Cognitive and Behavioral
Rigidity in Humans

I. Introduction: The Overlooked Foundational Variable

Discussion of artificial intelligence’s impact on human society has accumulated extensive literature across the technological dimension (exponential growth in model capabilities), the economic dimension (labor displacement and job restructuring), and the ethical dimension (bias, alignment, and governance). Yet a more foundational variable—the architectural characteristics of the human brain itself—has long occupied a marginal position in these discussions.

The core thesis of this paper is: the most fundamental challenge facing human society in the AI era lies not in the rate of AI capability growth, but in the structural contradiction between the human brain’s tendency toward rigidity and the cognitive flexibility demanded by AI. More precisely:

What the AI era eliminates is not “unintelligent people,” but those unable to release themselves from experiential lock-in.

Schultz and Searleman (2002), in a meta-analysis spanning a century of research, described “rigidity” as a multidimensional construct encompassing the tendency to form and persistently employ mental sets and behavioral sets^[1]. However, this traditional psychological construct has never been systematically integrated with contemporary neuroscience research on habit circuits, evolutionary psychology’s mismatch theory, and the social stratification dynamics of the AI era within a single explanatory framework. This paper attempts to accomplish that integration.

II. Methodological Statement

This paper is a theory paper employing abductive reasoning as its primary methodology—that is, starting from facts scattered across different disciplines and retroductively inferring the best unified causal structure that simultaneously explains all of them. Abductive reasoning is a core component of scientific methodology; McMullin (1992) called it “the inference that makes science”^[2], and Curiel (2019) argued that it is at least as important as standard deductive and inductive forms^[3].

The cognitive production process of this paper is itself an instance of its methodology: the human researcher used abductive reasoning to identify cross-domain probing directions, the AI system (Claude Opus 4.6) used inductive retrieval to supply empirical fragments from various fields, and the human researcher then used abductive reasoning to reassemble the fragments into a unified framework. This “adversarial human-AI collaboration” process runs throughout the paper. A point of transparency must be stated: the “Matrix Cognizer” concept defined in later chapters has a self-referential relationship with the paper’s production process—the production process can be viewed as an instance of the concept. This is both a feature and a limitation of the methodology: a single case can serve as the heuristic source for a concept but cannot substitute for independent empirical validation of that concept.

To ensure epistemic transparency, the entire text employs a three-tier marking system:

III. The Neurobiological Foundations of Cognitive and Behavioral Rigidity

3.1 Energy Economics: Rigidity as Optimal Design

ESTABLISHED Ali et al. (2022) demonstrated that when recurrent neural networks operate under energy-minimization constraints in predictable environments, predictive coding mechanisms emerge as a natural consequence of energy efficiency optimization—the networks self-organize into prediction units and error units, learning to suppress predictable sensory inputs^[4]. The brain’s rigidity tendency is not a design flaw but an efficient adaptive strategy in an energy-constrained information processing system.

ESTABLISHED Barton (2014) argued from an evolutionary perspective that the massive increase in information-processing demands imposed by bipedal locomotion conferred a selective advantage—accelerating the delegation of information processing from conscious, flexible control to more automated systems^[5]. Rigidity is not merely the cost of flexibility but its precondition.

ESTABLISHED Gilbert (1998) proposed from an evolutionary psychology perspective that cognitive biases (including rigid thinking) are natural consequences of employing fast, threat-sensitive defensive algorithms. Across various contexts, especially threat contexts, humans evolved for adaptive thinking rather than logical thinking^[6].

3.2 The Basal Ganglia: Physical Substrate of Habit Circuits

ESTABLISHED Yin and Knowlton (2006) established the modern framework for habit formation research in Nature Reviews Neuroscience: the neural correlates of habit formation manifest as a transfer of control from associative cortico-basal ganglia networks to sensorimotor cortico-basal ganglia networks. The dorsomedial striatum (DMS) mediates goal-directed behavior, while the dorsolateral striatum (DLS) mediates habitual behavior^[7].

ESTABLISHED Ashby et al. (2010) proposed that the development of automaticity is a gradual process in which control is transferred from subcortical procedural learning systems to purely cortico-cortical networks. The primary function of the basal ganglia may be to train the cortical representations that mediate automaticity^[8]. Once behavior is encoded in the DLS habit circuit, its execution becomes independent of outcome representations and therefore resistant to subsequent changes in outcome value.

ESTABLISHED Graybiel and Grafton (2015) further confirmed that habits, obsessive-compulsive disorder, and addiction share overlapping dorsal striatal circuits^[9], demonstrating that the “over-rigidification” of thought and behavior is not merely an everyday phenomenon but also a core mechanism of multiple psychiatric disorders.

3.3 The Life-Cycle Trajectory of Cognitive Flexibility

ESTABLISHED The emergence of cognitive flexibility is associated with the maturation of the prefrontal cortex (PFC) and the inferior parietal cortex. The frontal cortex matures slowly, continuing through adolescence into early adulthood^[10]. Different subregions within the PFC exhibit different maturational timelines—the orbitofrontal cortex matures earlier, while the dorsolateral PFC follows a more protracted developmental course^[11].

ESTABLISHED Cognitive aging is accompanied by structural changes including synaptic pruning, myelin degradation, and reduced neurogenesis, but adaptive compensatory mechanisms that promote continued performance also exist^[12].

IV. Reward-Weighted Memory and the Formation Mechanism of Experiential Lock-in

4.1 Dopamine Reward Prediction Error

ESTABLISHED Dopaminergic neurons encode the discrepancy between reward and expectation (reward prediction error, RPE) in the goal-directed system, and the discrepancy between the selected action and the habitual action in the habit system. These prediction errors respectively trigger learning about rewards and the formation of habits^[13]. Dopamine signaling in the DLS is critical for habit formation, while dopamine signaling in the DMS is essential for action-outcome learning^[14].

4.2 Synaptic Tagging and Capture

ESTABLISHED Hippocampal representations of high-reward contexts are preferentially reactivated during post-learning rest periods, and the number of reactivations predicts the preferential retention of high-reward content^[15]. Dopamine from the ventral tegmental area (VTA) modulates long-term potentiation (LTP) in the hippocampus, further strengthening memories that are temporally proximate to reward events^[16]. Rewarded information is preferentially replayed during sleep^[17].

4.3 Whole-Body Reinforcement via the Gut-Brain Axis

ESTABLISHED The brain-gut axis plays a modulatory role in reward processing. The gut harbors over 90% of the body’s serotonin and approximately 50% of its dopamine^[18]. Gut-derived serotonin influences central nervous system mood regulation and behavioral motivation through vagal afferent fibers^[19].

V. Evolutionary Mismatch in the AI Era: The Payoff Function Inversion of Experiential Lock-in

5.1 The Environment-Dependence of the Exploration-Exploitation Trade-off

ESTABLISHED In stable environments, the exploration-exploitation dilemma can be effectively resolved by first exploring and then continuously exploiting the optimal action; in volatile environments, exploration and exploitation must be continuously balanced. High levels of exploitation render agents unable to adapt to environmental changes, while excessive exploration reduces reward accumulation^[20].

ESTABLISHED The anterior cingulate cortex and the striatum causally regulate the level of exploration-exploitation, directing exploration toward reward-relevant targets during periods of high uncertainty, but operating in opposing directions^[21]—the brain system responsible for “whether to switch strategies” is itself engaged in a tug-of-war.

5.2 Evolutionary Mismatch Theory

ESTABLISHED Evolutionary mismatch refers to biological traits that were adaptive in ancestral environments but become maladaptive in modern environments because rapid change outpaces the speed of evolutionary adaptation^[22]. Boyd and Richerson’s (1985) gene-culture coevolution model demonstrated that the rate of cultural transmission can far exceed the rate of genetic adaptation^[23].

5.3 The Rate of Change in the AI Era: Empirical Anchoring

ESTABLISHED A 2026 systematic review (94 studies included in qualitative synthesis, 42 in quantitative extraction) found that the labor force exposure characteristics specific to this wave of LLMs are concentrated in non-routine cognitive work—qualitatively different from the routine manual and clerical task exposures documented in prior automation literature. Approximately 80% of the U.S. labor force has at least 10% of its tasks exposed to LLMs^[24].

VI. The Cognitive Polarization Hypothesis

6.1 The Self-Acceleration Mechanism

HYPOTHESIS What distinguishes cognitive polarization from wealth polarization is its positive-feedback structure. A 2026 paper described this mechanism: cognitive dependence diminishes individuals’ capacity to navigate complex information and weakens their resilience to labor market transitions requiring abstract thinking. Meanwhile, those who reflectively integrate AI gain advantages in education, employment, and political influence. Over successive technology adoption cycles, differences accumulate into a structural stratification of cognitive agency^[25].

6.2 The Cognitive Output Multiplier Model

HYPOTHESIS This paper proposes a simplified mathematical expression for AI as “multiplier rather than addend”:

The significance of introducing the exogenous variable E is to break the model’s circularity: B determines α, α determines C, C feeds back to reinforce B—if F were determined by B alone, the model would degenerate into the tautology “those with higher ability benefit more.” The introduction of E means that institutional design, educational policy, and AI interface quality can influence cognitive output trajectories independently of individual baseline ability. This model is positioned as conceptual formalization—it expresses the core mechanistic intuition of “multiplier rather than addend,” not a rigorous mathematical model amenable to direct numerical simulation. The latter would require empirical calibration of the specific functional forms of α, F, and E.

6.3 Indirect Evidence Chain for Brain Structural Divergence

ESTABLISHED Experience-dependent structural brain plasticity is well documented: experts exhibit significantly larger gray matter volume in relevant brain regions compared to novices^[26]; experience-induced changes can occur rapidly after short-term learning and may also dissipate when practice ceases^[27]; socioeconomic inequality has already been associated with measurable brain structural differences^[28].

VII. The Cognitive Attenuation Risk of Passive AI Use and the Tripartite Stratification Model

7.1 Early Evidence for Cognitive Degradation from Passive AI Use

HYPOTHESIS Multiple early studies suggest that passive AI use may impair cognitive function—a phenomenon referred to in media contexts as “AI brain rot”—though evidential limitations must be noted:

A longitudinal study from the MIT Media Lab (54 students, 4 months) found that students who relied entirely on AI for writing exhibited weaker brain connectivity and lower memory retention rates, with declines persisting after cessation of use^[29]. Limitation: Small sample size; full peer review not yet complete.

A joint Carnegie Mellon University study (319 professionals) found that 62% of participants reported reduced critical thinking effort when using AI^[30]. Lee et al. (2025) from Microsoft Research further confirmed that generative AI reduced participants’ perceived critical thinking effort^[31]. Limitation: Self-report data; the relationship between subjective perception and actual cognitive ability change requires further verification.

Eliav (2026) proposed the “delegation feedback loop” hypothesis: AI capability growth → lowered human delegation threshold for cognitive tasks → reduced cognitive practice → further capability weakening → further lowering of delegation threshold. Neither trend reverses spontaneously^[32].

7.2 The Paradox Structure and the Tripartite Stratification Model

HYPOTHESIS Synthesizing the above evidence, this paper proposes a tripartite stratification model:

The critical distinction of this model is: using AI per se does not determine cognitive trajectory; the usage mode does. It must be emphasized that A/B/C are within-task-domain usage mode classifications, not fixed population categories—the same individual may operate in Mode B (passive use) when writing, Mode C (adversarial use) when making investment decisions, and Mode A (non-use) when learning new technologies. The unit of stratification is “person × task domain × time period,” not the person as such. The reward-to-cost ratio of passive use is extremely high (minimal cognitive investment yields immediate results), making it more readily encoded by the basal ganglia as a strong habit circuit than non-use of AI—meaning that Mode B’s new rigidity may be harder to reverse than Mode A’s legacy rigidity.

VIII. The AI Multiplier Effect and Cognitive Stratification Amplification

8.1 The Innate Stratification of Human Cognitive Ability

ESTABLISHED Standardized IQ tests are designed with a mean of 100 and a standard deviation of 15. Approximately 68% of the population scores between 85 and 115, 95% between 70 and 130, and only 2.5% above 130^[33]. Over the past 30 years, skills in deductive reasoning, inference formation, and argumentation evaluation have declined by an average of 10–15% across the general population^[34]. It must be noted that standardized intelligence tests are a rough proxy for cognitive ability differences but cannot exhaustively capture effective AI-use ability. The latter is also influenced by openness to experience (the Big Five personality dimension), metacognitive ability, growth mindset (Dweck, 2006), educational background, cross-cultural experience, and domain expertise. In this paper’s multiplier model, these factors are compressed into the baseline ability variable B, but B ≠ IQ.

8.2 Empirical Support for the Multiplier Effect

ESTABLISHED A 2025 paper directly modeled this mechanism: information providers customize content complexity to maximize profits, delivering less cognitively enhancing content to users with lower cognitive ability. Individuals with higher cognitive ability benefit, while those with lower cognitive ability suffer adverse effects^[35].

HYPOTHESIS Combining this effect with the multiplier model proposed in this paper (Equation 1): when AI capability A(t) continues to grow exponentially, the distribution of baseline ability B (originally normal) is stretched through the positive feedback of α and F. The top cohort’s C(t) grows exponentially, while the bottom cohort’s C(t) stagnates or declines. The mathematical result is a distribution evolving from bell-shaped toward power-law, with the middle layer being compressed.

8.3 The Information Barrier Hypothesis

HYPOTHESIS This paper proposes that the information asymmetry created by AI differs from all previously technology-induced asymmetries: past gaps were “quantitative differences”—different amounts of information access but the same class of cognitive tools; the AI-era gap tends toward “qualitative difference”—the cognitive systems on either side of the barrier are no longer of the same kind (human brain + AI hybrid system vs. bare brain). For individuals lacking abstract reasoning ability, confronting AI output permits only “acceptance” (cognitive dependence) or “rejection” (reversion to legacy mode), but not “evaluation and improvement”—the latter requires metacognitive ability that is itself constrained by the cognitive normal distribution.

IX. Abductive Reasoning: Humanity’s Asymmetric Advantage

9.1 Three Forms of Reasoning and AI’s Capability Distribution

ESTABLISHED Peirce distinguished three forms of reasoning: deduction, induction, and abduction^[36]. In many formalized deductive tasks and large-scale pattern induction tasks, frontier AI systems have reached or exceeded the level of most non-expert humans, performing prominently in some specialized tasks. Induction—extracting patterns from large datasets—is the native strength of deep learning architectures.

HYPOTHESIS Abductive capability is constrained by current LLM architectures—a conditional probability distribution sampler can only output patterns within the training distribution and their interpolations; it cannot autonomously create cross-domain connections absent from the training data. AI can efficiently execute search and recombination once a direction has been specified, but the operation of “selecting which direction is worth probing”—especially locating cognitive gaps from real physical-world experience—currently remains a human capacity. The temporal window of this advantage depends on the rate of evolution of AI architectures.

9.2 “Matrix Cognizer”: Definition and Behavioral Indicator Framework

HYPOTHESIS This paper proposes the concept of “Matrix Cognizer,” describing the emergent cognitive paradigm that arises when abductive reasoning ability encounters AI tool amplification. Its triadic operational structure is:

Problem-selection intuition (locating cognitive gaps from physical-world experience) + Abductive reasoning (retroductively inferring unified explanatory frameworks from cross-domain fragments) + AI as inductive/deductive accelerator (rapid cross-domain information retrieval and verification)

To render this concept operationalizable, a preliminary behavioral indicator framework is proposed:

A necessary limitation must be stated: the Matrix Cognizer is not a capacity universally attainable through short-term training. The prerequisite abilities for abductive reasoning—abstract pattern recognition, cross-domain analogy, hypothesis generation and self-testing—are constrained by the normal distribution of cognitive abilities. This concept describes the apex manifestation of cognitive differentiation in the AI era.

X. Countervailing Variables and Framework Boundaries

The stratification framework constructed in preceding chapters derives from currently observable trends. But an intellectually honest theoretical framework must simultaneously identify countervailing variables that could decelerate, alter, or even reverse the predicted trajectory. This chapter discusses three such variables and assesses their impact on the core argument.

10.1 The Democratization of AI Interaction Interfaces

HYPOTHESIS Chapter VIII of this paper argued that the metacognitive abilities required for effective AI interaction are constrained by the cognitive normal distribution. However, this argument implicitly assumes that AI’s interaction paradigm remains in its current form. Historically, the democratization of technology has typically transformed operations mastered only by a select few into infrastructure accessible to the masses—from command line to graphical user interface, from programming to spreadsheets, every interface revolution has dramatically lowered the cognitive threshold for effective use.

Current research directions—which may be termed “Proactive Socratic Scaffolding”—aim to have AI systems actively elicit reflective engagement from users rather than passively awaiting high-quality input. If future AI systems could detect when users are in “passive reception mode” and proactively switch to a questioning, challenging, and guiding mode, this would lower the cognitive threshold for Mode C from the tool side—some users might enter adversarial usage mode under AI guidance, even when their spontaneous motivation is insufficient.

Impact assessment on core argument: This variable does not negate the existence of the multiplier effect but may narrow its amplification magnitude. It would shift some Mode B users upward to a transitional zone between Mode B and Mode C. However, the actual efficacy of this variable depends on two unknowns: (a) whether Socratic scaffolding can sustain its effectiveness in the absence of users’ spontaneous motivation—the behavioral decay evidence cited earlier^[29] suggests this is not easy; (b) whether commercial incentives support such designs—the competitive structure of current digital markets systematically selects for maximizing user engagement (i.e., catering to passive-use dopamine circuits) rather than maximizing user cognitive development^[39]. Making users feel “good” and “frictionless” always scores better on retention metrics than making them “think painfully”—scaling adversarial interfaces runs counter to the current commercial dynamics.

10.2 The Compensatory Effect of Cognitive Outsourcing

HYPOTHESIS Chapter VII of this paper unidirectionally defined passive AI use as a path to cognitive degradation. But a significant reverse mechanism warrants consideration: outsourcing low-level cognitive tasks to AI may free prefrontal cognitive load, allowing it to be reallocated to higher-dimensional tasks.

Historical analogies provide partial support: the proliferation of calculators weakened human mental arithmetic ability but did not lead to the wholesale destruction of mathematical capacity—on the contrary, it facilitated higher-dimensional mathematical and engineering development. Word processing software reduced people’s handwriting ability but freed more cognitive resources for content ideation. Every instance of cognitive outsourcing is accompanied by the degradation of specific sub-abilities and the potential enhancement of others.

Impact assessment on core argument: The critical question regarding the compensatory effect is not “is it theoretically possible” but “does it commonly occur in practice.” Are the freed cognitive resources actually reallocated to higher-dimensional tasks, or are they consumed by more passive content consumption? Current empirical data lean toward the latter—most adults interact with technology through a “substitution” rather than “amplification” lens^[34], with freed cognitive resources absorbed by short-form video, social media, and low-intensity information consumption rather than channeled toward deep thinking. The limitation of the calculator analogy is this: the cognitive resources freed by calculators had a natural destination (more complex mathematical problems were sitting there waiting), whereas the cognitive resources freed by AI face a meticulously engineered attention economy whose commercial logic is precisely to intercept these freed resources^[39]. Therefore, the compensatory effect is theoretically valid, but its realization requires deliberate institutional design to counteract the default gravity of the attention economy.

10.3 The Redirection Potential of Educational Intervention

HYPOTHESIS This paper’s stratification framework might be read as cognitive fatalism—implying that most people are “destined” to be unable to cross the information barrier. But educational intervention research suggests this reading is overly deterministic. The development of critical thinking is jointly influenced by physiological, psychological, sociocultural, technological, and educational factors^[41]. Small-scale programs—emphasizing learning through multidimensional, explanatory projects rather than relying on passive consumption of digital content—have shown early signs of improvement in 6th–12th graders’ cognitive fluency and individual critical analysis metrics^[34].

Impact assessment on core argument: Educational interventions can change an individual’s position within the tripartite stratification model but are constrained by two factors: (a) scale constraint—existing effective interventions are mostly small-scale programs, and large-scale deployment faces resource and institutional barriers; (b) temporal constraint—the data cited in this paper indicate that the AI impact window occurs at the 2–3 year grade level, while educational system reform cycles typically operate on decadal timescales. Interventions may be unable to match the speed of the impact.

10.4 Comprehensive Assessment: Countervailing Variables Alter the Slope, Not the Direction

The three countervailing variables discussed above—interface democratization, compensatory effects, and educational intervention—all have the potential to slow the stratification process predicted in this paper. Given that the current structure of commercial incentives, educational response speed, and default user behavior patterns do not undergo fundamental change, what they more likely alter is the speed of stratification (the slope), not its direction (the sign). As long as the following two foundational conditions remain unchanged, the direction of stratification will not reverse: (a) the brain’s energy-minimization default setting makes rigidity a baseline state that requires no external force to maintain; (b) the exponential growth of AI capabilities continues to widen the output gap between effective and ineffective users. In the best-case scenario, countervailing variables can expand Mode C’s boundaries, slow the rate of middle-layer collapse, and provide ladders for more people to cross the information barrier—but under the stated preconditions, they cannot eliminate the barrier itself, because the barrier’s physical foundation lies in the differential distribution of neuroplasticity and the biological properties of the basal ganglia. If either precondition undergoes fundamental change (e.g., AI interfaces evolve at scale into Socratic scaffolding mode, or educational systems successfully incorporate adversarial AI-use training into core curricula), partial reversals of the stratification direction may occur—but this itself would require overcoming the countervailing incentives of the current commercial logic.

10.5 Dimensions Not Covered by This Framework

This paper discusses “human cognition” as a relatively universal phenomenon, but at least two dimensions may produce structural corrections to the framework’s predictions and should be incorporated in future research:

Cultural differences. Cognitive flexibility, attitudes toward change, and the strength of conformity effects vary significantly across cultures. Conformity pressures in collectivist cultures (such as East Asia) may accelerate the social-level amplification of rigidity; exploratory incentives in individualist cultures (such as North America) may provide more opportunities to break rigidity. Different countries’ AI adoption patterns, educational system designs, and social incentive structures may all lead to regional variation in stratification trajectories. Under what cultural conditions this framework’s predictions are most applicable, and under what conditions they require modification, is a question that remains unanswered.

AI-native generations. The analysis in this paper primarily targets adult populations who completed cognitive development in non-AI environments before confronting the AI impact. However, the generation born after 2020, who coexist with AI systems from birth, will have basal ganglia trained in an AI environment from the outset. Their cognitive architecture may differ structurally from pre-AI generations—their “default settings” may already include basic circuits for AI collaboration. This variable may fundamentally alter the long-term (20+ year) predictions of this paper but does not affect the analytical validity for the current transition period (5–10 years).

XI. Conclusion: Testable Hypotheses and Future Research Directions

The six-layer causal chain constructed in this paper—from brain energy economics through reward-weighted memory, exploration-exploitation inversion, cognitive polarization, to brain structural divergence—provides a unified explanatory framework spanning from micro to macro. Its core theoretical thread can be compressed as follows:

Human experiential lock-in is essentially a behavioral model produced by the neural system’s reward-weighted rigidification of past returns. In low-change environments, it is a high-efficiency strategy; in the high-change environment created by AI, it becomes cognitive lock-in. Therefore, the true stratification of the AI era is not about whether one uses AI, but whether one can use AI to continuously break one’s own experiential lock-in.

To advance this framework from “theoretical model” to “verifiable scientific program,” this paper proposes three testable hypotheses:

Principal limitations of this framework: (a) H3 has not yet been directly tested by any longitudinal neuroimaging study; (b) the specific functional forms of parameters α, F, and E in the multiplier model (Equation 1) require empirical calibration; (c) the Matrix Cognizer behavioral indicator framework requires reliability and validity testing; (d) the boundary between Mode B and Mode C within the tripartite stratification is in practice a continuum rather than a discrete classification; (e) the framework does not adequately cover the potential structural effects of cultural differences and AI-native generations (see Section 10.5).

These limitations simultaneously constitute direct starting points for future research. What this paper can do is propose this framework through abductive reasoning and provide direction for its subsequent inductive verification and deductive derivation.