The Human Health Crisis from Uncertainty in the AI Era
The Dual Crisis of Input Overload and Regulatory Failure:
A Systems Dynamics Model from Evolutionary Mismatch to Socio-Cognitive Load
From Stone Age Brains to Algorithmic Overwhelm —
How AI-Era Uncertainty May Cascade into Chronic Disease
This paper proposes a “dual crisis” model of input overload and regulatory failure, exploring how social uncertainty in the AI era may increase human risk for multiple chronic diseases through evolutionary mismatch mechanisms. The “input side” of the core hypothesis posits that the AI era may significantly amplify the load on the social threat system across five dimensions (social evaluation, occupational-economic, cognitive-informational, identity, and relational uncertainty), with occupational-economic uncertainty identified as a potentially upstream primary driver. The “regulatory side” of the core hypothesis argues that traditional uncertainty-buffering mechanisms (religious faith, stable family structures, lifelong occupational identity, national identity consensus, and trust in scientific authority) are simultaneously weakening in the AI era — creating a pincer movement between “signal overload” and “buffer failure.” This dual burden, transmitted via the HPA axis, sleep disruption (identified as a core mediating variable), vagal nerve–gut-brain axis, and inflammatory pathways, may increase the risk of multiple psychosomatic diseases. At the same time, this paper argues that AI exerts a dual-phase effect on human health: it functions as a stress amplifier when it increases evaluative density and authenticity uncertainty, and as a stress buffer when it provides predictable support and reduces cognitive load. These effects are unevenly distributed along the socioeconomic status (SES) gradient. Distal outcomes such as cancer are presented as higher-order hypotheses. Based on multiple rounds of external review, this paper has completed an epistemological iteration from causal declaration to stratified hypotheses to a systems dynamics model.
Introduction: The Inversion of CertaintyWhere survival is assured but selfhood is under siege
For most of human history, the sources of uncertainty were rooted in the natural world — predators, famine, plague, extreme weather. The brain’s threat detection system (the circuit centered on the amygdala) evolved precisely to handle these acute, concrete, and actionable natural threats. When you face a tiger, the stress response activates, you fight or flee, the event ends, and the body recovers.
Industrialization completely upended this equation. In a modern city, no tiger chases you — but performance reviews do. Your dinner arrives by delivery in 30 minutes, but you do not know whether you will still have a job next month. Survival is certain, but life is uncertain. The body is safe, but the self is insecure.
The AI era pushes this “inversion of certainty” to its extreme. When environmental survival uncertainty approaches zero while social-environmental uncertainty continues to climb, the brain’s threat detection system does not shut down — it may redirect more processing resources toward social uncertainty, deploying ancient circuits designed for natural threats to process a signal type for which they were never optimized.
A Five-Dimensional Taxonomy of Uncertainty
To transform “uncertainty” from a vague umbrella concept into a researchable set of variables, this paper operationalizes AI-era social uncertainty into five dimensions:
| Dimension | Core Question | AI-Era Amplification Mechanism |
|---|---|---|
| Social Evaluation Uncertainty | Am I accepted, approved of, or rejected? | Real-time evaluative loops on social media (likes / follows / unfollows) |
| Occupational-Economic Uncertainty | Are my skills, income, and position secure? | AI automation replacement anxiety; WEF projects 41% of employers will downsize workforces by 2030 |
| Cognitive-Informational Uncertainty | Can I determine the veracity, source, and meaning of information? | Explosion of AI-generated content; deepfakes; information and misinformation becoming indistinguishable |
| Identity Uncertainty | Am I a creator, an executor, a displaced worker, or an AI collaborator? | AI erodes traditional occupational identities; human–machine boundaries become blurred |
| Relational Uncertainty | Am I interacting with a real person or with algorithmic / AI-generated feedback? | AI chatbots substituting for human relationships; “technology-induced shared psychotic features” |
These five dimensions are neither mutually independent nor strictly parallel; they may exist in a hierarchical driving relationship. Preliminary analysis suggests that occupational-economic uncertainty may serve as the upstream primary driver: fear of job loss drives status anxiety (social evaluation) → destabilizes occupational identity (identity uncertainty) → leads to excessive information consumption in search of a sense of control (cognitive-informational overload) → undermines the capacity and willingness to invest in authentic relationships (relational uncertainty). This hierarchical hypothesis requires testing through structural equation modeling or longitudinal panel data, but it provides a framework with greater predictive power and intervention specificity than a “five parallel dimensions” model.
AI-era social uncertainty possesses a unique combination of characteristics distinct from previous epochs: high-frequency, persistent, individualized, algorithmically driven, epistemically unstable, and identity-boundary-blurring. Whether this combination causes the total load to exceed historical peaks (such as during wars, famines, or pandemics) still requires cross-era comparative research. But its distinctiveness lies in this: it is not an intermittent acute crisis but a persistent chronic infiltration.
We humans are smart enough and long-lived enough to have created entire worlds of stress inside our own heads.
Evolutionary Mismatch: A Stone Age Brain in the AI EraWhen ancestral hardware meets algorithmic inputs
The Evolutionary Mismatch Hypothesis was systematically articulated by Li, van Vugt, and colleagues (2018) in Current Directions in Psychological Science. Its core proposition: human psychological mechanisms are evolved adaptive products, originally designed to process environmental inputs and generate behavior outputs favorable to survival. But the modern environment has diverged dramatically from the environment in which these mechanisms evolved — many inputs have changed in quantity and intensity, or no longer carry the same adaptive associations, causing these mechanisms to produce maladaptive outputs.
Over 99% of human evolution occurred in small-scale hunter-gatherer groups of 50–150 individuals. These groups had no bosses, production quotas, or pension plans, and no rigid separation between work and private life. Only after the agricultural revolution — the final 1% of human evolutionary history — did human societies grow dramatically in scale and complexity. Giphart & Van Vugt (2018), in their book Mismatch, argue that this growth produced many social arrangements harmful to human psychological well-being.
The Brain’s “Hardware Specifications”
Robin Dunbar’s (1998) Social Brain Hypothesis proposes that primate brains evolved not to solve ecological problems but to navigate social complexity. A species’ neocortex size correlates directly with its social group size. For humans, this upper limit is approximately 150 individuals (Dunbar’s number), validated across 23 studies of personal social networks and ethnographic communities, spanning diverse cultures and historical periods over the past 2,000 years, with a maximum sample size of 61 million individuals (Dunbar, 2020).
The average size of a Neolithic English village was 160 people, the average Roman military unit was 150, and even modern Christmas card networks max out at 153. This is not coincidence — it is the repeated expression of the hardware ceiling of the neocortex’s processing capacity.
Analysis of a 1.7-million-user Twitter dataset (Gonçalves et al., 2011) further validated this ceiling: even in the digital world, users can maintain at most 100–200 stable relationships. The “attention economy” in the online world is constrained by the cognitive and biological limits predicted by Dunbar’s theory. The brain cannot be upgraded — no matter how many apps you install.
The Great Migration of Threat: From Nature to SocietyWhen the brain treats a performance review like a predator
Eisenberger, Lieberman, and Williams (2003) published a landmark study in Science using fMRI to demonstrate that brain activation during social exclusion partially overlaps with activation during physical pain in the dorsal anterior cingulate cortex (dACC) and anterior insula — social threat can recruit neural networks associated with pain, salience detection, emotional distress, and threat monitoring.LEVEL A · Peer-Reviewed fMRI
Eisenberger and Lieberman (2004) proposed an evolutionary explanation in Trends in Cognitive Sciences: across mammalian species, the social attachment system co-opted the computational processes of the pain system to prevent the potentially harmful consequences of social separation. A precise formulation is necessary: social pain and physical pain exhibit partial neural overlap, not neural equivalence — they share certain processing pathways but also have distinct components. The amygdala’s response to social evaluative threat differs measurably from its response to physical threat.LEVEL A · Systematic Review
Social Pain Outlasts Physical Pain
Meyer, Williams, and Eisenberger (2015) published findings in PLOS ONE revealing a critical asymmetry: the experience of social pain can be re-experienced or “relived” long after the painful event has passed, whereas physical pain cannot be easily relived once the painful event subsides. This means social threat possesses an “infinite replay loop” quality — you will relive a humiliation at 3 a.m., and each time the brain responds as if it were happening for the first time.
Social Evaluative Threat: No Wound, but the Immune System Is Already Mobilized
The meta-analysis by Dickerson and Kemeny (2004) found that what elicits the strongest cortisol response is not physical pain or cognitive challenge, but social evaluative threat — the experience of being scrutinized and judged by others. The body is in no physical danger whatsoever, yet the entire stress response system is running at full capacity. The brain has not yet evolved the ability to distinguish between “a tiger is chasing you” and “your boss is evaluating you.” To the brain, both are threats.
An important clarification: social threat can recruit neural circuits that partially overlap with those activated by physical threat (dACC, anterior insula), causing social anxiety to produce stress responses at the neurophysiological level similar to those elicited by physical threat. However, “similar” does not mean “identical” — the multivoxel pattern analysis by Woo et al. (2014) demonstrated that social exclusion and physical pain, while activating the same brain regions, exhibit distinguishable fine-grained representational patterns. The brain “borrows” pain circuitry to process social threat, but does not run it in an identical manner.
The Cascade Collapse: A Chain Reaction from Nervous to Digestive to Immune SystemsHow social signals become somatic disease
Slavich and Irwin (2014) proposed the “Social Signal Transduction Theory of Depression” in Psychological Bulletin, which is the most important theoretical framework for understanding this causal chain. Core hypothesis: experiences of social threat and adversity directly upregulate the expression of genes involved in inflammation within the immune system. The innate immune system’s “Conserved Transcriptional Response to Adversity” (CTRA) — originally developed to combat physical threats from predators and hostile conspecifics — is now activated by social conflict, evaluation, exclusion, and isolation in the contemporary social environment.
Three-Tier Evidence Stratification
Social stress → HPA/SNS activation → changes in sleep, mood, and inflammatory markers. Supported by Slavich and Irwin’s (2014) Social Signal Transduction Theory, the 30-year meta-analysis by Segerstrom and Miller (2004), and the cortisol meta-analysis by Dickerson and Kemeny (2004).LEVEL A
Chronic stress → intestinal barrier / microbiome / immune dysregulation. Mechanistic evidence is provided by the 2023 Nature study on brain–gut circuits and the vagal nerve review by Bonaz et al. (2018), but much of the key data comes from animal models, and longitudinal human evidence is still accumulating.LEVEL B
Chronic inflammation and immune dysregulation may influence risk pathways for certain chronic diseases and cancer. However, cancer is a multifactorial disease (genetic mutations, environmental carcinogens, aging, viruses, stochastic DNA replication errors) and cannot be reduced to a linear “social stress → cancer” causal chain. This tier should be treated as a hypothesis requiring longitudinal cohort study validation, not as an established conclusion.LEVEL C
A landmark 2023 Nature paper, “Stress-activated brain-gut circuits disrupt intestinal barrier integrity,” further confirmed how psychological stress disrupts intestinal barrier function via brain–gut circuits. The gut houses approximately 70% of the body’s immune cells, and impairment of its barrier function may allow pro-inflammatory antigens and microbial compounds to translocate from the intestinal lumen into the bloodstream. A 2025 review in the International Journal of Biological Sciences concluded that chronic stress persistently reduces microbial diversity, depletes beneficial bacteria, and promotes the growth of pro-inflammatory microbes, potentially amplifying systemic inflammation and disrupting immune regulation.
Sleep — The Core Mediating Pathway from Social Stress to Somatic Disease
Within the cascade pathway described above, there exists a previously underappreciated core mediating variable: sleep. Sleep may be the most important, most measurable, and most intervenable bridge through which AI-era social uncertainty enters the body.LEVEL A
The mechanistic chain: social uncertainty → anticipatory rumination → difficulty falling asleep / sleep fragmentation → cortisol rhythm disruption → immune dysregulation → elevated inflammatory markers → metabolic disturbance. The meta-analysis by Segerstrom and Miller (2004) demonstrated that one of the most robust immune effects of chronic stress is mediated precisely through sleep disruption. The AI era presents a unique problem: AI content generation operates 24/7, algorithmic push notifications remain active late at night, blue light exposure suppresses melatonin, and the ruminative thinking produced by occupational replacement anxiety is particularly active during nighttime hours — it is precisely the “you don’t know whether your job will still exist tomorrow” type of unactionable uncertainty that most potently erodes sleep. The public health significance of elevating sleep to a core mediating variable lies in this: it is the most easily measured (polysomnography, wearable devices) and most easily targeted for intervention (sleep hygiene, Cognitive Behavioral Therapy for Insomnia [CBT-I], light exposure management) node in the entire cascade pathway.
A methodological caveat is warranted: chronic inflammation is a confirmed risk factor for multiple chronic diseases, including cardiovascular disease, metabolic syndrome, and autoimmune disorders. Epidemiological associations also exist between chronic inflammation and certain cancer types, but the pathogenesis of cancer is an extraordinarily complex multifactorial network. The pathway from social stress to a specific cancer traverses multiple causal levels, the effect size at each step has not been precisely quantified, and it depends on interactions among cancer type, immune context, genetic susceptibility, and environmental exposures. This pathway should be understood as a distal hypothesis requiring longitudinal cohort study validation.
Universe 25: A Laboratory Preview of Social ExtinctionWhen all threats are removed except the social kind
John B. Calhoun’s (1968–1973) Universe 25 experiment at the National Institute of Mental Health (NIMH) constitutes an animal experimental analog of the theoretical framework presented above. The experimental conditions precisely replicated this paper’s core proposition: eliminate all natural threats (unlimited food, constant temperature, no predators, no disease), leaving social density as the sole variable.
The Inversion of Certainty and Uncertainty
Universe 25 featured a highly certain survival environment paired with highly uncertain population dynamics and social structure. The initial 8 breeding pairs enjoyed low density and low social-relational complexity. But as the population doubled every 55 days, the number of social relationships each mouse needed to track grew exponentially — far exceeding the processing capacity of their social cognitive systems.
Key findings: after the social hierarchy collapsed, traditional mouse social roles disintegrated. Males could no longer defend territories or establish reproductive hierarchies. Maternal care collapsed, and infant mortality soared above 90%. Social interaction overload — too many social contacts occurring too frequently and unpredictably — prevented the mice from forming stable social hierarchies or engaging in reproductive behavior.
Mice born into the chaotic environment never learned complex social behavior, remaining trapped in a juvenile social state. Calhoun termed this phenomenon “Death Squared” — the death of the spirit preceding the death of the body. Necropsies revealed adrenal hypertrophy — the hallmark physiological marker of chronic HPA axis overload.
The last known conception occurred on Day 920, when the population stood at 2,200 — far below the enclosure’s capacity of 3,000. Unlimited food, water, and resources remained available, but none of that mattered anymore. Their spirits had died — they were no longer capable of executing the complex behaviors compatible with species survival.
The most profound lesson: even during the population decline phase, social uncertainty did not decrease — because the social structure had completely collapsed, each death altered the topology of the remaining network, and the social cognitive hardware of surviving individuals had been permanently damaged by chronic overload. Even “Beautiful Ones” relocated to entirely new environments failed to recover normal social and reproductive behavior.
The Limits of the Analogy: Physical Density ≠ Digital Isolation
The structural limitations of the Universe 25 analogy must be acknowledged: the core destructive variable in Calhoun’s experiment was ultra-high-density physical contact and spatial deprivation — mice were forced to encounter excessive numbers of conspecifics in physical space. By contrast, the core destructive variable facing humans in the AI era is informational overload and existential anxiety at the virtual level — humans often endure digitally mediated social stress in physical isolation. “Contact overload in crowding” and “informational overload in isolation” are fundamentally different in behavioral terms. Furthermore, Kessler (1966) failed to observe the same population collapse at even higher densities in replication experiments, and the medical historian Ramsden has argued that the collapse may have been driven by “excessive social interaction” rather than density per se. Therefore, Universe 25 should be understood in this paper as a heuristic structural analogy — the two share the abstract mechanism of “social signals exceeding cognitive bandwidth” — rather than as a direct causal prediction. Humans and rodents differ fundamentally in cognitive complexity, cultural transmission, and institutional capacity, and simple one-to-one correspondences should not be drawn.
From Village to AI: Four Compressions of Attention BandwidthHow the signal-to-noise ratio collapsed across four epochs
Georg Simmel first diagnosed this problem in 1903 in “The Metropolis and Mental Life”: the defining characteristic of the metropolis is its “rapid crowding of changing images, the sharp discontinuity in the grasp of a single glance, and the unexpectedness of onrushing impressions” — bombarding the individual with dense, fleeting stimulation until emotional reactivity is exhausted. Simmel’s “blasé attitude” is the human version of Universe 25’s “Beautiful Ones.” Milgram (1970) formally proposed the “urban overload hypothesis” in Science, and Lederbogen et al. (2011) confirmed in Nature via fMRI that the amygdala of urban residents is significantly more reactive under social stress.
| Era | Daily Social Signal Volume | Rate of Change | Amygdala Load | Adaptive Strategy |
|---|---|---|---|---|
| Village Era | ~150-person stable network | Generational timescale | Baseline | Not needed |
| Urbanization Era | Thousands of “familiar strangers” | Annual / monthly timescale | Elevated (Lederbogen 2011) | Simmel’s “blasé attitude” |
| Algorithmic Recommendation Era | Infinite feeds, interruptions every 6 min | Hourly / minute timescale | Persistent overclocking | Brain rot / information avoidance |
| AI Era | AI-generated content explosion + existential threat | Accelerating in real time | Risk may be elevated (exploratory) LEVEL C | To be observed |
Each layer does not replace the previous one but stacks on top of it. Contemporary humans simultaneously bear the social stimulus overload of urban density, the attentional hijacking of algorithmic recommendation, and the existential uncertainty brought by AI. Three layers of stress act simultaneously on the same brain that was shaped in a Neolithic village.
Operationalizing “Attention Bandwidth Overclocking”
“Attention bandwidth overclocking” as a concept needs to be upgraded from metaphor to a researchable framework. To this end, this paper decomposes it into five measurable sub-dimensions:
| Sub-Dimension | Operationalized Indicator | Evidence Source |
|---|---|---|
| Social Relationship Maintenance Capacity | Number of stable social relationships (Dunbar layers: 5/15/50/150) | Dunbar Social Brain Hypothesis LEVEL A |
| Attentional Switching Frequency | Context switches per hour; recovery time after interruption (~25 minutes) | Gloria Mark, attention research LEVEL B |
| Social Threat Monitoring Load | Amygdala social stress response intensity (fMRI measurement) | Lederbogen et al. 2011 LEVEL A |
| Information Veracity Judgment Load | Proportion of AI-generated content; cognitive resources required for veracity assessment | Emerging field LEVEL D |
| Cortisol Recovery Capacity | Post-stress cortisol return-to-baseline time; HPA axis negative feedback efficiency | McEwen, allostatic load theory LEVEL A |
It must be emphasized that Dunbar’s number primarily describes the cognitive ceiling for maintaining stable social relationships and does not directly equate to the total processing ceiling for information feeds, news, AI content, short-form video, work tasks, and occupational uncertainty. “Attention bandwidth overclocking,” as an integrative concept, must go beyond Dunbar’s number itself and construct a more generalized “socio-cognitive load model” — Dunbar’s number can serve only as one piece of evidence for “finite social cognitive capacity,” not as sole proof of AI-era overload in its entirety.
Regarding how the five sub-dimensions integrate into a unified “bandwidth” metric, this paper acknowledges that this remains an untested modeling choice. Possible models include: a bottleneck model (the weakest sub-dimension determines the overall load ceiling), a weighted additive model (dimensions accumulate linearly with assigned weights), or an interaction model (certain dimensional combinations produce supra-additive effects). Which model is more accurate requires experimental designs with simultaneous multi-dimensional measurement. At this stage, the five sub-dimensions should be understood as “multiple operationalization entry points converging on the same construct,” rather than an already-integrated scale.
The New Spectrum of AI-Era Mental DisordersEmerging psychopathologies the diagnostic manual has not yet caught
A landmark 2025 paper published in Frontiers in Digital Health, “Redefining Psychopathology in the Context of Digital Overload,” was the first to propose a taxonomy of 19 novel digital-era mental disorders, organized into four diagnostic categories: cognitive fragmentation and digital overload disorders, social media and immersive technology-induced disorders, technology integration disorders, and symptom-driven disorders requiring diagnostic adaptation.
Important caveat: the following diagnostic categories are currently at the descriptive phenomenological stage or early provisional diagnostic exploration and have not been formally incorporated into mainstream psychiatric classification systems such as the DSM-5-TR or ICD-11. They represent emerging symptom patterns observed at the clinical frontline, not formally validated disease classifications that have undergone complete reliability and validity testing. This paper cites these concepts to present the state of the art in AI-era mental health, not to treat them as definitive medical diagnoses.
| Diagnostic Name | Abbreviation | Clinical Recognition Rate |
|---|---|---|
| Continuous Partial Attention Disorder | CPAD | 85.3% |
| Digital Anxiety Disorder | DAD | 82.7% |
| Doomscrolling Disorder | DD | 78.7% |
| AI Identity Diffusion Disorder | AIDD | Statistically significant |
| AI-Induced Psychosis | AIP | Clinical cases increasing |
| Algorithmic Dependency Disorder | ADD | Emerging diagnosis |
| Notification Hypervigilance Syndrome | NHS | Emerging diagnosis |
76% of cases (171 of 225) documented symptoms consistent with digital-era psychopathology, with CPAD alone appearing in 36.4% of medical records. These are not imagined illnesses — they are flooding into clinics.
The American Psychiatric Association’s official publication, Psychiatric News, published a feature report on AI-induced psychosis in 2025. Keith Sakata, a psychiatrist at the University of California, San Francisco, reported 12 patients hospitalized for AI-related psychosis in August 2025. Innovations in Clinical Neuroscience published the first clinically validated case of new-onset AI-related psychosis.
Statistical Evidence: An Accelerating CrisisThe numbers behind the epidemic of stress
(2025 composite study)
(APA 2024; up from 32% in 2022)
significantly impacts stress (APA 2025)
stress increase (APA 2025)
(1990–2019, reaching 301 million)
(17 years earlier than avg. age 42)
Psychosomatic Syndromes of the AI Era
Technostress research has systematically documented AI-related somatization symptoms: anxiety, mental fatigue, depression, and sleep problems as psychological symptoms; gastrointestinal disorders, hypertension, irritable bowel syndrome, headaches, neck and shoulder pain, and insomnia as somatic symptoms (Molino et al., 2023; KDVI, 2024). Heavy users of AI chatbots report greater loneliness, dependency, and less real-world social interaction (Fang et al., 2025, four-week RCT).
Global productivity losses due to mental health problems amount to $438 billion (Gallup, 2025). Depression is projected to become the world’s largest disease burden by 2030. This is no longer an individual psychological problem — it may be a systemic health risk escalation driven by social structure.
Data quality disclosure: the statistics cited above come from diverse sources with varying evidence levels. APA annual surveys and WHO Global Burden of Disease data are Level A (large-sample, peer-reviewed, representative sampling). The Mind Share Partners and Workhuman corporate surveys are Level C–D (corporate-commissioned, potentially non-representative samples, potentially leading question phrasing). Figures such as “82% burnout risk” and “age 25 burnout peak” require independent verification of original sample sizes, country coverage, and survey methods before citation. This paper presents these figures to indicate directional trends, not to provide precise epidemiological data.
Conclusion: Dual Crisis and Dual-Phase EffectsThe complete picture — neither doom nor salvation
This paper has traced a hypothesized causal chain from evolutionary biology to AI-era mental disorders. Its core architecture is the dual crisis model:
The AI era, through five-dimensional uncertainty (with occupational-economic uncertainty as the potential primary upstream driver), may significantly amplify the load on the social threat system. This load, transmitted via the HPA axis, sleep disruption (core mediator), and inflammatory pathways, increases the risk of anxiety, depression, and insomnia.
Traditional uncertainty-buffering mechanisms (religion, family, occupational identity, national identity, scientific authority) are simultaneously weakening in the AI era. The pincer movement of input overload and regulatory failure may produce supra-additive effects.
Chronic immune dysregulation may influence certain chronic disease risk pathways. However, distal outcomes depend on multifactorial interactions and should serve as hypothesis directions for longitudinal cohort studies.
AI exerts both amplifier and buffer effects on human health simultaneously. It functions as an amplifier when it increases evaluative density and authenticity uncertainty; it functions as a buffer when it provides predictable support and reduces cognitive load. These effects are unevenly distributed along the SES gradient — the most vulnerable populations bear the greatest risk and possess the fewest buffering resources.
The value of this paper lies not in claiming that “AI will destroy human health,” but in arguing that the structure of social uncertainty in the AI era may significantly amplify the load on the stress systems shaped by human evolution, while traditional buffering mechanisms are simultaneously failing — this dual-crisis hypothesis warrants serious scientific investment, scrutiny of technology design, and public health policy attention. The most urgent action is not to wait for perfect causal proof, but to begin intervening at the lowest-cost, highest-certainty nodes — protecting sleep, managing notifications, screening for technostress — intercepting the earliest links in the cascade pathway before more ambitious institutional reforms are completed.
The Dual Crisis: Input Overload and Regulatory FailureWhen signal overload meets buffer collapse
The preceding sections focused on the “input side” — how the AI era increases the quantity and rate of change of social uncertainty signals. But this is only half of the crisis. This section introduces a critical structural dimension: humanity’s traditional uncertainty-buffering mechanisms are simultaneously weakening in the AI era — creating a “dual pincer” of input overload and regulatory failure.
The Simultaneous Failure of Traditional Buffers
Throughout history, humans have developed multiple institutional tools for reducing uncertainty. Their essential function is to transform an unpredictable world into a predictable psychological space, thereby freeing cognitive resources for other purposes. But these tools are facing unprecedented simultaneous erosion in the AI era:
| Buffer | Traditional Function | AI-Era Erosion Mechanism |
|---|---|---|
| Religious Faith | Provides an ultimate meaning framework, reducing existential uncertainty | Continued secularization; science and AI further compress the space for supernatural explanations |
| Stable Family Structure | Provides predictable intimate relationships and a social support baseline | Increasing atomization; rising rates of single-person households; AI companions substituting for some emotional functions |
| Lifelong Occupational Identity | “I am a teacher / doctor / engineer” provides stable self-definition | AI automation disrupts occupational boundaries; skill half-lives shrink dramatically |
| National Identity Consensus | Shared collective identity reduces “who are we” uncertainty | Algorithmic push intensifies polarization; consensus space fragments |
| Trust in Scientific Authority | “Experts know the answer” reduces cognitive-informational uncertainty | AI-generated content blurs the boundary between authority and noise; post-truth environment erodes trust |
| Community Belonging | Dunbar-scale acquaintance networks provide social predictability | Urban atomization + digital socialization replaces face-to-face interaction |
When input-side uncertainty is increasing (AI amplifies five-dimensional social threat) while regulatory-side buffering capacity is simultaneously declining (traditional uncertainty-reducing institutional tools are failing), the combined pincer may produce supra-additive effects — the net load borne by the system is not the sum of the two components but potentially greater than their sum. This is the complete structure of health risk in the AI era.
The Dual-Phase Effect Model of AI: Amplifier and Buffer
However, the intellectual honesty demanded by scientific argumentation requires acknowledging that AI’s effect on human health is not unidirectional. It simultaneously exhibits both stress-amplifying and stress-buffering modes:
| AI as Stress Amplifier | AI as Stress Buffer |
|---|---|
| Increases social evaluative density (like/follow cycles) | Provides low-risk interaction practice space for the socially anxious |
| Creates occupational replacement anxiety (WEF: 41% workforce reduction) | Reduces repetitive cognitive labor, theoretically freeing creative resources |
| AI-generated content explosion → information veracity becomes indeterminate | LLM summarization functions are reducing information entropy in some domains |
| Blurs human–machine boundaries → identity uncertainty | AI-assisted diagnostics reduce medical uncertainty (Level B evidence) |
| AI companions substitute for real social interaction → deepening loneliness | Provides 24/7 support for marginalized groups (chronic illness / disability / immigrants) (Level C evidence) |
Research from HBS in 2025 found that interaction with AI companions can reduce loneliness to levels comparable to human interactionLEVEL B; yet a four-week RCT by Fang et al. (2025) found that heavy users experienced greater lonelinessLEVEL B. This contradiction is itself evidence of the dual-phase effect: dosage and usage patterns determine whether AI falls on the “amplifier” or “buffer” side.
The Class Gradient — The Unequal Distribution of AI Risk
Marmot’s Whitehall studies long ago demonstrated that health risks are distributed along a social class gradient. The uncertainty risks of the AI era follow the same pattern, and may further exacerbate inequality:
The elite class that commands capital and AI computing power is not only less exposed to AI replacement threats but actually gains productivity advantages from AI tools — for them, AI is a buffer. Blue-collar workers and entry-level white-collar employees face direct job displacement, skill devaluation, and economic insecurity — for them, AI is an amplifier. Attention bandwidth overclocking also exhibits a class gradient: those with financial resources can purchase “digital detox vacations,” private psychological counseling, and low-stimulation living environments; those lacking economic resources are forced into continuous exposure to high-density digital social signals with no opt-out. The distribution of health risk in the AI era may recapitulate Marmot’s findings: those at the bottom bear the greatest risk, and the health gradient across society widens further.
Neuroplasticity: The Circuits Can Still Be Recalibrated
The analysis above should not be read as fatalism. Andy Clark’s (2015) predictive processing framework and Eisenberger’s (2015) research both demonstrate that anxiety circuits can be recalibrated through neuroplasticity — CBT, mindfulness-based stress reduction, and exposure therapy are effective in principle.LEVEL A Humans possess the capacity to build cultural scaffolding, and historically every major technological shock has ultimately spawned new adaptive institutions. But “can be recalibrated” does not mean “is being recalibrated” — institution-building takes time, and AI’s acceleration waits for no one.
The complete picture presented in this paper is not “AI will destroy humanity” nor “AI will save humanity,” but rather: the health risks of the AI era stem from the dual structure of input overload and regulatory failure; AI itself is simultaneously a participant in both processes — it increases uncertainty along some dimensions while decreasing it along others; and the entire effect is unevenly distributed along the class gradient. The most vulnerable populations bear the greatest risk and possess the fewest buffering resources.
Breaking the Loop: An Action Framework from Individual to InstitutionalThree tiers of intervention, three levels of feasibility
Based on the dual crisis model presented in this paper, concrete intervention pathways can be proposed across three levels and three priority tiers. The core principle is: reduce the information entropy of social signals on the input side, repair buffering mechanisms on the regulatory side, and restore the recovery cycle of attention bandwidth.
| Priority | Type | Specific Measure | Feasibility Assessment |
|---|---|---|---|
| Tier 1 Immediate Pilot |
Individual | Aggressive notification permission reduction (each notification source eliminated ≈ 25 min saved in cognitive recovery cost) | Zero cost, immediately executable |
| Individual | Sleep protection protocol: no screens 90 minutes before bed, no devices in the bedroom, fixed sleep schedule | CBT-I has Level A evidence support | |
| Institutional | Incorporate “technostress screening” into routine occupational health examinations | Low cost, can be embedded in existing check-up workflows | |
| Tier 2 Requires Policy Assessment |
Technical | “Attention Impact Assessment” (AIA) for algorithmic recommendation systems, analogous to Environmental Impact Assessment | Requires legislative framework; EU AI Act as potential reference |
| Institutional | Prohibit infinite scroll designs targeting minors; mandatory interaction limits for AI chatbots | Involves balancing product ethics and user autonomy | |
| Tier 3 Theoretical Advocacy |
Technical | Adopt “minimum user time spent” rather than “maximum engagement” as the core product KPI | Conflicts with platform business models; requires market incentive restructuring |
| Institutional | Establish an AI-era mental health epidemiological surveillance system (WHO-level) | Requires international coordination and long-term funding commitment |
The shared characteristic of Tier 1 measures is that the evidence is relatively robust, cost is low, and they can be initiated without relying on institutional reform. If the hypothetical framework of this paper is correct, then the most urgent intervention is not to change AI technology itself, but to protect humanity’s most vulnerable physiological node — sleep — and the most immediately actionable behavioral node — notification management. The cost of these two interventions is virtually zero, yet they may interrupt the cascade pathway at its earliest links.
Limitations and Methodological DisclosureWhat this paper can and cannot claim
Evidence Tier Definitions
| Tier | Definition | Application in This Paper |
|---|---|---|
| LEVEL A | Meta-analyses, systematic reviews, large-sample longitudinal studies, human experiments replicated in independent samples | Eisenberger fMRI, Segerstrom meta-analysis, Lederbogen Nature, Dunbar 23-study validation |
| LEVEL B | Peer-reviewed experimental studies, fMRI studies, medium-sample human RCTs, peer-reviewed animal mechanistic studies | Nature 2023 brain–gut circuit, Fang 2025 AI-RCT, Bonaz vagal nerve review |
| LEVEL C | Pilot studies, cross-sectional surveys, case series, early clinical observations, small-sample mixed methods | Frontiers 2025 digital psychopathology taxonomy (75 clinicians), AI user satisfaction surveys |
| LEVEL D | Corporate reports, industry surveys, media reports, theoretical extrapolation, expert commentary | Atlassian developer survey, Workhuman corporate report, AI cognitive offloading theoretical extrapolation |
| Limitation | Description | Mitigation |
|---|---|---|
| Causal Chain Closure | Inferential leap from plausible mechanisms to established causation | Three-tier hypothesis stratification; structured dual crisis model |
| Conceptual Operationalization | Core variables require more precise definitions | Five-dimensional taxonomy + five sub-dimensions + driver hierarchy + integrated model declaration |
| Countervailing Evidence | Discussion of positive AI effects | “Dual-phase effect” model integrating positive and negative evidence |
| Class Differences | Unequal distribution of AI risk | SES gradient analysis |
| Regulatory-Side Analysis | Argument for traditional buffer failure | “Simultaneous weakening of traditional buffers” core argumentation |
| Sleep as Mediator | Core bridge variable needing elevation | Sleep as core mediator in the cascade pathway |
| Dunbar Critique | Lindström et al. (2021) methodological critique of Dunbar’s number (95% CI: 4–520) | Cited opposing literature; acknowledged Dunbar’s number as one piece of evidence, not sole pillar |
| Residual Determinism | Overall narrative remains primarily risk-alerting | Paper positioning: systemic risk alert, not balanced review |
The strongest version of this paper does not claim that “AI will destroy human health,” but rather argues that “the structure of social uncertainty in the AI era, under the dual operation of input-side overload and regulatory-side failure, may significantly amplify the load on the stress systems shaped by human evolution — this hypothesis warrants serious scientific investment and policy attention.” All inferences should be subjected to testing in subsequent longitudinal cohort studies, randomized controlled trials, and epidemiological surveillance.