Social Sciences

Abstract: The “creative” outputs of artificial intelligence systems—from GPT composing poetry to AlphaFold predicting protein structures—raise a philosophical question: where does creativity come from? Is it simple reproduction of training data, or a chance product of stochastic algorithms? This paper proposes “The Creativity Conjecture”: the creative output of an AI system is neither pure reproduction of training data nor pure random generation, but an emergent phenomenon arising from “experience training” and “random perturbation” under a “multi-level data fusion” mechanism. The conjecture comprises three conditions: (1) empirical density—the system has internalized sufficient pattern structure from training data; (2) random perturbation—the system introduces randomness of appropriate intensity during generation; (3) fusion depth—the system possesses the capacity for cross-level, cross-modal data fusion. All three are necessary: randomness without experience is noise; experience without randomness is copying; experience plus randomness without fusion is collage. Only when all three fuse does “structured surprise”—i.e., creativity—emerge. This paper argues for the conjecture on three levels: structural analysis (the mechanism of each element), fusion mechanism (how the three couple to produce creative emergence), and experimental verifi- cation (the effects of randomness intensity and fusion depth on creative output). We further propose the concept of “weak creativity” to delineate the epistemological status of machine creativity—it transcends random generation and data reproduction, yet does not reach the level of human phenomenal creativity. This framework provides a structural foundation for understanding AI creativity and offers a new path for the philosophy of creativity, moving from “genius theory” to “structural theory.”

Abstract: This integrative literature review examines how generative artificial intelligence is shaping human thinking, cognitive processes and writing practices, with a sustained focus on education across schools, universities, adult learning and professional settings. It considers how AI-supported composing alters the ways writers plan, generate, organise, revise and evaluate text, and how these changes reshape cognition, agency, authorship and learning. Rather than treating generative AI simply as a tool for productivity, the review approaches it as a relational and quasi-agentic technology that participates in meaning-making, decision-making and textual production. Research is synthesised across education, cognitive psychology, writing studies, digital literacy and human-computer interaction. Albert Bandura's social cognitive theory provides the analytical lens, particularly the concepts of triadic reciprocal causation, self-efficacy, observational learning, self-regulation and human agency. The review asks how generative AI mediates the relationship between person, behaviour and environment, and how it reshapes both confidence and dependence in writing. The discussion develops a critical account that recognises clear benefits while warning against cognitive offloading, uncritical trust, homogenised expression and weakened metacognitive control. The article concludes by proposing guidelines for educators across sectors to support critical AI awareness, reflective writing practice and agentic human-AI engagement.

Abstract: Skilled readers are thought to rely on bottom-up processing to activate word meanings in long-term memory from their form (i.e., spelling or sound) in a process described as ‘context-free’ lexical processing. However other research suggests skilled readers recognize words faster when they appear in predictable and/or plausible contexts, despite efficient bottom-up processing [1]. Research on individual differences in skilled comprehension suggests skilled readers only use context to support word recognition when bottom-up processing is slow [2]. Therefore, highly skilled readers may not rely on context, but less-skilled readers may continue to rely on context during word recognition due to poor bottom-up processing. The current study investigated whether individual differences in lexical quality would influence contextual processing during word recognition. Participants eye movements were tracked as they read sentences containing either a high frequency (e.g., baby) or low frequency (e.g., elbow) target word preceded by either a related prime 20 word (mother – baby) or an unrelated control word (teeth – elbow). Results show relatedness had a significant effect on readers across the range of lexical knowledge, however the specific effect depended on both vocabulary and target word frequency. The results suggest that context supports word recognition and comprehension when readers have adequate vocabulary.

Abstract: This literature review synthesizes the growing body of scholarship on cognitive hacking, a form of cyberattack that targets human perception and decision-making rather than technical infrastructure. Cognitive hacking sits squarely within Libicki’s framework of semantic attack, and countermeasures against such attacks are expected to constitute an important area of research in the science of intelligence and security informatics (Thompson, 2004). Drawing on foundational works originating from the Semantic Hacking Project, contemporary studies in behavioral cybersecurity, and recent empirical scholarship on artificial intelligence (AI)–driven disinformation, this review identifies four overarching themes: the definition and taxonomy of cognitive attacks; the psychological mechanisms underlying cognitive exploitation; the technological tools that amplify such attacks; and the countermeasures available to researchers and practitioners. Recent experimental evidence indicates that generative AI systems now produce disinformation that is more compelling than human-written content (Spitale et al., 2023) and can out-persuade human interlocutors when supplied with minimal personal information about their targets (Salvi et al., 2025). With generative AI enabling faster, cheaper, and more convincing tailored disinformation, scholars and policymakers are urgently seeking coordinated ways to regulate and mitigate the impact of deepfakes (Romanishyn et al., 2025). At the same time, meta-analytic and experimental work on psychological inoculation demonstrates that resilience to manipulation can be cultivated at scale (Roozenbeek et al., 2022; van der Linden, 2022). The review identifies significant research gaps, particularly at the intersection of generative AI and cognitive resilience, and concludes that a multidisciplinary approach integrating cognitive psychology, security informatics, and public policy is essential for developing robust defenses.

Abstract: Two people can view the same information but can take different actions. Most importantly, while thinking about moral, political issues, such as abortion, depending on the variation in the psychological systems, a perception of threat can arise. To exploit this aspect of the psychological systems, state and non-state actors make use of propaganda techniques to influence any target audience. Although the spread of disinformation is extensively studied in the information eco-system, an understanding of the effects of propaganda that result from interactions between the schemata of the individuals and schemata of the perceptual stimuli is not accounted for in systemic analysis of the disinformation threat vector. Moreover, while aiming to undermine the norms, values, and institutions in free societies, propaganda does not target the material objects or networks; it targets the psychological systems. Therefore, to improve the systemic understanding of disinformation propaganda, this research scrutinized the possibility of using the moral foundation theory (MFT) to include the psychological systems in a target audience analysis. The moral sentiments of an information operation were analyzed and the MFT dictionary was used to determine the moral scheme of the information streams. The words and hashtags were clustered by using the MFT dictionary; then, the moral cluster and word frequency analyses in the context of disinformation propaganda were conducted. The moral cluster frequency approach can augment the target audience analysis; it can ameliorate the counter information operation endeavors and threat vector analyses. To further study this, a quantum open system model of interaction between an information stream and the target audience cognitive system was built. The effects of non-selective measurement were studied via modeling and simulating a quantum open system based cognitive model.

Abstract: The article examines the transition from traditional to digital experiences by analyzing the relationships between attitudes toward digital and non-digital environments in four domains: navigation, work, public transportation, and shopping. In four studies, participants’ attitudes were examined toward a parallel digital experience (a navigation app, remote work, the use of digital means in public transportation, and online shopping) compared to the traditional experience. In three contexts (navigation, remote work, and online shopping), a negative correlation between attitudes was found, indicating that the technology is perceived as a substitute for the traditional environment. In contrast, in the context of public transportation, a positive relationship was found, as digital means are perceived as a complementary component to the physical experience. The findings are interpreted using the exemplar approach to memory, according to which digital and traditional experiences are stored in separate or overlapping sets of exemplars. The article proposes an extension to classical technology acceptance models and highlights practical implications for designing implementation processes that emphasize experiential continuity between the traditional and digital environments.

Abstract: Ultrasound mid-air haptics (UMAH) can create touch sensations without physical contact, but it remains unclear whether they evoke steady-state somatosensory evoked potentials (SSSEPs) that could serve as objective markers of user experience. This study tested whether 20 Hz UMAH delivered with a commercial device at maximum available intensity elicited SSSEPs comparable to those produced by vibrotactile stimulation (VTS). Electroencephalography was recorded from 26 participants during three stimulation conditions: full-intensity VTS, subjectively matched-intensity VTS, and full-intensity UMAH. Signal-to-noise ratio (SNR) and power spectral density (PSD) at 20 Hz were analyzed over contralateral and ipsilateral somatosensory regions using linear mixed-effects models, with baseline estimates derived from no-stimulation intervals. Full-intensity VTS produced clear contralateral SSSEPs, and subjectively matched-intensity VTS yielded weaker but significant responses. In the present setup, UMAH did not yield a detectable 20 Hz SSSEP relative to baseline in either SNR or PSD. These findings support SSSEPs as sensitive markers of contact vibrotactile stimulation, but suggest that, with the present apparatus and analysis approach, they are not yet a robust objective measure for evaluating UMAH experience.

Abstract: Background: The Analytic Thinking Score (ATS) has been interpreted as a linguistic marker of organized thought. Written language typically shows higher ATS than spoken language. From the perspective of active inference under the free energy principle, we proposed a neurocomputational model of the mechanism underlying this writing-over-speaking advantage. We propose that ATS reflects precision-weighted inference over latent conceptual-organization (CO) states during language production. We hypothesize that written production supports higher ATS when it increases posterior confidence in high-CO states. Methods: University students described Thematic Apperception Test images in spoken and written modalities. ATS values were obtained from the resulting language samples. Participants were modeled as active-inference agents using a two-timestep Markov decision process (MDP) in which observed speaking and writing cues updated beliefs about latent CO states. Belief updating was formalized through variational message passing and interpreted in terms of prediction-error signaling and precision-weighted neuronal synaptic gain. An attention-related parameter (AP) controlled the precision and directionality of the mapping between latent CO states and observed production cues. Bayesian model selection was used to assess the model’s construct validity. Results: Written responses showed higher ATS than spoken responses. The group-level AP estimate indicated that writing cues supported posterior inference toward high-CO states stronger than speaking cues. Bayesian model selection favored the active-inference MDP over a Variational Laplace linear model. Conclusions: Writing may increase ATS by providing production cues that support precision-weighted inference toward high-CO states. ATS is therefore interpreted as a downstream linguistic trace of latent CO inferred during language production.

Abstract: Social anxiety (SA) negatively impacts myriad aspects of an individual’s life. Although research with adults and children highlights an important link between SA and social-cognitive abilities (e.g., reasoning about others’ thoughts and emotions), findings are mixed. We hypothesized that these mixed findings stem from the various combinations of social-cognitive components of SA under investigation and the different types of measures used. Understanding these relationships in middle to late childhood is especially important, given that it is a period of substantial social-cognitive development and a common onset age for SA. Seventy-eight children (Mage=8.15 years, SD=1.61) and their parents completed measures capturing different components of anxiety (i.e., social worry, fear of negative evaluation, and social avoidance) and social cognition (i.e. emotion recognition, mental state understanding, and social perspective taking). Contrary to our expectations, measures of social cognition were only weakly correlated. Consistent with our expectations, associations between social cognition and social anxiety were measure-dependent. Self-reported fear of negative evaluation emerged as a positive predictor of accuracy in a behavioral measure of mental state understanding but a negative predictor of parent-reported mental state understanding. In addition, social avoidance accounted for additional variance only when predicting lower self-reported perspective-taking. Together, our findings underscore the multifaceted nature of social cognition and SA and highlight the need for distinguishing these facets in future work.

Abstract: Despite the dramatic advances made in artificial intelligence (AI) and other fields of computer science towards implementing “intelligent” systems expert in specific tasks, the goal of devising algorithms and machines able to interact with human beings just as naturally as other humans do is still elusive. As this naturalness is arguably a consequence of the similarity of the underlying ‘hardware’ (the human brain), it is reasonable to claim that only artificial systems closely inspired by the actual functioning of the human brain and mind have the potential to render this possible. More specifically, the aim of this paper is to propose a new, biologically inspired computational model able to mimic, in a more accurate way than existing ones, the set of functionalities know as Theory of Mind. This is a set of mental processes that allow an individual to attribute mental states to others. In human social interactions this mechanism is crucial, as it allows one to explain the observed behaviour of others, to guess their intentions and to effectively predict their future conduct. This happens by modelling and selecting the most likely (unobservable) mental states of the considered person, which are the primary causes of everyone’s observed actions. The proposed model combines a number of concepts, including those of hierarchical structure, hypotheses pre-activation, and the notion of agent class or ‘stereotype’. It rests on one of the main psychological approaches to Theory of Mind, termed Simulation Theory (ST), and is supported by significant neuroscientific evidence. Crucially, unlike previous efforts in AI, the proposed model puts the learning element at the forefront, in the belief that simulations of other intelligent being’s reasoning processes need to be learned from experience. In this perspective, a possible implementation of the model in terms of deep, reconfigurable neural networks, trained in a reinforcement learning setting, is outlined.

Abstract: The increasing adoption of generative artificial intelligence (AI) in higher education has created new opportunities to enhance Learning Management Systems (LMS) with personalized feedback, adaptive assessment, and learning analytics. Despite these advances, many LMS platforms remain primarily focused on content delivery and grade management, with limited support for metacognitive assessment and intelligent feedback. This study presents CONF.i, a confidence-informed assessment and AI feedback framework integrated with Canvas LMS using Google Apps Script and Google Gemini AI. Developed through a design-based research approach, the framework combines traditional assessment scores with student self-reported confidence levels to support personalized formative feedback and diagnostic learning insights. The proposed system integrates Canvas LTI standards, a Google Apps Script backend, and Gemini AI services to automate scoring, confidence tracking, and AI-generated educational feedback within existing institutional infrastructure. A prototype implementation was evaluated using simulated learner profiles representing different combinations of performance and confidence patterns. The framework identified four illustrative assessment profiles: aligned mastery, underconfident competence, overconfident struggle, and aligned struggle. These patterns demonstrate how confidence-informed assessment can reveal metacognitive dimensions of learning that are not visible through conventional grading alone. Preliminary usability observations indicated positive perceptions regarding the integration within the familiar Canvas environment and the relevance of AI-generated feedback, while also identifying limitations related to response latency and feedback specificity. The findings suggest that integrating confidence-informed assessment with generative AI may support more personalized and reflective learning experiences without requiring major institutional infrastructure changes or commercial licensing costs. This study contributes an exploratory prototype framework for AI-enhanced formative assessment in higher education and provides a practical model for institutions seeking to extend existing LMS platforms with confidence-aware analytics and personalized feedback capabilities.

Abstract: Digital Transformation (DT) increasingly relies on project-based organizing to develop and deploy new capabilities, yet corporate innovation projects frequently stall not for lack of ideas but because of recurring governance and resource-commitment bottlenecks. This study presents a micro-longitudinal, AI-enabled, and human-reviewed analysis of 711 episodes drawn from 28 weekly project governance meetings across two corporate startup initiatives participating in the same internal incubation program, conducted between November 2024 and April 2025. Employing a six-stage analytical pipeline that combines episode-level segmentation, linguistic tension markers, and a large language model (LLM) classifier, we identify 28 decision-relevant governance tensions, which are then abductively grouped into 13 project governance dilemmas and mapped onto Teece's dynamic capabilities framework (sensing, seizing, reconfiguring). The key finding is that 62% of dilemmas are structural in nature—reflecting persistent governance design tensions between autonomy and control, compliance and agility, and centralization and decentralization—and that 69% concentrate at the seizing stage, corresponding to resource-commitment and execution decisions. This pattern indicates a governance choke point in corporate DT projects that is structural and decisional rather than ideational. By shifting attention from lagging indicators (overruns) to governance-tension leading indicators, the approach supports earlier interventions to reduce decision latency and protect project delivery performance. We further synthesize two incubation-specific meso-level governance dilemmas—stakeholder engagement and compliance vs. agility—that serve as transmission mechanisms between macro structural constraints and micro-level decision bottlenecks. The AI-enabled pipeline is proposed as a replicable early-warning system for project governance tensions in organizations pursuing digital transformation.

Abstract: Artificial intelligence is increasingly embedded in decision-making across organizational and societal contexts, yet it remains unclear whether individuals remain cognitively aligned with decisions generated under algorithmic conditions. Existing research has emphasized trust, fairness, and transparency, but provides limited insight into the cognitive mechanisms that sustain coherent human judgment during system-mediated decision processes.Here we introduce perceptual integrity as a measurable construct capturing the extent to which individuals maintain interpretive coherence and decision authorship in human–AI interaction. We test this framework in a controlled experiment (N = 602) comparing algorithmic imposition with interpretive autonomy. Algorithmic imposition significantly reduced perceptual integrity relative to interpretive autonomy (t(600) = 4.21, p < 0.001, Cohen’s d = 0.38). Perceptual integrity was a significant predictor of trust in AI-assisted decisions (β = 0.36, p < 0.001) and partially mediated the relationship between decision condition and trust (indirect effect = 0.17, 95% CI [0.09, 0.27]).These findings identify perceptual integrity as a cognitive mechanism linking decision structure to trust under system-mediated conditions. More broadly, they suggest that effective integration of algorithmic systems depends not only on performance accuracy but on preserving cognitive alignment during decision formation. This work provides a generalizable framework for understanding how humans remain engaged with decisions in increasingly automated environments.

Abstract: Biological cognition depends on learning structured representations in ambiguous environments. Computational models of structure learning typically frame this as an inference problem, but often overlook the temporally extended dynamics that shape learning trajectories under ambiguity. In this paper, we reframe structure learning as an emergent consequence of constraint-based dynamics. Informed by a literature on the role of constraints in complex biological systems, we develop a constraint-based approach to computational cognitive modelling and provide a proof-of-concept model. The model consists of an ensemble of components, each comprising an individual learning process, whose internal updates are locally constrained by both external observations and system-level relational constraints. This is formalised using Bayesian probability as a description of constraint satisfaction rather than epistemic inference. Representational structure is not encoded directly in the model equations but emerges over time through the interaction, stabilisation, and elimination of components under these constraints. Through a series of simulations in environments with varying degrees of ambiguity, we demonstrate that the model reliably differentiates the observation space into stable representational categories. We further analyse how global parameters controlling internal constraint and initial component precision shape learning trajectories and long-term behavioural alignment with the environment. We discuss the formal relationship between the present approach and Bayesian inference accounts, and argue that a constraint-based approach offers a conceptually distinct foundation for relating computational models to biological systems.

Abstract: Large language models (LLMs) have made visible a long‑standing philosophical tension: sophisticated symbolic cognition can arise from large‑scale pattern extraction even in the absence of consciousness. This observation motivates a minimalist conceptual framework grounded in an ontological distinction between conscious regulation and symbolic structures. Language is treated as a crystallized form of human cognition—an externalized, culturally accumulated substrate created by conscious agents over millennia—while the human brain is understood as a biological system that evolved to operate over this symbolic layer. Within this view, consciousness and symbolic cognition are not different degrees of the same process but distinct kinds of cognitive organization: consciousness generates, grounds, and regulates symbols, whereas symbolic cognition manipulates them.LLMs illuminate this asymmetry by reproducing symbolic reasoning without conscious access, motivation, or subjective experience. Their performance therefore raises epistemological questions about the nature of meaning, grounding, and cognitive stability. The proposed framework situates these questions within a broader account of human cognitive evolution shaped by gene–culture coevolution and the emergence of culturally scaffolded symbolic systems. Finally, the article introduces an information‑theoretic constraint (the AI Theorem) suggesting that purely computational systems inevitably accumulate drift in the absence of a regulatory layer, offering a philosophical explanation for why artificial cognition may remain structurally distinct from biological minds.

Abstract: The digital age has fundamentally dissociated the creation of fundamental intellectual frameworks, such as novel theories, methodologies, and paradigms, from their widespread application and economic value realization. The fundamental reason why the creators of such meta-intellectual labor often receive disproportionate returns to the enormous long-term social and commercial value created by their work is that we cannot accurately measure, attribute, and automatically trade the value contained therein. In this paper, we propose a new integrated framework for automated valuation and liquidation of knowledge contribution based on the principle of fusion intelligence. This problem is formalized as a Knowledge Contribution Valuation and Liquidation (KCVS) system, with the dual formalization mechanism as its operational core, and the nine steps of intellectual integration as the maturity model of value creation. It shows how AI systems themselves, especially large language models, can be repositioned as impartial measuring instruments, automated traders, and transparent governance within this framework. Through the analysis of real cases of DeepSeek and Qianwen in scientific research and commercial applications, it is clarified that their underlying architectures have instantiated dual formal mechanisms, thus providing empirical support for the theoretical basis of the system proposed in this paper. This is followed by a blueprint consisting of three pillars: (1) an AI-driven knowledge contribution index for dynamic, multi-dimensional impact measurement; (2) a decentralized micropayment and clearing layer based on smart contracts; and (3) a transparent governance protocol for auditability using distributed ledgers. A simulated economic model is used to assess the feasibility of the framework and demonstrate its potential in building a sustainable, equitable, and self-optimizing ecosystem for foundational intellectual labor. This paper provides a theoretical and practical roadmap for aligning the incentives of knowledge creators with the structure of AI-driven economies, ensuring that future innovation is both dynamic and fair.

Abstract: Radial Analysis (RA) is a methodological framework that transforms radial category theory from static structural mapping into dynamic trajectory modeling. Building on the Trace & Trajectory Framework's (TTF) non-representationalist architecture, RA provides researchers with practical tools for analyzing indexicality, identity navigation, and meaning dynamics in discourse. This paper presents RA as an applied methodology rather than a foundational theory. The framework employs hexagonal geometry (the SpiderWeb architecture—a board game model based on hexagonal tessellation) to formalize navigational patterns: how speakers move through identity space, what these movements cost informationally, and how trajectorial patterns reveal underlying dynamics invisible to categorical approaches. Core innovations include: (1) the three-level terminology (Hexid/Hex/Hxp) for precise analytical description; (2) formally grounded metrics (hexagonal distance, trajectory cost, Temporal Dissipation Rate) enabling principled relational comparison; (3) the λ/ς/σ parameter system distinguishing structural granularity, semiotic depth, and epistemic access; (4) the depth parameter (ς) governing semiotic visibility through shading mechanics; (5) semiotic coherence (SC) as the constitutive principle underlying positional significance; (6) stratified epistemic barriers (Hxₙ) and hex bands (Hx⁽ⁿ⁾) structuring radial distance into qualitatively distinct reference domains with characteristic cost profiles; and (7) direct application to epistemic appropriation dynamics including flattening, internalization, and trajectorial refraction. RA addresses phenomena that categorical frameworks handle only through ad hoc mechanisms: simultaneous multi-level positioning, asymmetric intersubjective dynamics, and the geometric constraints that institutional power imposes on identity navigation. Applications span personal deixis, temporal reference, identity navigation dynamics, and—through integration with recent work on epistemic appropriation—the formal analysis of internalized oppression in clinical and educational contexts.

Abstract: Consciousness remains one of the most persistent problems in philosophy and cognitive science. Despite substantial advances in neuroscience, no consensus exists regarding how physical processes give rise to subjective experience. This paper proposes an evolutionary hypothesis according to which consciousness emerges from the interaction of three fundamental dimensions: neural integration of sensory information, the continuous influence of internal bodily states, and the capacity to simulate and anticipate possible scenarios. Within this framework, conscious experience is interpreted not as a mere byproduct of neural processing but as the phenomenological manifestation of a biological system capable of integrating multiple streams of information to construct a model of the organism situated in its environment. The paper develops the thesis that human consciousness can be understood as the result of an evolutionary simulation system that integrates external perception and internal bodily states within a phenomenal field structured around a bodily located self. Once established, this system may have exceeded its original adaptive functions, supporting complex forms of self-reflection, symbolic language, and cumulative culture. The proposal is also intended as an analysis of how contemporary cognitive science constructs explanatory models of consciousness, connecting neural mechanisms, embodied processes, and evolutionary function.

Abstract: Existing adult-report survey measures provide crucial information about children’s executive function (EF) development across contexts, but lack cultural relevance and ecological validity. To address these limitations, we introduce the Executive Function From Observation and Reflection Tool (EFFORT), a publicly available, open-source item bank designed for cross-cultural adaptation that includes 32 parallel items for caregivers and teachers across six EF domains: sustained attention, response inhibition, interference suppression, working memory, cognitive flexibility, and planning/organization. EFFORT additionally includes 10 assessor-report items intended for use following a structured, standardized assessment session. This study presents the first multinational validation of the tool across seven countries (Argentina, Australia, Bangladesh, Haiti, South Africa, Sri Lanka, United States) leveraging caregiver, teacher, and assessor observations of 1,738 children (aged 3–11 years). Findings revealed acceptable fit for a six-factor structure for caregiver and teacher reports that were not empirically distinct, but yielded highly reliable composites. We further validated a 12-item short form for caregiver and teacher that demonstrated strong unidimensionality, gender invariance, and age-related increases. We demonstrated significant convergence of a short-form caregiver and teachers composite with the assessor-reported measures, as well as convergence of all three adult reports with direct assessments of children’s EF skills. This new tool holds promise to advance the science of how children develop and apply EFs to accomplish everyday goals across different cultural settings and in understudied populations.

Abstract: The proliferation of complex conceptual systems developed in interaction with artificial intelligence agents poses an epistemological problem not anticipated by classical theories of falsification: in such systems, the external validation agent is simultaneously a structural generator of narrative coherence, inducing a functional collapse between the roles of creation and assessment. This collapse is not reducible to Popperian immunization or to the adjustment of auxiliary hypotheses in the Lakatosian sense, since it does not arise from deliberate defensive strategies but from an architectural asymmetry between the way such systems produce coherence and the way their human creators interpret it. This paper proposes the concept of epistemic delusion to designate the methodological state in which the operational conditions of falsification disappear as the cumulative effect of conceptual drift mechanisms, and argues that in AI-mediated self-referential systems this process exhibits a specific vector — systemic narrative induction — not yet systematized in the literature. The paper examines the mechanisms of conceptual drift, the modes of epistemic closure, and a set of methodological safeguards whose normative foundation is derived from the distinction between internally generated coherence and empirically independent corroboration.