Social Sciences

Abstract: Artificial intelligence is approaching a problem that cannot be reduced to the accumulation, retrieval, or representation of knowledge: heterogeneous bodies of knowledge must be assembled so that their local semantics can differ while their deeper structural relations remain operationally compatible. This paper proposes Artificial Science as a framework for this problem. Its central premise is that Integration Scienceidentifies portable mathematical structures across scientific domains, whereas Artificial Science converts such structural portability into executable mechanisms of knowledge transformation, composition, control, and stabilization. We develop the first unified model of this program throughgauge-symmetry knowledge assembly and formulate it as a theory of Big Knowledge Dynamics. The framework begins with a global-to-local geometric architecture in which knowledge contexts form a base space, structural symmetries act on internal representations, and connections, covariant derivatives, curvature, and holonomy govern transport across changing local frames. Within this architecture, four mechanism-specific sectors form a staged knowledge dynamics. provides the minimal model of covariant single-charge knowledge transport; provides a non-Abelian model of internally mediated multi-component assembly; supplies a doublet control architecture for externality-driven state transformation; and spontaneous symmetry breaking together with Higgs-like stabilization provides a model of locally stable yet adaptive rational organization. These sectors are not asserted to be ontologically identical to physical gauge theories. Their role is structural: each supplies a mathematically constrained interface that can be mapped into domain-specific knowledge systems and re-grounded in their own semantics. We combine these mechanisms into a staged assembly operator, A_BKD = S_stab ∘ X_SU(2) ∘ B_SU(3) ∘ T_U(1), representing transport, internal binding, externally induced transformation, and stabilization. The model separates representation change from content change, local curvature from global memory, internal mediation from external control, and stability from rigidity. Finally, we formulate a machine-facing layer in which formal objects are translated into computational objects, experimental manipulations, observable metrics, and falsifiable predictions. Gauge symmetry is thereby recast not as an analogy imported from physics but as a candidate structural technology for operational knowledge assembly. Artificial Science is proposed as the science of converting portable deep structures into executable and testable knowledge dynamics.

Abstract: Artificial intelligence is rapidly moving from the accumulation and representation of knowledge toward a more difficult problem: the assembly of heterogeneous bodies of knowledge. Knowledge accumulation is not knowledge integration, and knowledge integration is not yet knowledge assembly. Distinct scientific domains differ in objects, semantics, methods, and local representational frames, even when they share deeper mathematical structures. Building on the program of Integration Science, this paper proposes that such cross-disciplinary structures can be transformed in Artificial Science from descriptive correspondences into operational interfaces for knowledge assembly. Gauge symmetry is developed here as the first structural technology of a broader theory of Big Knowledge Assembly Dynamics. The framework begins with a global-to-local structural principle: for cross-domain knowledge assembly, identification of a common structural group precedes the construction of domain-dependent local gauge transformations. A principal knowledge bundle is introduced as the geometric setting in which a global knowledge connection is represented locally by domain-specific gauge potentials. Covariant derivatives distinguish changes of knowledge content from changes of local frame, while curvature measures irreducible local path dependence in assembly. Domain-specific dynamical phases, already developed in prior models of market, reasoning, economic, and AI dynamics, are lifted into a cross-domain geometric setting in which Berry-type phases represent geometric memory. Yang’s non-integrable phase factor supplies the path-dependent parallel-transport operator, and closed-loop transport yields knowledge holonomy. The framework therefore separates local curvature from global holonomy and allows the possibility of global path memory even when local curvature vanishes. To coordinate the geometry of the knowledge base space with transformations in internal knowledge fibers, a Total Knowledge Connection is proposed by combining an affine connection with an internal gauge connection. This structure supports geodesic and variational descriptions of knowledge-assembly paths. A Geodesic-Pareto principle is then formulated cautiously: when a multi-objective assembly problem induces an effective geometric action, stationary efficient paths may be represented by geodesics of the induced geometry. The resulting concept of covariant Pareto efficiency provides a candidate optimization principle for AI knowledge assembly. The paper concludes with an Artificial Standard Model program in which U(1), SU(3), SU(2), and Higgs-type differentiation organize subsequent studies of single-charge knowledge transport, three-component non-Abelian assembly, doublet coupling, and the emergence of stable local rationality. The proposal is not that AI literally contains physical gauge fields, but that gauge geometry supplies a mathematically disciplined architecture for converting shared structures discovered by Integration Science into operational technologies of Artificial Science.

Abstract: The preceding papers of Principles of Artificial Science I developed a global gauge architecture for Big Knowledge Dynamics and three mechanism-specific models: U(1)-based transport, SU(3)-based internal binding, and SU(2)-based externality-driven transformation. The present paper addresses the final structural problem of this gauge-symmetry sequence: stabilization. How can a knowledge system, after transport, binding, and transformation, settle into a stable local organization without becoming rigid? The physical scaffold is spontaneous symmetry breaking and the Higgs mechanism, but the proposal is structural rather than ontological: human rationality and artificial intelligence are not asserted to be physical Higgs systems. Earlier work on ordinary rationality identified eight principles - high selectivity, subjective certainty, null action, sunk cost, hesitation, better life, face, and emotion - and characterized ordinary rationality by minimum-energy, inertial, and degenerate-background properties. The present paper converts these inherited correspondences into a multi-layer Artificial Science stabilization model. The state is represented by Z=(Φ, M, S), separating a rational configuration, history/memory, and social-affective context. The eight principles are assigned to distinct dynamical roles rather than inserted as undifferentiated terms in a single potential. A symmetry-selected state Φ₀ is described by a vacuum manifold and residual stabilizer H=Stab(Φ₀). Global symmetry breaking is explicitly distinguished from gauge-coupled Higgs stabilization. The latter generates an anisotropic resistance spectrum over transformation directions. Local stability, recoverability, adaptation threshold, and return time are then defined as candidate AI observables. The resulting theory treats artificial ordinary rationality as a stability architecture that suppresses destructive drift while permitting information-bearing change. Stable intelligence is therefore not immobility but symmetry-constrained adaptability.

Abstract: The preceding papers of Principles of Artificial Science I developed a global gauge architecture for Big Knowledge Assembly Dynamics, a U(1)-based mechanism for single-charge knowledge transport, and an SU(3)-based mechanism for non-Abelian internal binding. The present paper develops the next mechanism: externally induced transformation in doublet knowledge systems. The physical scaffold is weak-isospin dynamics, but the proposed theory is not a claim that economic, cognitive, or artificial systems literally instantiate electroweak physics. Rather, SU(2) is used as a mathematically disciplined control architecture for two-state coupling, externally induced transition, order-sensitive intervention, and post-transition memory. Earlier work on economic externality modeled policy, regulation, and other third-party influences as weak-like forces and introduced two isospin structures: achievement versus fear-of-failure, and market behavior versus market aftertaste. Related Integration Science work extended the same architecture to cognitive achievement/fear and reasoning behavior/problem-solving residuals. The present paper upgrades those correspondences into a Big Knowledge Control Dynamics model. A domain externality is lifted by an explicit map from a semantic externality space into the Lie algebra associated with SU(2); the reversible component of transformation acts on a Bloch-type knowledge state, while dissipation, forgetting, noise, and residual feedback are represented by an open-system term. Macroscopic market or logical charge becomes an observable q(ρ) = Tr(ρQ̂), finite intervention order is measured by a group commutator, externality curvature is defined on an intervention parameter space, and closed-loop memory is represented by gauge-covariant holonomy with a gauge-invariant Wilson-type scalar. Aftertaste is modeled as a separate residual state coupled to the doublet dynamics, allowing net residual memory to be distinguished from accumulated memory burden. A supplementary U(1) residual phase provides an Aharonov-Bohm-type representation of path memory without conflating it with the SU(2) transition connection. The Better-Life Principle is reformulated as an asymmetric transition-rate hypothesis rather than an unsupported lifetime assertion. Finally, three testable AI diagnostics are introduced: externality susceptibility, finite-order sensitivity, and long-horizon drift, with extensions to multi-agent interaction and control. The resulting framework specifies state, control, dynamics, memory, and observable quantities, thereby moving from structural analogy toward an operational theory of Artificial Science and Big Knowledge Dynamics.

Abstract: The first paper of Principles of Artificial Science I proposed a global gauge architecture for Big Knowledge Assembly Dynamics. The present paper develops its first concrete mechanism: knowledge assembly under U(1) symmetry. The theoretical scaffold is quantum electrodynamics (QED), but the purpose is not to identify economic or cognitive systems with physical electromagnetic systems. Rather, QED provides a minimal mathematical structure for a single-charge dynamical system: a one-dimensional Abelian phase symmetry, a gauge potential, a covariant derivative, a field strength, and path-dependent phase transport. Previous work has independently constructed U(1)-type local dynamics in market dynamics and reasoning dynamics, within broader programs of economic dynamics and higher-cognition theory. The present paper does not reconstruct those local theories. It asks a new question: can previously separated U(1)-structured knowledge blocks be assembled through one reusable structural interface? We represent a local knowledge block by and introduce a common U(1) assembly template Each domain is related to this common template by a structural map Knowledge assembly therefore does not require QED, market dynamics, and reasoning dynamics to become identical. It requires designated structural relations to be preserved across domain-specific realizations. We formulate a preliminary criterion of U(1)-assemblability in terms of compatibility with covariant differentiation, curvature, and holonomy.The resulting framework is then applied to four contemporary AI problems: compositional generalization, continual learning and catastrophic forgetting, long-horizon trajectory drift, and cross-domain frame mismatch. In each case, the U(1) model suggests a common principle: knowledge should be transformed without confusing changes of representational frame with changes of structural content. The paper therefore interprets U(1) symmetry not merely as a shared mathematical form across disciplines, but as the first candidate technology for covariant AI knowledge assembly.

Abstract: The first two papers of Principles of Artificial Science I developed a gauge-theoretic architecture for Big Knowledge Assembly Dynamics and then implemented its first concrete mechanism through U(1)-based single-charge knowledge assembly. The present paper advances from Abelian transport to non-Abelian internal assembly. Quantum chromodynamics (QCD) supplies the mathematical scaffold: a three-dimensional internal representation, an SU(3) connection, noncommuting generators, path-ordered transport, interaction-dependent curvature, and composite states built from fractionally charged constituents. The purpose is not to identify economic or cognitive systems with subatomic matter, but to convert shared structures previously developed in Integration Science into an operational interface for Artificial Science. Earlier sub-economic dynamics modeled achievement and fear-of-failure as up- and down-type impulses, introduced fractional market charges, represented Freudian Id-Self-Superego structure as a three-color internal space, and assigned consciousness a gluon-like mediating role. The present paper combines these ingredients with a parallel sub-cognitive triplet - truth, belief, and justification - in which language plays the mediating role. A three-index state with indices (a, f, g) separates color-like internal position a, flavor-like type f, and generation-like intensity g. The generation index is linked to Yang’s trichotomy, grounded in earlier mental-logic experiments in which inference-schema weights formed three empirical clusters. We formulate an SU(3) knowledge-assembly interface, weak and strong criteria of assemblability, a Fractional-Charge and Dominance Assembly Principle, and a non-Abelian account of order-sensitive composition, curvature, and holonomy. Proton-like uud states are interpreted as achievement-dominant but risk-aware active states; neutron-like udd states as fear-dominant but still aspirational standing-by states. Their different persistence is connected to the Better-Life Principle of ordinary rationality. The framework concludes by preparing the transition from strong-like internal binding to SU(2)-type externality-driven state transformation in Artificial Science I (4).

Abstract: Artificial intelligence is moving from an external tool of scientific research toward an internal participant in knowledge production. Mathematics is among the first disciplines in which this transition has become visible. In 2026, Terence Tao publicly described AI as a practically useful research tool in mathematics and theoretical physics; later, together with twenty-four other Fields Medalists, he signed “A Severe Misalignment of AI in Mathematics,” warning that an AI culture centered on rapid problem solving could weaken mathematical understanding, intellectual transmission, and the training of future mathematicians. He subsequently highlighted the Foundation for Science and AI Research (SAIR) and its effort to develop open scientific AI infrastructure. We interpret this sequence not as a contradiction, but as evidence that human science and artificial intelligence are moving from a tool relation toward a coupled relation.On this basis, the paper distinguishes three levels: science using AI, science jointly produced by humans and AI, and Artificial Science - the science of artificially constructed knowledge-generating systems. We then introduce Coupled Knowledge Dynamics as a framework for the co-evolution of human science and artificial science. Two conceptual quantities are proposed: cognitive position p, describing where artificial cognition is located within scientific practice, and cognitive momentum q, describing how artificial science drives changes in the trajectory of knowledge. At the formal level, these are represented by self-adjoint operators P̂ and Q̂. If order of intervention changes the resulting scientific state, then [P̂,Q̂] ≠ 0. For a state |Ψ⟩ we define a state-dependent human-AI cognitive action scale κ_HAI(Ψ)=|⟨Ψ|(1/i)[P̂,Q̂]|Ψ⟩|, yielding ΔP·ΔQ ≥ κ_HAI(Ψ)/2.To show that the proposal is mathematically non-vacuous, we construct a minimal two-state toy model in which noncommutativity, order effects, and a saturated Robertson uncertainty relation arise explicitly. The model is then represented projectively on ℂP¹ ≅ S². The framework is quantum-like rather than quantum-physical: its claim concerns noncommuting cognitive operations and coupled scientific dynamics, not microscopic quantum mechanisms.

Abstract: Background/Objectives: Misophonia is characterized by disproportionate emotional and physiological responses to specific "trigger" sounds, often accompanied by self-reported difficulty maintaining focus. Here, we sought to provide objective evidence of such cognitive disruptions by combining behavioral measures of auditory working memory with electroencephalographic (EEG) indices of neural processing. Methods: Fifty-five participants (23 Misophonic, 32 control) completed a pitch-matching WM task under three distractor conditions: individualized trigger sounds, neutral sounds, and silence. EEG was analyzed using a time-frequency approach to assess alpha (8-12 Hz) oscillatory activity during working memory retention at the source level with Independent Component Analyses (ICA). Results: Behaviorally, Misophonic participants exhibited reduced overall working memory performance compared to controls, with no difference between neutral and trigger conditions, suggesting a general deficit in distractor inhibition. EEG results showed lower alpha power across all conditions in the misophonia group, consistent with diminished inhibitory control rather than trigger-specific effects. This lower Alpha appeared to occur for sources in parietal and somatomotor areas. Group differences were apparent in alpha suppression around trigger onset times and alpha enhancements occurring later in retention intervals. Conclusions: Results are consistent with claims made in the literature that Misophonics show impaired attentional control and motor regulation. Implications for Misophonia theoretical models and directions for future research are discussed.

Abstract: Language-processing studies routinely average over the observers whose timing generated the effect. Observer-Dependent Entropy Retrieval (ODER) predicts that semantic access is organized in observer-indexed time, making observer identity part of the empirical object. Using ERP CORE and Natural Stories, this paper tests whether observer-indexed relations can be reconstructed from disjoint observations and which measurement transformations preserve or destroy them. In ERP CORE, timing estimated from one trial half and temporal localization estimated from the other were reunited only through correct participant identity. The symmetric observer-paired relation was r = .393, with a 95% bootstrap interval of [.207, .554] and a shared observer-pairing probability of p = .0004. The relation recurred across repeated trial allocations, survived amplitude control, and remained present without zero rectification. Pooling preserved the canonical semantic-negativity profile while eliminating the observer coordinate required to estimate the relation. Natural Stories revealed two behavioral surfaces. The locked clipped concentration family produced a positive but heterogeneous reader-paired relation. A prospectively specified signed temporal-allocation family produced substantially stronger recovery: all six signed-safe specifications were positive, and the family median was r = .491, with an interval of [.407, .577] and p = .0002. A prospectively frozen compression experiment then reduced ERP peak timing from its native resolution to three ordered categories. Contrary to prediction, compression did not weaken recovery: the native-minus-three-bin contrast was Δz ≈ −.004, with an interval of [−.155, .172]. The result also persisted without rectification. Pooling therefore preserved the canonical effect while destroying the observer relation, whereas severe temporal compression discarded most timing distinctions while preserving it. Measurement transformations must be evaluated by the scientific relations they preserve, not by resolution or retained detail alone. The compression interval does not establish equivalence across resolutions.

Abstract: During dyadic interaction, another person’s unfolding action may proceed successfully as expected, succeed after correction, or ultimately fail, with different implications for how the observer should respond. It remains unclear whether the brain differentiates these action courses while the interaction is still unfolding and whether extensive dyadic motor experience sharpens this differentiation. Here, we recorded EEG from expert table tennis players and novices while they observed three-stroke interactions containing uninterrupted success, corrected success, or unsuccessful action. Experts showed a stronger ERP distinction between corrected success and unsuccessful action than novices. Time-frequency and multivariate analyses further revealed temporally evolving and decodable information about action course, particularly in experts, although expertise-related enhancement was not consistent across measures. Experts were more accurate overall in predicting landing location, but this behavioral advantage did not vary across action courses. These findings show that unfolding action courses can be neurally differentiated during dyadic observation and that extensive motor experience selectively sharpens some of these distinctions. More broadly, neural monitoring during dyadic interaction appears sensitive to how another person’s action develops and to its relevance for how the interaction is likely to continue.

Abstract: Large language models (LLMs) are increasingly used for qualitative coding, yet it remains unclear how closely their coding outputs align with frameworks developed by human experts. Existing studies tend to report overall similarity but rarely distinguish thematic granularity from dimension-level semantic alignment or systematically identify which constructs are stably recovered. This study therefore compares human expert coding with three zero-shot LLM-generated qualitative analyses using a structured framework-level mapping method. The method maps human-coded dimensions to LLM-generated qualitative frameworks and classifies each relationship as direct, partial or merged, or not recovered. To quantify alignment, three evaluation metrics are used: Direct Coverage (DC), which measures strict boundary recovery; Broad Coverage (BC), which includes partial or merged mappings; and Frequency-Weighted Mapping (FWM), which weights each dimension by its human coding frequency to reflect construct prominence. The method is demonstrated using a human-derived emotional-intelligence framework developed from 34 Chinese behavioral-event interviews and three zero-shot frameworks generated from the same corpus by DeepSeek V4.0 PRO, Qwen 3.6, and GPT 5.6 under an identical prompt. Two senior experts jointly evaluated 36 dimension-level relationships. The three LLM outputs achieved direct coverage ranging from 75.0% to 91.7%, broad coverage from 83.3% to 100.0%, and frequency-weighted mapping from 80.3% to 94.9%. Eight of the 12 human-derived dimensions were directly recovered across all three LLM frameworks, whereas guidance and motivation, big-picture awareness, care and support orientation, and teamwork showed variable alignment. The framework-level comparison distinguishes thematic organization from construct-boundary recovery and supports the use of LLMs as complementary analytical assistants while retaining human responsibility for contextual interpretation and construct definition.

Abstract: Understanding how visual exploration evolves over time during complex decision making remains challenging, as the dynamics of underlying cognitive processes are difficult to observe. Hidden Markov models are widely used to model eye movement sequences, but their assumption of stationary transition probabilities limits their ability to identify successive phases. This study investigated whether Switching Hidden Markov Models (SHMMs), trained only on fixation positions, could segment visual exploration during an obstacle-avoidance task requiring a left/right decision. Individual SHMMs with two or three high-level (HI) states were trained without using reaction times. HI states were characterized through transition-matrix variability, saliency maps, cumulative dwell times in regions of interest, and their association with reaction times. Both models produced temporal segmentations consistent with decision making. Early HI states showed greater interindividual variability and exploration focused on obstacles, whereas later states were more homogeneous and reflected decision confirmation followed by reorientation toward the center of the screen. The transition to the final HI state of the three-state model best predicted reaction time. Furthermore, cumulative log-likelihood significantly improved this prediction, indicating that SHMM confidence provided complementary behavioral information. These findings show that SHMMs trained only on fixation positions can identify phases consistent with the organization of the decision-making process, including early evidence accumulation and late decision confirmation.

Abstract: People often recognize spoken words more accurately and more quickly than written words. From the perspective of active inference under the free-energy principle, the present study develops a formal account of this speaking-over-writing advantage in production-modality retrieval. We modeled retrieval under spoken and written production-modality states, together with the selection of an early or late response policy according to its expected sensory consequences. University students spoke or wrote words presented on a computer screen and later judged the modality in which each word had been produced. Responses were followed by explicit visual feedback indicating whether each judgment was correct or incorrect, thereby providing an observable sensory consequence of the recognition response. Response times were implemented as early and late policies within a partially observable Markov decision process (POMDP). The model included modality-specific likelihood parameters linking response timing to feedback outcomes. Spoken words were identified more quickly and accurately than written words, and early responding was more strongly associated with correct feedback under the speaking state than under the writing state. Bayesian model selection favored the POMDP over a simpler linear model estimated using Variational Laplace, indicating that the POMDP provided a better balance of fit and complexity. These findings show how production modality, response timing, and expected sensory consequences can be represented within a common generative framework. The framework provides a basis for future extensions that model uncertainty about production modality, preserve continuous response timing, incorporate learning across trials, and compare active-inference accounts with alternative models of recognition memory.

Abstract: Physical exercise improves working memory (WM) across the lifespan, yet the neural mechanisms underlying this benefit remain incompletely understood, and it is unclear whether intervention-induced neuroplastic changes and long-term exercise-related differences converge on common neural circuits. We systematically searched PubMed, Web of Science, PsycINFO, and CNKI through June 2026 and included 11 task-based fMRI studies (6 longitudinal interventions and 5 cross-sectional comparisons; 403 par-ticipants) reporting whole-brain activation coordinates. Activation likelihood estima-tion (ALE) meta-analyses were performed separately for activation increases and de-creases within each design. Longitudinal studies revealed exercise-induced activation increases in the bilateral cerebellum (posterior lobe, cerebellar tonsil) and decreases in the right thalamus. Cross-sectional studies revealed greater activation in the left middle temporal gyrus (BA 21) and reduced activation in the right cingulate gyrus (BA 24) and left caudate body among long-term exercisers relative to controls. Critically, the two designs yielded spatially non-overlapping patterns. These findings support a du-al-mechanism model in which exercise strengthens task-positive network engagement while optimizing the suppression of task-irrelevant processing, and they suggest that exercise shapes working memory circuitry across distinct, timescale-dependent neural circuits.

Abstract: The hypothesis defended here is as follows. Before the emergence of predication, language consisted only of pre-words used exclusively to request something or call someone over, with only one pre-word occurring in each message and always with the same intonation. How did these meanings become detached from their invariant function and intonation and thereby turn into genuine words? The most plausible possibility is that a recipient was puzzled upon hearing a request for something that was no longer available or a call addressed to someone who was not in the area. The recipient would then repeat the pre-word, but without its conative function and intonation, using it instead as a thema, and would add another former pre-word that now functioned as a rhema or predicate. The transformation this entailed soon required a further innovation: vocal signifiers had to split into two mutually detachable planes, the articulatory-phonetic and the intonational. In this way, each signifier could be perceived as exactly the same both when it constituted a message used to call someone over or to ask for something and when it was integrated with one or more other signifiers into a single intonational pattern.

Abstract: Teacher creativity plays a central role in fostering creative learning environments. However, most evaluations of creativity training focus on performance outcomes and self-report measures rather than changes in underlying cognitive organisation. We investigated whether an eight-month certified teacher-training course based on the Scientific Creativity in Practice (SCIP) framework was associated with changes in Austrian secondary teachers’ creative performance, semantic organisation, and professional mindsets. Seventeen teachers completed pre- and post-training assessments, including the Alternative Uses Task, Verbal Fluency, Word-Sentence-Construction, Behavioural Forma Mentis associations with valence ratings, and written reflections. We assessed originality in the AUT with two independent AI-based systems. Semantic and affective changes were modelled through cognitive networks, mindset streams, and measures of emotional framing. At POST, teachers had higher AUT originality scores and generated largely different concepts than at PRE. They produced broader conceptual repertoires and larger semantic networks, although other structural changes differed across tasks. Affective changes were more modest. Positive valence predominated at both measurement points and became slightly more common, while negative valence declined. Separate PRE and POST mindset streams showed substantial pathway turnover but remained predominantly positive. These findings suggest that creativity training may support changes in how teachers organise and connect professional knowledge, and does not only impact how creatively they perform. Furthermore, our findings show that cognitive network analyses can complement traditional creativity assessments by revealing changes in semantic organisation and professional mindsets that remain invisible in performance scores alone.

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.