Technological progress is eliminating effort from cognitive tasks, reducing cognitive friction—the resistance, effort, or difficulty encountered in performing tasks or making decisions using products, services, or systems. While eliminating friction improves efficiency and immediate performance, it also raises growing concern about loss of competence, a decrease in the satisfaction derived from addressing challenges, and an alienation from our evolutionary cognitive capacities. However, an emerging perspective suggests that we may be witnessing the emergence of new forms of competence through the adaptation and reallocation of cognitive resources. In this paper, we highlight the “good” and “bad” aspects of the use of AI systems, paying particular attention to the concepts of: Performance (immediate and easy results), Learning (acquisition and long-term maintenance), and Expertise (mastery and adaptability). We review the empirical evidence on cognitive offloading—spanning neuroscience, cognitive psychology, and human-computer interaction—and introduce the concept of critical friction as a design principle for AI systems. Based on Polanyi++, a neurosymbolic architecture founded on tacit knowledge extraction, ontology-driven context engineering, and heuristic-driven inference, we propose a layered architecture for critical friction in collaborative human/AI systems, illustrate it through a worked example, and outline a research agenda connecting cognitive science evidence with engineering design.
Post-Traumatic Stress Disorder (PTSD) flashbacks can be triggered by neutral cues that were present during trauma. This paper presents a second-order adaptive temporal–causal network that models cue-driven replay, metaplastic control of learning speed, and therapy-induced gating of both sequence entry and emotion regulation. The adopted causal and agentic modeling approach used in the paper, serves as a conceptual tool for understanding the structure and dynamics of complex adaptive processes. A four-epoch scenario (trauma, pre-therapy trigger, therapy, post-therapy trigger) is instantiated with an air-raid siren. Simulations show that therapy reduces post-therapy affect and action without erasing memory: replay still initiates, but control rises sooner, peaks lower, and decays faster. The model is compiled into role matrices for direct execution, and an AI-generated short animation is used as a communicative front-end to the dynamics. As a validation check, therapy is split into entry-gating and regulation routes: only regulation produces durable post-therapy gains under current parameters. Together, the model and the animation translate a technical mechanism into an interpretable narrative and yield concrete predictions about timing, generalization, and how to strengthen prevention at the entry point.
This paper contributes a conceptual and formal (philosophical) analysis of how causal pathways created by the world drive the development of several types of agency and various causal pathways within agents, for example, those constituting brains and minds. More specifically, it is analysed how especially in context-dense environments the world continuously creates causality in the form of new causal pathways and changes existing ones, which provides a dynamic and adaptive concept of (contextual, embedded) causality. Moreover, it is analysed how such created causal pathways form the basis for the creation of various types of agency by the world, including human agency with its brains and minds. The considered perspective is conceptually based on universal causality structures that also serve as convergence points for evolutionary processes, thus explaining, for example, the phenomenon of convergent evolution describing similar structures that are created via independent evolutionary pathways. These universal causality structures are formalised mathematically by the new concept of temporal causality algebra based on the area of (many-sorted) universal algebra within mathematics.
This paper introduces a formal cognitive architecture in which thinking is modeled as the self-organization of semantic structures within a unified and evolving dynamical environment. In contrast to classical cognitive architectures such as ACT-R, SOAR, and LIDA, which decompose cognition into predefined functional modules, the proposed framework treats cognition as a continuous process emerging from the coupled evolution of activation dynamics and relational structure.The architecture is formalized as a class of ontology-constrained, co-evolving dynamical systems defined over adaptive semantic networks. Within this formulation, both activation states and network topology are dynamical variables, coupled through nonlinear transformations and structurally admissible update operators. This enables the representation of cognitive processes — such as insight, intuitive reasoning, and goal formation — not as externally specified mechanisms, but as emergent properties of the system’s internal dynamics.Unlike symbolic architectures with fixed structures or neural models with static topologies, the proposed framework operates over an evolving configuration space, integrating structural plasticity, activation propagation, and instability-driven reorganization within a single formal system. This provides a unified basis for modeling cognition as a self-organizing process in which structure, meaning, and control co-evolve.The framework is positioned as complementary to existing cognitive architectures while providing a mathematically grounded foundation for modeling intrinsically generated cognitive phenomena. In this perspective, goals, transitions, and conceptual structures arise from the constrained evolution of the system rather than from externally imposed representations or control schemes.
In this article, we argue that Hernández-Orallo and Vold’s (2019) definitions and characterizations of AI systems as cognitive extenders fail to capture the complexity and dynamics involved in AI-enabled cognitive extensions (AI-EXT). More specifically, they assume that AI systems can extend cognition only insofar they are not autonomous and that any form of AI-EXT necessarily enhances the cognitive capacities of the user. Thus, the aim of this article is to overcome these two limitations, offering a more comprehensive framework of AI-EXT capable of accommodating AI autonomy and the variety of effects of cognitive extensions, not reducible to cognitive enhancement. To do so, we first present and argue in favor of an integrationist and multidimensional account of the extended cognition thesis (EXT). We then apply Fasoli’s (2017) taxonomy of cognitive artifacts to distinguish three types of relationships between the agent and the AI system (constitutive, complementary, and substitutive), and we analyze the relationship between AI autonomy and its potential for cognitive extension. We argue that our framework offers a more fine-grained and flexible account of the varieties and dynamics of cognitive extensions via AI systems. To support this claim, we conclude the article presenting real-world cases of AI-EXT.
The literature on Turing's Imitation Game and the “Turing Test” has largely been dominated by debates about the nature and definition of intelligence, or by attempts to classify the proposal either as a ready-to-run empirical experiment or as a philosophical thought experiment. This paper argues that both readings miss a crucial structural feature. Focusing on the 1950 gender-based Imitation Game, we reconstruct it as an idealised experiment with the design-level structure of a Fisherian hypothesis test. Unlike later jury-based formats, the Imitation Game incorporates a human–human control condition and a symmetric decision task that neutralises conservative anti-machine bias. Indistinguishability is measured indirectly, via the interrogator's strategic behaviour relative to a baseline. In this reconstruction, the game does not define intelligence but probes a more specific dispositional ability in linguistic interaction. By shifting attention from definitional debates to experimental structure, the paper offers a novel methodological interpretation of Turing's proposal and its limitations. We argue that this reconstructed framework provides a rigorous methodological template that contemporary AI researchers should adopt—or explicitly depart from—when designing behavioural evaluations of machine intelligence.
What counts as memory depends on authorization by a governing regime, where persistence alone cannot settle reportability. A jurisdictional model separates candidate formation from mnemonic standing. Storage, reconstruction, prediction, and retrieval explain how candidates arise without settling why one becomes reportable while another stays silent, a gap closed by admissibility: the condition under which content counts as recall. The primary object is engineered memory in large language model systems, with biological remembering entering as a comparative constraint that tests the framework against substrates not directly inspectable in the same architectural sense. Standing in LLM architectures is assigned across three layered loci, each operating under a separate authority that admits a distinct mode of revision. Parametric priors fix durable regularities that operate ahead of any user interaction, whereas at runtime dynamic buffers rely on scope and policy rules to govern re-entry, and interactional persistence forms a third, user-facing locus in which standing becomes partly available to user control. Persistent memory in ChatGPT provides a clear product-level case in which saving and deletion alter future eligibility without changing weights, whereas activation inserts stored content as an operative constraint before local ranking is completed. The framework clarifies why declared control can depart from effective priority, and why reportability can vary independently of representational support, a divergence that the jurisdictional model accommodates without displacing mechanistic or functionalist accounts.
We investigated the neuro-behavioural correlates of cognitive adaptability in 36 young, healthy volunteers using an ecologically valid visuo-spatial working memory paradigm. All volunteers showed a stable performance (approximate to 71% accuracy) with consistent reaction times across loads, suggesting effective adaptation to increasing task demands (2-8 items in memory). Various aspects of EEG dynamics were studied (power spectra, frequency sliding, and functional connectivity) for correct trials during encoding and delay periods of the task. Consistent changes observed include: theta power increased during encoding in the fronto-central and tempero-parietal clusters relative to the baseline and decreased (globally) during the delay period. The fronto-central alpha power increased during encoding and delay periods, and the theta frequency dynamics were shifted to a faster rhythm with a predominately parietal distribution during both the encoding and delay periods. The alpha operating frequency was shifted to slower frequencies in the posterior regions during encoding, but the alpha rhythm was shifted to a faster frequency in the fronto-central regions during the delay period. Functional connectivity showed a network reconfiguration that facilitated global processing associated with behavioural performances. Such a flexible brain network integration and temporal oscillatory dynamics would ensure cognitive adaptability to perform the task with an optimal working memory capacity.
In this article, we explore the possibility that machines can develop consciousness and analyze various approaches to defining and modeling consciousness. We present seven theoretical approaches, including ‘consciousness’ as language ability, sensory perception, social interaction, or as an emergent phenomenon in complex systems. We examine these perspectives in terms of their feasibility in artificial systems, for example through neural networks, embodied cognition, or self-referential processes. However, the very success of such efforts leads us to a paradox: on the one hand, there is no reason not to implement consciousness processes in technical systems, but on the other hand, the gap between objective modelability and subjective introspective quality remains unbridgeable. But since this also applies to the observation of our brains – no one has yet discovered subjective ‘introspection’ in physiological processes – there is no reason why silicon-based cognitive systems cannot also ‘experience’. This has ethical implications: if machines develop sentience, this requires a reassessment of our moral responsibility. Our contribution shows that AI with consciousness poses not only technical challenges, but also philosophical and ethical ones. We advocate an interdisciplinary discourse that combines technical innovation with social responsibility in order to reflect on how to deal with potentially ‘conscious’ machines.
In general sum multi-player competitive game theory, players dynamically form and disband coalitions as they identify joint utility. However, human players can invalidate this assumption by biasing strategies with emotions and perceptions. This paper shows how machine agents can leverage a perceptual and interpretive team cognitive model to capture these biases and adapt their action strategies. Opponents are represented as non-cooperative team members and the problem is recharacterized as a general sum, multi-player game with imperfect information over player intentions. Agent performance is analyzed via qualitative and quantitative methods in the game of Diplomacy. The results demonstrate that these agents can form cognitive representations of other players that encode elements of structural, perceptual, and interpretive emergence. Additionally, the results suggest that these cognitive representations construct a causal thread throughout game play, that provide intent projections about other players when joining or defecting from coalitions.
The training process of feed-forward neural networks is a slow and computationally intensive procedure mainly due to the iterative nature of most algorithms. A solution to this problem was the creation of the extreme learning machine (ELM) algorithm for single-layer neural networks (SLNNs). This method uses a very fast approach where the hidden-layer weights and thresholds are randomized, and the output layer's weights are analytically calculated using the Moore-Penrose pseudo-inverse. Although it provides good generalization performance, it is restricted in traditional neuron types where each neuron's input is multiplied by its corresponding weight. On the other hand, most ELM variants have focused on algorithmic enhancements such as robustness or parameter tuning, without considering the integration of higher-order or multi-cube neurons into the hidden and output layers while preserving ELM's single-pass training speed. This paper introduces six ELM architectures for single-layer neural networks (SLNNs) that replace low-order units with higher-order (sigma-pi) and multi-cube variants (the latter enable controllable expressivity without exponential weight growth). The advantage of higher-order units lies in their ability to utilize more weights than traditional neurons, thereby overcoming the linear separability limitation of low-order units. The experimental results indicated that higher-order SLNN variants demonstrated better generalization performance compared to SLNNs trained with the traditional ELM and online sequential ELM (OS-ELM) algorithms. This observation experimentally verified across 15 classification datasets and eight regression datasets.
Human reasoning has long been studied through two lenses: normative logical frameworks that prescribe ideal inference and descriptive cognitive models that explain how people actually reason with limited knowledge and resources. In parallel, artificial intelligence (AI) has evolved along symbolic and learning-based paths. This paper advances a three-part integration by developing a unified framework for reasoning under limited knowledge and resources. Building on bounded and resource rationality, we introduce the Relative Rationality Principle (RRP): reasoning quality should be evaluated relative to an agent’s informational state and computational budget, rather than against unconstrained ideals. We operationalize RRP as a metacognitive control problem in which systems allocate budget across inference, evidence acquisition, verification, and stopping/abstention, yielding testable design propositions under a unified budget definition. Mechanistically, we propose a unified inference core grounded in Non-Axiomatic Logic (NAL). At the methodological level, we specify a reproducible evaluation protocol that characterizes bounded reasoning along three complementary dimensions: counterfactual stability, systematic generalization, and confidence reliability/justifiability. This paper provides implementable objects and auditable metrics intended to enable future comparative evaluation and failure-mode diagnosis under consistent task and budget specifications.
Analogical reasoning is one of the core mechanisms underlying human cognition and creative thinking, yet modeling its dynamic similarity remains a key challenge for artificial intelligence. Existing models typically treat context as static background information and model dynamic similarity through weight assignment and parameter adjustment. However, such approaches tend to overlook the reconstruction and updating of context during the reasoning process. From a dynamic semantics perspective, this paper systematically examines contextual information in analogical reasoning, distinguishing between context as background information and three types of context that function as objects of discursive influence. On this basis, AGM belief revision theory is introduced to characterize the interactive updating mechanisms among contexts conceived as belief sets in analogical reasoning. Analogical reasoning is thus reconstructed as a process in which stage-wise consensual beliefs emerge through the expansion, contraction, and revision of contextual belief sets themselves. This perspective deepens the understanding of the dynamic and context-sensitive nature of analogical reasoning and provides a more explanatory analytical framework for erroneous analogies, such as over-analogization.
We present a cognitive-architecture framework for integrating large language models (LLMs) with self-regulated learning (SRL)-informed tutoring, instantiated in a web-based Virtual Tutor for undergraduate essay writing. The framework embeds an LLM within the emotional Biologically Inspired Cognitive Architecture (eBICA), enabling feedback and dialogue acts to be guided by an explicit learner state rather than generated ad hoc. This state incorporates task goals, writing progress, interaction history, and affective indicators. Tutoring policies are represented as moral schemas that encode pedagogical narratives and socio-emotional norms, supporting consistent, context-sensitive interventions such as planning prompts, rubric-based self-monitoring, and reflective questioning. The system includes four modules: (1) essay-structure and rubric-coverage visualization; (2) an interface for iterative drafting and clarification; (3) cognitive reasoning for learner-state updating and SRLaligned strategy selection; and (4) LLM-based generation of explanations, examples, and revision suggestions. We evaluated the tutor in a within-subject classroom pilot with 50 students, each writing one tutor-assisted and one independent essay. Essays were automatically scored with GPT-4.1 using a 16-criterion analytic rubric, with mean overall rubric score as the primary outcome. A linear mixed-effects model showed a significant tutor effect of 3.44 points, 95% CI [1.56, 5.32], p = 0.0006, Cohen's dz = 0.48, with no significant interactions, including Tutor & times; Order. The solution is browser-accessible, requires no specialized hardware, and is implemented in Python for scalable deployment. Future work will strengthen longitudinal evaluation of retention and transfer, refine learner modeling, and incorporate multimodal signals to better infer intensions and affect.
Traditional crowd simulation models prioritize evacuations, neglecting the cognitive mechanisms underlying helping behaviors. We present CogHelp, a cognitive architecture using a game-theoretic utility function to arbitrate among helping, observing, and evacuating. The model quantifies emotional states via real-time risk and contagion, formalizes curiosity using information gap theory, and assesses social evaluation through the bystander effect and social identity theory. Empirical reconstruction of a real-world fire incident and forward predictive simulations validate the architecture. Bivariate parameter coupling analysis reveals a dual-layered mechanism: the engagement growth parameter temporally regulates helping initiation, while conscientiousness gates the mobilization scale. A threshold effect produces a cognitive stagnation zone below critical parameters, suppressing helping. Predictive simulations forecast sustained observing states, demonstrating that low conscientiousness limits mobilization. CogHelp provides a computational tool for inferring cognitive drivers of helping behavior in high-risk environments.
In this paper, we employ the Minimal Cognitive Grid (MCG), a framework created to evaluate the cognitive plausibility of artificial systems, to offer a systematic assessment of leading computational models of analogy and metaphor, including the Structure-Mapping Engine (SME), CogSketch, METCL, and Large Language Models (LLMs). We present a formal and quantitative operationalization of the MCG framework and, through the analysis of its three main dimensions (Functional/Structural Ratio, Generality, and Performance Match), examine how well each system aligns with standard cognitive theories of the modeled phenomena, thus allowing for comparison of the models with respect to their cognitive plausibility, according to consistent and generalizable mathematical criteria.
Working memory (WM) is an important component of human cognition that shapes reasoning, planning, learning, language understanding, and language generation. Given its importance to humans, we believe that WM could also play a key role in robot cognition, especially for language-based processes. In this paper, we take inspiration from human WM models to propose three different robot WM configurations, each of which differently distributes information across a robot cognitive architecture. In addition to defining these configurations, we propose a set of recommendations for parameterizing the forgetting mechanisms that manage these WM systems in order to optimize WM-facilitated referring expression generation, by promoting lexical entrainment while avoiding the communication of outdated information. Finally, we step through proofs-of-concept that demonstrate how our approach achieves these aims while also improving space and time efficiency, and provides the foundation for future experimental and architectural research.
This article introduces a two-agent model to simulate conflict between an anxiously attached partner (Agent A) and an avoidantly attached partner (Agent B). The agent model is based on attachment theory and integrates elements from affective neuroscience and adaptive dynamical systems for showing how internal working models, emotional regulation and interpersonal feedback influence relational conflict. The agent model was designed as a multi-order adaptive architecture with states that represent emotional, behavioral and regulatory components. The agent model learns from co-activated relational experiences. The simulation explores the fight phase of conflict escalation which demonstrates specific patterns that occur when anxious and avoidant partners interact. Agent A shows quick sensory reactivity, schema activation and protest preparation, but Agent B shows emotional shutdown and belief-driven withdrawal. The agent model demonstrates certain interaction mechanisms between partners where distress from one partner strengthens defensive behaviors in the other partner which matches with observed relational patterns from empirical research. The addressed scenario and related simulation outcomes have been used as scripts for AI-generated videos to show the patterns in an illustrative manner. The research demonstrates how at a detailed level adaptive causal network modeling of the underlying dynamical systems helps to establish more insight and knowledge about attachment-based interpersonal behavior and when displayed in a relatable manner could lead to future therapeutic simulation development.
The motivation to belong and the capacity for social consciousness are key elements of human cognition. This paper presents a computational model that enables an artificial agent to assess its sense of belonging to a group and update this belief based on social feedback. The model is based on three central ideas: firstly, belief updates depend on the detection of inconsistencies between internal assumptions and external feedback; secondly, consciousness of such inconsistencies only arises when an evidence threshold is exceeded; and thirdly, the credibility of the feedback depends on source reputation. Drawing on theories of mind and false belief, as well as theories of consciousness, social cognition, and motivation, the model provides a functional framework through which to model the process by which agents become aware of their inclusion. The proposed computational model is validated through a case study, implemented and described using the ODD+D protocol, in which agents revise their beliefs about group inclusion through social interaction. The results demonstrate how consciousness, feedback, and social reasoning can be combined to produce behavior that is more akin to that of humans in artificial systems.
The motivation to belong is a fundamental social and psychological need that influences individuals' well-being and behavior in society. There is currently significant research activity in this field, driven by increasing interest in areas such as social robotics, the Internet of Things (IoT), and complex systems. This article presents a computational model that quantifies a weighted Belonging Score (B) as a function of seven socio-cognitive features: group identity, social interaction, acceptance/rejection, social memory, reciprocity, reputation, and utility. The aim is to propose a model that explains how agents' motivation to belong emerges from their social interactions. The behavior of each model component is formally defined through mathematical functions. The inputs of these functions are normalized and dynamically updated in real time based on agent-agent interactions, ensuring is an element of [-1, 1] and enabling comparability across contexts. To validate the proposed model, realistic, nuanced, and context-sensitive scenarios were simulated using a large language model (LLM). In this setup, interactions among agents naturally vary the values of the variables determining each agent's internal belonging score. Consequently, agents dynamically assess their perception of inclusion or exclusion within the group.
