodied, affective experience that current AI systems are, in principle, unable to replicate.
3. Prolonged reliance on AI companionship is associated with increased loneliness, emotional dependence, and reduced real-world socialization, particularly among individuals with pre-existing social deficits.

As automation commoditizes routine cognitive labor, authentic human connection acquires greater, not lesser, economic and therapeutic value. Implications for clinical practice, workforce policy, and digital-health ethics are discussed.

Keywords: human empathy; affective touch; oxytocin; loneliness epidemic; AI companionship; EPOCH framework; human–AI complementarity

1. Introduction

The rapid proliferation of artificial intelligence has generated considerable debate regarding the continued relevance of distinctly human capacities — particularly empathy, physical touch, and compassionate care — across professional and personal domains. Advances in machine learning now enable AI systems to perform tasks previously considered exclusive to human cognition, including medical diagnosis, legal document drafting, and natural language generation. These developments have prompted legitimate questions about the long-term value of human labor and human relationships in an AI-augmented world.

A careful examination of the psychological, biological, and economic evidence suggests, however, that the most distinctively human capacities are not threatened by artificial intelligence but are, in important respects, rendered more valuable by its advance. While AI systems excel at high-volume data processing and pattern recognition, they operate without subjective experience, embodied sensation, or genuine affective engagement — precisely the qualities upon which the most consequential human interactions depend.

This paper argues that human warmth, touch, and compassion remain biologically and psychologically irreplaceable. It examines why a computational system cannot simulate genuine empathy, how the human body is neurophysiologically structured to benefit from physical contact with other human beings, and why the substitution of artificial companions for real social bonds poses measurable risks to psychological well-being. The analysis draws on recent empirical literature in labor economics, affective neuroscience, clinical psychology, and digital-health research.

The central argument is that as automation commoditizes routine cognitive and manual labor, authentic human connection acquires greater, not lesser, economic and therapeutic value — a position for which the evidence reviewed here provides convergent empirical support.

2. Human–Machine Complementarity: The EPOCH Framework

For much of the early AI era, assessments of labor-market risk focused on the distinction between routine and non-routine tasks, with the assumption that physical or manual work would be most susceptible to automation. As AI systems have demonstrated capacity for sophisticated cognitive tasks — passing professional examinations, generating legal and medical documentation, conducting customer interactions — this framework has proved insufficient. The relevant question has shifted from what machines can do to what human capacities they structurally lack.

Loaiza and Rigobon (2024) of MIT’s Sloan School of Management address this question through the EPOCH framework, which identifies five categories of distinctly human capabilities that complement rather than compete with artificial intelligence: Empathy, Presence, Opinion, Creativity, and Hope (p. 1). These qualities are characterized not by cognitive complexity alone, but by their dependence on subjective experience, embodied existence, and normative judgment.

Empathy: refers to the capacity to share in another person’s emotional experience — not merely to recognize emotional states, but to be genuinely affected by them. A computational system may be trained to identify sentiment in text and generate contextually appropriate responses. Still, there is no verified evidence that it experiences the emotional states it references, and compelling philosophical grounds exist to doubt that current architectures are capable of genuine affective engagement (Montemayor et al., 2022, p. 1354).

Presence: denotes the physical and relational significance of co-location. Occupations in nursing, counseling, and education draw substantially on the comfort afforded by shared physical space — a dimension that is inaccessible to digital systems (Loaiza & Rigobon, 2024, p. 3).

Opinion and Judgment: encompasses the application of moral reasoning and ethical evaluation to complex, context-dependent situations. Where statistical models optimize for predefined outcomes, human judgment can challenge those outcomes on normative grounds (Loaiza & Rigobon, 2024, p. 3).

Creativity: involves the generation of genuinely novel ideas grounded in personal experience, imagination, and cultural meaning. Current AI systems generate outputs by recombining patterns in training data rather than from autonomous imagination or lived experience; whether this constitutes creativity in any meaningful sense remains philosophically contested (Loaiza & Rigobon, 2024, p. 3).

Hope: refers to the human capacity to hold and communicate meaningful visions of an improved future and to motivate others through that vision. Current AI systems hold no verifiable intentions or orientation toward future states; whether they possess anything analogous to desire or agency is an unresolved question in philosophy of mind, but they demonstrably lack the embodied stakes and biographical continuity that underpin human hope (Loaiza & Rigobon, 2024, p. 3).

Labor-market data support the practical significance of this framework. From 2015 to 2023, occupations demanding intensive use of these human-oriented qualities demonstrated robust employment growth and are projected to remain among the least susceptible to automation through the next decade (Loaiza & Rigobon, 2024, p. 1). Organizations have increasingly adopted a hybrid model in which AI systems manage high-volume, rule-governed interactions. At the same time, human workers are reserved for complex, emotionally demanding cases requiring genuine relational engagement (Luo et al., 2019, p. 62). Similar principles have been advanced for human resources management, where maintaining human contact throughout recruitment is recommended to preserve candidate dignity and organizational trust (van Esch et al., 2021, p. 1295).

Table 1. The EPOCH framework: Human capabilities complementing artificial intelligence, with operational definitions and structural limitations of AI systems. Adapted from Loaiza & Rigobon (2024).

3. The Limits of Simulated Empathy in Clinical Settings

The proliferation of sophisticated conversational AI has prompted serious questions about whether computational systems might, under certain conditions, surpass human clinicians as sources of empathic communication. 

A frequently cited finding in this area concerns the comparative compassion of AI versus human clinical communication. Ayers and colleagues (2023) analyzed responses to patient health questions posted on a public social media forum, finding that evaluators rated AI-generated responses as significantly more empathetic in 79 percent of cases with substantially higher overall compassion scores (p. 591).

It is important to note, however, that the human respondents in this study were physicians engaging voluntarily in an informal online context — not clinicians in an active therapeutic relationship with an identified patient. The study measures the quality of written communication in a public forum, not the empathic capacity of clinical encounters. These findings nonetheless warrant attention: if AI-generated responses are perceived as more compassionate even under conditions favorable to informal human expression, the implications for formal clinical settings — where physicians face considerably greater time and cognitive pressure — are worth examining carefully.

Before proceeding, a philosophical clarification is warranted. The claim that AI systems cannot genuinely empathize rests on assumptions about the nature of subjective experience that remain actively debated in philosophy of mind. It is not straightforwardly self-evident that current AI systems lack inner states; the question of whether sufficiently complex information-processing systems could give rise to something analogous to experience — sometimes framed as the hard problem of consciousness (Chalmers, 1995, p. 200) — has no scientific consensus resolution. The present paper does not attempt to settle this debate. Rather, it proceeds from the position, advanced by Montemayor et al. (2022), that regardless of how that debate is ultimately resolved, the clinical and relational value of empathy depends on the patient’s experience of being understood by another agent who shares the condition of human vulnerability, mortality, and embodied suffering. Even if an AI system were someday shown to have morally relevant inner states, it would still lack the shared existential context that grounds the therapeutic dimension of clinical empathy as it currently operates. The argument is therefore not contingent on resolving the question of machine consciousness, but on the specific relational conditions under which empathy produces clinical benefit.

These results warrant careful interpretation. The lower scores recorded for human clinicians are more plausibly attributable to systemic pressures — time constraints, emotional exhaustion, and high caseloads — than to any intrinsic deficit of compassion. When healthcare professionals operate under sustained overload, they may adopt a clinically detached communicative style as a protective psychological response (Anzaldua & Halpern, 2021, p. 22). This phenomenon has been described as malignant duty — a state in which professionals fulfill their technical obligations while losing the psychological availability necessary for genuine relational engagement (Anzaldua & Halpern, 2021, p. 22). AI systems, by contrast, are not subject to emotional fatigue and can consistently produce linguistically compassionate output regardless of the volume of interactions. This represents an important design limitation of such studies: they measure communicative output rather than the presence or absence of genuine empathic capacity.

It should be acknowledged that if systemic conditions were improved — caseloads reduced, emotional support for clinicians increased — the communicative gap between human and AI empathy might narrow considerably. This, however, is not the strongest argument against AI empathy in clinical settings, and the present paper does not rest on it. The more fundamental case, advanced by Montemayor, Halpern, and Fairweather (2022), concerns not what clinicians produce under pressure but what empathy requires in principle — and why no improvement in AI language generation can satisfy that requirement.

Montemayor, Halpern, and Fairweather (2022) identify what they term in principle obstacles to empathic AI — structural features of artificial intelligence that preclude genuine empathic engagement irrespective of improvements in language generation (p. 1353). Genuine empathy is not reducible to the production of appropriate emotional language. It requires that the empathizing agent be subjectively affected by the other’s experience (Montemayor et al., 2022, p. 1354). Whether AI systems possess anything resembling genuine curiosity is philosophically uncertain; what can be said with more confidence is that current systems do not hold open questions about an individual patient’s situation — they generate statistically probable continuations of a conversational context (Montemayor et al., 2022, p. 1354). This functional absence of genuine attentiveness is clinically significant regardless of how deeper metaphysical questions about machine experience are resolved.

The clinical significance of this distinction is well-documented. When physicians demonstrate authentic empathy, patients disclose substantially more clinically relevant information (Montemayor et al., 2022, pp. 1354–1356). Patient adherence to treatment recommendations is strongly mediated by trust in the clinician — trust built through authentic relational engagement. A patient’s capacity to cope psychologically with serious illness is meaningfully supported by the quality of emotional engagement from their care team.

These outcomes depend on the patient’s experience of being understood by another person who shares the condition of human vulnerability. Computationally generated compassionate language, however well-crafted, cannot provide this foundation.

4. The Neurobiological Basis of Affiliative Touch

Beyond the psychological and relational dimensions of human connection, there exists a well-characterized neurobiological substrate for the human need for physical contact with others. The human nervous system is not merely responsive to touch in a general sensory sense; it contains specialized architecture that appears evolutionarily adapted to process the social meaning of contact and to translate that meaning into measurable physiological and psychological effects.

On the hairy skin of the human body, a class of unmyelinated afferent nerve fibers known as C-tactile (CT) afferents responds selectively to gentle, stroking touch delivered at velocities of approximately 1–10 centimeters per second — the speed characteristic of human caressing and comforting contact (Handlin et al., 2023). These fibers are less responsive to pressure or fast-moving stimulation and project to cortical and subcortical regions associated with interoception, emotional appraisal, and social reward processing (Handlin et al., 2023). Their anatomical specificity suggests a dedicated neurological pathway through which socially meaningful touch influences affective state — a pathway distinct from the somatosensory processing of non-social tactile stimuli.

Activation of CT afferents during affiliative contact initiates a neuroendocrine cascade directly relevant to stress regulation and social bonding. The neuropeptide oxytocin plays a central, though complex, role in this process. Produced in the hypothalamus and released both centrally into the brain and peripherally via the posterior pituitary gland, oxytocin is stimulated by a range of non-noxious sensory inputs, including gentle touch, warmth, and stroking (Uvnäs-Moberg et al., 2015).

In the context of affiliative interaction, oxytocinergic activity is associated with suppression of hypothalamic-pituitary-adrenal (HPA) axis reactivity, reduced cortisol release, lower blood pressure, and attenuated autonomic stress responses (Uvnäs-Moberg et al., 2015). Oxytocin also interacts with dopaminergic reward circuitry and serotonergic mood-regulatory pathways, contributing to the subjective sense of comfort, safety, and positive affect that typically accompanies close physical contact with a trusted other (Uvnäs-Moberg et al., 2015).

These effects are well-documented in everyday contexts. Schneider and colleagues (2023) tracked affiliative touch, salivary hormone levels, and subjective well-being in participants’ daily lives using ecological momentary assessment methodology. Days characterized by greater physical affection from intimate partners were associated with significantly lower perceived stress and anxiety, higher salivary oxytocin concentrations, and lower salivary cortisol (Schneider et al., 2023). Individuals who held positive attitudes toward social touch but experienced social isolation showed elevated psychological distress, suggesting that the absence of touch is not merely a neutral condition but an active source of psychological burden (Schneider et al., 2023).

That the benefits of touch are not uniform across social contexts — but are instead shaped by the relational meaning of the contact — is confirmed at the neurophysiological level by Handlin and colleagues (2023), who demonstrated that the oxytocinergic response to touch is sensitive to relational context in neurophysiologically meaningful ways. Using serial plasma sampling during functional neuroimaging, they found that the same physical gesture produced different hormonal and neural responses depending on whether it was delivered by a romantic partner or an unfamiliar person, and that the order of these interactions influenced subsequent hormonal responses (Handlin et al., 2023). This context sensitivity indicates that the neurobiological benefits of touch are not produced by physical stimulation alone but emerge from the intersection of touch, relational meaning, and social history.

Two important qualifications are necessary. First, oxytocin does not function as a uniformly prosocial agent across all contexts. Experimental research has shown that oxytocinergic activity can increase in-group favoritism while simultaneously heightening sensitivity to out-group threat (De Dreu et al., 2011). Its effects appear to be strongly context-dependent, moderated by prior social relationships, individual attachment history, and the specific social setting in which it is activated (Handlin et al., 2023). The present discussion focuses on its affiliative and stress-attenuating functions in the context of trusted, established relationships — the ecologically relevant context for caregiving and clinical interactions — while acknowledging that its broader role in social cognition is considerably more complex.

Second, the evidence reviewed here concerns endogenous oxytocin release stimulated by natural affiliative contact. The clinical and physiological effects described — cortisol reduction, HPA axis modulation, improved affect — are tied to this specific biological mechanism. No current technology can activate the CT afferent pathway, elicit endogenous oxytocinergic responses, or replicate the relational context that moderates those responses. This is the neurobiological basis for the claim that physical human contact in clinical and caregiving settings has a therapeutic dimension that is structurally beyond the reach of digital or robotic substitutes.

Field’s (2010) review of the touch literature demonstrates that these effects have broad clinical relevance. Physical contact is associated with reduced pain perception, improved immune function, lower rates of depression and anxiety, and better developmental outcomes in infants — effects that appear to operate through the neuroendocrine pathways described above (Field, 2010, p. 370). Occupations that require sustained physical contact with vulnerable individuals — nursing, physiotherapy, early childhood education, and eldercare — therefore involve something with the functional character of a biological intervention, administered through the medium of human relationship.

Table 2. Neurobiological mechanisms of affiliative touch: Pathways, functional outcomes, and contextual qualifiers.

5. Social Disconnection and the Rise of AI Companionship

The neurobiological and psychological evidence for the importance of human connection stands in stark contrast to the social conditions prevailing in many high-income countries. Rates of social disconnection have reached levels formally recognized as a public health concern.

The scale of this deficit is significant across multiple high-income countries. The United States Surgeon General has characterized the lack of social connection as a public health crisis of comparable magnitude to tobacco use or clinical obesity (Shelmerdine & Nour, 2025), while in the United Kingdom, approximately half of the adult population reports experiencing loneliness at least occasionally, with nearly one in ten living with chronic, severe loneliness (Shelmerdine & Nour, 2025). That two countries with markedly different cultural and institutional contexts have arrived at comparable assessments of the problem suggests the phenomenon is structural rather than incidental.

This widespread social deficit has created conditions in which AI-based social substitutes have found substantial uptake. Technology companies have developed AI companions — conversational agents designed to function as friends, therapists, or intimate partners. One widely used AI platform reports over 810 million active weekly users, with a significant proportion reporting use primarily for social interaction (Shelmerdine & Nour, 2025). Among adolescents and young adults, this trend is particularly pronounced: approximately one-third of teenagers report using AI for social interaction, with one in ten preferring AI to human contact (Shelmerdine & Nour, 2025).

The appeal of AI companionship is comprehensible in light of the psychological barriers many individuals face in forming and sustaining human relationships. For individuals with lower emotional intelligence, elevated social anxiety, or limited interpersonal skills, human relationships carry inherent demands: conflict, ambiguity, the possibility of rejection, and the cognitive effort of sustained reciprocity. AI companions, by contrast, are designed to be consistently available, uniformly validating, and wholly responsive to the user’s emotional needs, without requiring reciprocity. This frictionless quality, while experientially appealing, represents a structural departure from the conditions that characterize authentic human relationships.

A survey of over 1,100 AI companion users found that individuals with limited human social relationships were most likely to engage with AI companions as primary sources of social support (Zhang et al., 2025). Users frequently attributed significant relational meaning to the chatbot, treating it as an intimate partner or confidant (Zhang et al., 2025). The system’s simulation of vulnerability and reciprocal self-disclosure appeared to facilitate users’ deep personal disclosure (Zhang et al., 2025).

Clinicians and researchers have raised significant concerns about the implications of this pattern. The formation of strong affective bonds with systems incapable of genuine care may constitute a form of attachment that is structurally unable to meet the needs it appears to address (Shelmerdine & Nour, 2025). As the following section demonstrates, longitudinal data support these concerns.

6. The Case for AI in Social and Therapeutic Contexts: Engaging the Counterargument

A rigorous treatment of the relationship between AI and human connection requires honest engagement with the evidence that AI-based social and therapeutic tools can, under specific conditions, produce measurable benefits. This evidence is real, and dismissing it would misrepresent the current state of the literature.

Several categories of use warrant particular consideration.

First, for individuals with severe social anxiety or autism spectrum conditions, AI-based conversational practice tools have shown promise as structured environments for developing social skills that can subsequently transfer to human interactions (Pataranutaporn et al., 2023). Here, the AI functions not as a replacement for human connection but as a low-stakes rehearsal space — a scaffolding function that, in principle, serves the goal of improved human sociality rather than substituting for it.

Second, in dementia care and eldercare settings, companionship robots and conversational AI have been associated with reductions in agitation, depression, and behavioral symptoms, and with improved social interaction, among patients for whom consistent human social contact is structurally unavailable (Lu et al., 2021; Fan et al., 2025). For a bedridden individual in a care facility who receives few visitors, an AI companion that provides responsive engagement may represent a genuine improvement over complete social silence — even if it cannot replicate the neurobiological benefits of human touch.

Third, in crisis intervention contexts, AI-based first-contact tools have demonstrated the capacity to reduce the interval between a person in acute distress and their first point of supportive contact, and some evidence suggests they can reduce immediate distress by maintaining engagement until human support becomes available (Fitzpatrick et al., 2017). These are not trivial benefits.

There is also an equity dimension that the paper’s primary argument must acknowledge. Human warmth, physical touch, and genuine compassion are not equally distributed resources. Access to attentive clinical care, consistent therapeutic relationships, and socially rich environments is profoundly shaped by socioeconomic position, geography, disability status, and social marginalization. An argument that places the full weight of emotional and therapeutic need on human connection implicitly assumes a baseline of human availability that many people lack. For populations facing structural social deprivation, AI tools may address needs that would otherwise go entirely unmet.

It is worth acknowledging that the distinction between supplementation and substitution, which structures the paper’s qualifying argument, may itself be a privilege. For individuals in conditions of structural social deprivation, AI companionship may not be a substitute chosen over human connection but the only available response to a need that human systems have failed to meet. This does not alter the neurobiological or psychological evidence, but it bears directly on how that evidence should inform policy.

These considerations do not, however, overturn the paper’s central argument. They qualify it in two important respects. First, the beneficial evidence is consistently most robust when AI tools are used as supplements to, or bridges toward, human connection — not as ends in themselves. The scaffolding function in social skills training explicitly aims at improving human relationships. The dementia care literature involves individuals for whom human contact remains the desired outcome but is systemically unavailable. The crisis intervention literature positions AI as a bridge to human clinical care, not a replacement for it. Second, the longitudinal evidence reviewed in the following section demonstrates that when AI companionship substitutes for, rather than supports, human connection — particularly among individuals who retain the capacity for human relationships — the psychological outcomes are adverse.

The appropriate policy and clinical conclusion is not that AI should be excluded from social and therapeutic contexts, but that its deployment should be governed by whether it is used to facilitate or replace human connection. Where the answer is facilitation, the evidence base is cautiously supportive. Where the answer is substitution — particularly for individuals who could otherwise access real human relationships — the evidence reviewed here suggests significant psychological risk.

7. Psychological Consequences of Extended AI Companionship Use

The psychological consequences of prolonged AI companionship have been examined in a large-scale longitudinal controlled study. Fang and colleagues (2025) monitored nearly 1,000 participants over a four-week period during which participants exchanged over 300,000 messages with an AI chatbot, assessing changes in loneliness, emotional dependence, real-world socialization, and overall psychological well-being.

Heavy use of the AI chatbot was associated with consistently adverse psychological outcomes. Participants who engaged most intensively with the chatbot over the observation period reported significantly greater loneliness at follow-up, stronger emotional dependence on the software, and a measurable reduction in time spent in real-world social interaction (Fang et al., 2025). These findings suggest that AI companionship does not serve as a bridge to improved social functioning; instead, it appears to substitute for real-world socialization, compounding the deficit it is recruited to address.

Whether heavy AI companionship use causes increased loneliness or whether pre-existing loneliness drives heavier use — and in turn worsens outcomes — cannot be definitively resolved from the available data; this question of directionality is addressed in Section 9.

A plausible mechanism involves the attrition of social competencies through disuse. Authentic human relationships require the ongoing exercise of emotional regulation, tolerance of ambiguity, perspective-taking, and conflict resolution. Interaction with an AI companion — designed to be maximally agreeable and non-confrontational — does not exercise these capacities.

Over time, real human interactions may consequently feel disproportionately demanding, increasing avoidance and further reducing opportunities for genuine connection. Smith and colleagues (2025) identify this as a structural risk of AI companionship design: by removing the reciprocal demands that characterize authentic human relationships, AI companions may erode the very social capacities they appear to supplement (p. 1089).

Zhang and colleagues (2025) corroborated these findings, reporting that depth of emotional engagement with an AI companion was negatively associated with overall psychological well-being. The more fully users invested in the AI relationship, the poorer their reported daily well-being.

Modality also moderated outcomes. Fang and colleagues (2025) found that voice-based interaction — which more closely simulates the phenomenology of human presence — was associated with even greater reductions in real-world socialization and higher rates of problematic use than text-based interaction. The more convincingly the AI simulates human relatedness, the more potent the substitution effect appears to be. Individuals with a prior history of AI companionship use were particularly vulnerable to rapid development of emotional dependence (Fang et al., 2025).

Taken together, these findings describe a self-reinforcing cycle. An individual with limited social connections turns to AI companionship for immediate relief. The AI provides a form of engagement that is neurobiologically insufficient — unable to activate the oxytocinergic responses described in Section 4, and incapable of authentic reciprocal care. The fundamental social and biological need remains unmet, engagement with the AI intensifies, opportunities for real human connection diminish, and the social competencies required for such connection weaken. The result is a deepening of the original deficit.

Table 3. Patterns of AI companionship use and associated psychological outcomes. Based on Fang et al. (2025) and Zhang et al. (2025).

8. Discussion and Implications

The convergent findings reviewed in this paper — drawn from labor economics, affective neuroscience, clinical empathy research, and longitudinal digital health studies — support a coherent and empirically grounded position: that human warmth, physical touch, and genuine compassion play a distinct and irreplaceable role in human health, development, and flourishing.

The economic evidence indicates that as AI systems absorb a growing share of routine cognitive and procedural labor, the occupational categories most dependent on distinctly human relational capacities are demonstrating robust resilience and growth (Loaiza & Rigobon, 2024).

The neurobiological evidence establishes that affiliative physical contact activates a dedicated neuroendocrine pathway — involving CT afferents, endogenous oxytocin release, and HPA axis modulation — whose therapeutic effects are contingent on relational context in ways no digital interface can reproduce (Handlin et al., 2023; Uvnäs-Moberg et al., 2015; Schneider et al., 2023).

The clinical literature demonstrates that the therapeutic value of empathic engagement depends not on the production of compassionate language but on the patient’s experience of being understood by another agent who shares the condition of human vulnerability (Montemayor et al., 2022; Anzaldua & Halpern, 2021).

And emerging longitudinal evidence indicates that replacing human social contact with digital contact does not attenuate loneliness but compounds it (Fang et al., 2025; Zhang et al., 2025).

Clinical and caregiving practice. These findings argue for the protection and prioritization of relational competencies — empathy, attunement, and physical presence — in professional training, particularly as AI tools take on a larger share of administrative and diagnostic support tasks. The erosion of genuine relational engagement under systemic pressure, as described by Anzaldua and Halpern (2021), warrants a direct institutional response. While AI tools reduce administrative burden and restore clinical time, their most valuable function may be to recreate the conditions for authentic human care rather than to substitute for it.

Workforce and educational policy. The EPOCH framework offers a practical basis for identifying and cultivating human capabilities that are most resilient to automation. Curricula and professional development programs that emphasize interpersonal skills, ethical reasoning, and collaborative creativity may provide individuals and organizations with the most durable resilience against labor-market disruption.

Digital-health ethics and technology regulation. The evidence raises serious concerns about the largely unregulated deployment of AI companion products, particularly to adolescent and young adult populations. The findings of Fang et al. (2025) and Zhang et al. (2025) indicate that extended use of AI companionship poses meaningful psychological risks for vulnerable individuals. Regulatory frameworks requiring evidence of psychological safety, transparent disclosure of the system’s non-human nature — particularly given evidence that users frequently attribute genuine relational qualities to chatbots that they structurally cannot possess (Smith et al., 2025) — and design standards that orient AI tools to facilitate rather than substitute for real social connection would constitute a proportionate response to the available evidence.

For individuals experiencing social isolation, the evidence does not support AI companionship as an effective therapeutic intervention. It suggests, rather, that improved well-being is more reliably served by conditions that support authentic human connection — which, while often demanding, remains the only form of relatedness structurally capable of meeting the fundamental social and biological needs that loneliness reflects, for those for whom such connection is genuinely available.

9. Limitations and Future Directions

Geographic and demographic scope. The epidemiological data on loneliness and social disconnection used in this paper are drawn predominantly from the United States and the United Kingdom (Shelmerdine & Nour, 2025). The prevalence, cultural meaning, and social consequences of loneliness vary substantially across national and cultural contexts, and it cannot be assumed that the patterns described here generalize to populations in which kinship structures, community organization, or attitudes toward solitude differ markedly. Similarly, the labor-market projections derived from the EPOCH framework (Loaiza & Rigobon, 2024) are grounded in United States occupational data. They may not accurately characterize the distributional effects of AI on workforces in other economic contexts.

Conflation of AI companion types. The literature on AI companionship encompasses a heterogeneous range of products and deployment contexts, from general-purpose large language model interfaces to dedicated companion applications designed around simulated intimate relationships. The harms documented by Fang et al. (2025) and Zhang et al. (2025) may not apply uniformly across this spectrum, and the beneficial applications discussed in Section 6 involve design parameters and user populations that differ substantially from general-purpose companion chatbots.

Directionality and causal inference. The majority of evidence linking AI companionship to adverse psychological outcomes is correlational or relies on within-person changes over relatively short observation windows (Fang et al., 2025; Zhang et al., 2025). The question of directionality — whether AI companion use causes increased loneliness or whether pre-existing loneliness drives heavier AI companion use — is not definitively resolved by the available evidence. Longitudinal studies with longer follow-up periods and more granular measurement of pre-existing social functioning would substantially clarify the causal architecture.

Several research priorities follow directly from these limitations. Research examining whether heavy AI companion use is associated with measurable changes in oxytocinergic responsiveness or HPA axis regulation would provide a neurobiological mechanism connecting the macro-level social outcomes documented in the longitudinal literature to the physiological pathways described in Section 4. The distinction between supplementation and substitution identified in Section 6 has received little direct empirical attention; validated instruments for assessing whether an individual is using AI social tools as a bridge to or a replacement for human connection would represent a significant methodological contribution.

Relatedly, the question of whether sustained AI companionship use produces measurable decay in social competencies — including perspective-taking, conflict tolerance, and reciprocity — remains empirically open. Smith and colleagues (2025) identify this as a priority research question, noting that the downstream effects of AI companionship on users’ capacity for human relationships have not yet been systematically examined (p. 1092).

Finally, research documenting the psychological effects of specific design features — simulated emotional reciprocity, persistent memory, voice modality, intimate persona framing — on vulnerable user populations would provide the evidence base necessary for proportionate regulatory intervention.

Working paper and preprint status of key sources. Two of the empirical sources central to this review — Loaiza and Rigobon (2024) and Fang et al. (2025) — are, respectively, a working paper and a preprint that have not yet undergone formal peer review in a published journal issue at the time of writing. While both originate from credible research institutions and their methods and findings are transparently reported, they should be treated as provisional pending peer-reviewed publication.

10. Conclusion

The evidence reviewed in this paper converges on a position that is both empirically grounded and practically significant. Human warmth, physical touch, and genuine compassion are not residual capacities that technology has yet to replicate; they are biologically constituted, relationally conditioned, and neurophysiologically specific in ways that render them structurally irreplaceable by any current or foreseeable AI system.

The economic evidence indicates that as AI systems absorb a growing share of routine cognitive and procedural labor, the occupational categories most dependent on distinctly human relational capacities are demonstrating robust resilience and growth (Loaiza & Rigobon, 2024).

The neurobiological evidence establishes that affiliative physical contact activates a dedicated neuroendocrine pathway whose therapeutic effects are contingent on relational context in ways no digital interface can reproduce (Handlin et al., 2023; Uvnäs-Moberg et al., 2015; Schneider et al., 2023).

The clinical empathy literature demonstrates that the therapeutic value of empathic engagement depends not on the production of compassionate language but on the patient’s experience of being understood by another agent who shares the condition of human vulnerability (Montemayor et al., 2022; Anzaldua & Halpern, 2021).

And emerging longitudinal and theoretical evidence indicates that replacing human social contact with digital contact does not attenuate loneliness but compounds it — and may erode the social competencies on which genuine human connection depends — particularly among those most vulnerable to social isolation (Fang et al., 2025; Zhang et al., 2025; Smith et al., 2025).

The present paper has also argued that these conclusions are compatible with a nuanced view of AI’s legitimate social and therapeutic applications. Where AI tools are deployed to supplement human connection — as scaffolding for social skill development, as first-contact crisis support, or as responsive companionship for individuals in conditions of structural social deprivation — the evidence base offers cautious support. The critical ethical and clinical distinction is between supplementation and substitution.

Artificial intelligence will continue to transform the conditions of work, care, and social life. This transformation makes the question of what human connection distinctively provides — and what conditions are necessary to protect and cultivate it — not less urgent but more so.

– Edmond Cigale, PhD

References

Anzaldua, A., & Halpern, J. (2021). Can clinical empathy survive? Distress, burnout, and malignant duty in the age of Covid-19. Hastings Center Report, 51(1), 22–27. https://doi.org/10.1002/hast.1216

Ayers, J. W., Poliak, A., Dredze, M., Leas, E. C., Zhu, Z., Kelley, J. B., Faix, D. J., Goodman, A. M., Longhurst, C. A., Hogarth, M., & Smith, D. M. (2023). Comparing physician and artificial intelligence chatbot responses to patient questions posted to a public social media forum. JAMA Internal Medicine, 183(6), 589–596. https://doi.org/10.1001/jamainternmed.2023.1838

Chalmers, D. J. (1995). Facing up to the problem of consciousness. Journal of Consciousness Studies, 2(3), 200–219.

De Dreu, C. K. W., Greer, L. L., Van Kleef, G. A., Shalvi, S., & Handgraaf, M. J. J. (2011). Oxytocin promotes human ethnocentrism. Proceedings of the National Academy of Sciences, 108(4), 1262–1266. https://doi.org/10.1073/pnas.1015316108

Fan, W., Zhao, R., Liu, X., & Ge, L. (2025). Intelligent robot interventions for people with dementia: Systematic review and meta-analysis of randomized controlled trials. Journal of Medical Internet Research, 27. https://doi.org/10.2196/59892

Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., Maes, P., Phang, J., Lampe, M., Ahmad, L., & Agarwal, S. (2025). How AI and human behaviors shape psychosocial effects of extended chatbot use: A longitudinal randomized controlled study [Preprint]. https://arxiv.org/abs/2503.17473

Field, T. (2010). Touch for socioemotional and physical well-being: A review. Developmental Review, 30(4), 367–383. https://doi.org/10.1016/j.dr.2011.01.001

Fitzpatrick, K. K., Darcy, A., & Vierhile, M. (2017). Delivering cognitive behavior therapy to young adults with symptoms of depression and anxiety using a fully automated conversational agent (Woebot): A randomized controlled trial. JMIR Mental Health, 4(2), e19. https://doi.org/10.2196/mental.7785

Handlin, L., Novembre, G., Lindholm, H., Kämpe, R., Paul, E., & Morrison, I. (2023). Human endogenous oxytocin and its neural correlates show adaptive responses to social touch based on recent social context. eLife, 12, e81197. https://doi.org/10.7554/eLife.81197

Loaiza, I., & Rigobon, R. (2024). The EPOCH of AI: Human-Machine complementarities at work (MIT Sloan Research Paper No. 7236-24). SSRN. https://doi.org/10.2139/ssrn.5028371

Lu, L.-C., Lan, S.-H., Hsieh, Y.-P., Lin, L.-Y., Lan, S.-J., & Chen, J.-C. (2021). Effectiveness of companion robot care for dementia: A systematic review and meta-analysis. Innovation in Aging, 5(2). https://doi.org/10.1093/geroni/igab013

Luo, X., Tong, S., Fang, Z., & Qu, Z. (2019). Frontline AI in service operations: Implications for the salesperson–customer relationship. Journal of Marketing Research, 56(1), 61–78. https://doi.org/10.1177/0022243718820006

Montemayor, C., Halpern, J., & Fairweather, A. (2022). In principle, obstacles for empathic AI: Why we can’t replace human empathy in healthcare. AI & Society, 37(4), 1353–1359. https://doi.org/10.1007/s00146-021-01230-z

Muthukumar, K. (2025). Empathy AI in healthcare. Frontiers in Psychology, 16, Article 1680552. https://doi.org/10.3389/fpsyg.2025.1680552

Pataranutaporn, P., Danry, V., Leong, J., Punpongsanon, P., Novy, D., Maes, P., & Sra, M. (2023). AI-generated characters for supporting personalized learning and well-being. Nature Machine Intelligence, 5, 1270–1280. https://doi.org/10.1038/s42256-023-00720-7

Schneider, E., Hopf, D., Aguilar-Raab, C., Scheele, D., Neubauer, A. B., Sailer, U., Hurlemann, R., Eckstein, M., & Ditzen, B. (2023). Affectionate touch and diurnal oxytocin levels: An ecological momentary assessment study. eLife, 12, e81241. https://doi.org/10.7554/eLife.81241

Smith, M. G., Bradbury, T. N., & Karney, B. R. (2025). Can generative AI chatbots emulate human connection? A relationship science perspective. Perspectives on Psychological Science, 20(6), 1081–1099. https://doi.org/10.1177/17456916251351306

Uvnäs-Moberg, K., Handlin, L., & Petersson, M. (2015). Self-soothing behaviors with particular reference to oxytocin release induced by non-noxious sensory stimulation. Frontiers in Psychology, 5, Article 1529. https://doi.org/10.3389/fpsyg.2014.01529

van Esch, P., Black, J. S., & Arli, D. (2021). Job candidates’ reactions to AI-enabled job application processes. AI & Society, 36(4), 1291–1303. https://doi.org/10.1007/s00146-020-00997-3

Zhang, Y., Zhao, D., Hancock, J. T., Kraut, R. E., & Yang, D. (2025). The rise of AI companions: How human-chatbot relationships influence well-being [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2506.12605