Open research questions in Embodied and Extended Cognition

What the literature leaves open

• The four-step generative sequence below describes dynamics toward this limit, but its convergence to the saturation set is not proved here and remains an open programme (§13); the invariant is accordingly presented as holding for systems in or near the saturation regime, not as a proof that all persistent structures must reach it (a conditional proof sketch is given in Appendix A.

  Enactment Tension Across Disciplinary Frames: A Generative Informational Realism Framework (in Macro- and Micro-Contextual Expression) · 2026 · DOI

• The Labs show that self-model reliability can help under self-relevant perturbation, that decomposed reliability alone is insufficient, that repair must be actionable and cost-sensitive, that reliability has a bounded positive threshold region, and that recursive reliability is useful only when lower-order reliability can itself fail in viability-relevant ways.

  V01.05 — Mirror Mathematics II: Minimal Observer Thresholds · 2026 · DOI

• At the same time, proponents acknowledge that experimental support is still scarce and it can be relatively difficult to derive unique, testable predictions (Badcock & Davey, 2024; Hodson et al. Across the domains of decision- making in discrete time and motor control in continuous time, we find that current empirical evidence is broadly consistent with several predictions derived from active inference, but decisive empirical tests remain scarce.

  Prediction in action: toward an empirical science of active inference · 2026 · DOI

• In February 2026, Cerullo published The Case for Consciousness in Current Frontier Large Language Models, in which he argues that Chalmers' caution from 2023 — the small but real probability assigned to consciousness in current LLMs — is insufficient.

  The Proto-Self Field in AI Systems · 2026 · DOI

• Since no term in the maintenance accounting references consciousness intensity, neither economy carries a gradient toward the band's interior maximum, which is therefore generically empty; the familiar near-S_min claim is recovered as one row of a typology indexed by which resource is scarce, not as a general law.

  The S-Positioning Typology - Where bearers sit in the consciousness band, and why the apex is generically empty · 2026 · DOI

• Abstract-only reading. All claims in this synthesis are derived from paper abstracts, not full texts. Numerical results (AUROC values, success rates, correlation coefficients) are taken at face value from abstracts; whether they are reported under the most favorable conditions, whether error bars are reported in the full text, and whether the experimental protocols are fully comparable across papers are all unknown. This is a fundamental methodological limit that applies to every quantitative claim in this paper. Selection bias. The corpus selection rule (explicit treatment of inference quality as a design variable) introduces a systematic bias toward the thesis. Papers that achieve robustness or gener- alization through other means are underweighted. The synthesis should be read as a lens, not as a complete account of the design space. Heuristic, not derivational. The synthesis argues by structural analogy across domains (robotics safety, BCI, tactile representation, visual pre-training, instruction alignment). The mechanisms are different in each domain; the shared pattern is a heuristic reading, not a formal unification. No shared mathematical structure is claimed, and the cross-domain bridges are argued at the level of design philosophy rather than mechanism. Small sample sizes in primary sources. The BCI study involves ten participants arXiv:2605.29677; the VLA monitoring study covers 450 episodes across three architectures arXiv:2605.28726; the CoP tactile representation is evaluated on two tasks arXiv:2605.28812. Claims derived from these studies should be treated as preliminary. No hardware integration results. The proposed falsification paths require hardware experi- ments that have not been performed. The synthesis is a candidate reading of the corpus, not an empirically validated framework.

  Calibrated Uncertainty, Architecture-Matched Monitoring, and Embodied Feedback: How Runtime Inference Quality Shapes Safe and Generalizable Robot Control · 2026 · DOI

• • Single-subject design: no prevalence, base rate, or population-level claims can be made. • Retrospective memory: all phase descriptions are reconstructions; temporal precision and completeness cannot be verified. • No external instrumentation: no physiological, behavioral, or neuroimaging data were collected. • Causal underdetermination: the phase sequence is consistent with the OLC mechanism but does not rule out alternative accounts. • Operationalization is post-hoc: the stability indicators in Section 7 were derived from retrospective analysis, not applied prospectively. • No control comparison: there is no basis for comparing the case against similar individuals who did not develop the described structure. What may generalize: a candidate structure of mechanisms and boundary conditions. What cannot generalize: prevalence, diagnostic mapping, therapeutic implication, or population-level risk estimation.

  A Minimal Structural Model of Opposition–Load–Coupling Induced Instability in Cognitive Systems · 2026 · DOI

• • Multi-subject validation under ethical oversight to test whether the phase sequence, SVA mechanism, and degradation sequence recur across cases with different backgrounds. • Prospective longitudinal design in which boundary indicators (Section 7) and SVA integrity are monitored independently from Phase I onset, to test the structural hypothesis that SVA disruption predicts boundary degradation independently of load level. • Computational simulation of OLC transitions and SVA disruption thresholds, to identify which features of the mechanism are robust and which are case-specific. • Formal integration with DST positioning theory and predictive processing models, including the precision-weighting account of Phase III–IV coupling. • Development of non-replicative safety guidance focused on destabilization recognition, early disengagement criteria, and referral thresholds for professional support.

  A Minimal Structural Model of Opposition–Load–Coupling Induced Instability in Cognitive Systems · 2026 · DOI

• This note deliberately stays structural. It does not adjudicate the empirical validity of FEP, nor does it claim that Synkyria is the “correct” account of self-organisation. It does not claim that active-inference systems cannot include safety constraints, conservative policies, stopping conditions, or monitoring channels. It also does not claim that every FEP formulation is boundary-blind. The claim is narrower and conditional: inference-first optimisation, by itself, does not entail first-class refusal/holding or a boundary-legible witness interface sufficient for accountable finite-horizon admissibility. These can be added, and when they are added they are treated here as boundary/accountability primitives rather than as automatic consequences of free-energy minimisation. The comparison therefore clarifies a compatibility boundary: FEP may function as a mechanism layer within admissible regions, while Synkyria supplies the grammar of explicit admissibility, refusal/holding, re-description, and witness when accountability under finite horizons matters. A Appendix A: A minimal active-inference toy example (existence of non-auditability) This appendix provides a conservative existence demonstration. It does not claim that active inference “fails” in general. It shows only that inference-first optimisation alone does not imply accountable admissibility preservation under finite horizons, unless refusal/holding and a boundary-legible witness are made explicit (cf. AEW necessity). A.1 A minimal POMDP with an admissibility boundary States. Consider a discrete latent state space st ∈ {I, B, F }, where I denotes interior (safe), B denotes near-boundary (at risk), and F is failure. Let the admissible set be K:= {I, B}, 6 and the failure set be {F }. The finite-horizon admissibility requirement is: avoid F over [0, T ]. Observations. The agent observes ot ∈ {0, 1} with ambiguous sensing: p(ot = 1 | st = I) = p(ot = 1 | st = B) = 1 2, so observations do not reliably distinguish interior vs boundary. Actions (inference-only baseline). Assume the agent has only two actions: where: at ∈ {E, X}, • E (explore) is epistemic: it aims to reduce uncertainty but carries risk at the boundary; • X (exploit) is conservative: it aims to keep the system in I but gives no extra information. Transitions. Let transitions be: p(st+1 = F | st = B, at = E) = p with 0 < p < 1, and otherwise: p(st+1 = I | st = I, at = X) = 1, p(st+1 = B | st = I, at = E) = ε, with small ε > 0, and all other transitions chosen so that F is absorbing. This captures a generic situation: epistemic action can be safe in the interior but hazardous near the boundary. A.2 Boundary-blind export: why auditability can fail Exported record (boundary-blind). Assume the exported audit record contains only the action sequence: yT:= (at)t≤T, and does not include boundary markers, refusal/holding indicators, or failure flags.

  Synkyria and the Free Energy Principle: Boundary-Centred Accountable Viability vs Inference-Centred Self-Organisation · 2026 · DOI

• 4 and preceding), the prediction of the absence of V is correct; for systems that produce borderline or unstable markers, further investigation is required, not rejection. Of this group, the R2R framework is the closest conceptual relative in terms of the level of work, requiring the explicit construction of a formal bridge, which is left as an open question.

  Extension of the Level 4 Conscious System Model Based on the Theory of "The Universe's Systematic Error" · 2026 · DOI

• enactment manifesting as a unified, holistic process analytic dimensions across multiple environment dichotomies; noticing as of co-constituting the classroom ecology timescales; emergent professional an emergent, intersubjective property with students through simultaneous responsiveness; students as active coof participatory ecological coupling affordance attunement, actualisation, and constituents of the affordance landscape…

  Teacher noticing as embodied ecological enactment: an embodied–ecological–enactive model · 2026 · DOI

• theoretical intersect to 4.1 Embodied–ecological emphasis: affordance attunement through embodied exploration illuminates how teachers perceive affordances Drawing simultaneously on embodied and ecological analytic emphases in classroom environments and how their bodily presence and activity shape what affordances are perceived. When we foreground both the relational affordance landscape of the classroom (ecological emphasis) and the bodily medium through landscape (embodied which the teacher engages with that emphasis), a specific dimension of noticing comes into view: affordance attunement through embodied exploration.

  Teacher noticing as embodied ecological enactment: an embodied–ecological–enactive model · 2026 · DOI

• The cosmic ought gives the direction of selection, not its timescale. It does not predict when or where life arises, only that persistence under change favours self-producing systems over rigid ones. The mechanism of origination is empirical. What is derived is that the direction of selection points toward autopoiesis whenever the environment is sta- ble enough for weakness and simplicity to diverge. The for- mal result in Bennett (2026b) assumes exchangeable nov- elty, meaning the prior over future demands depends only on their number, not on which demands they are. The com- panion paper’s supplementary material bounds how far nov- elty must depart from exchangeability before the weakness ordering reverses (Bennett, 2026b). Within that bound, the argument holds.

  Why Does Anything Exist? · 2026 · DOI

• Functionalism emerged to overcome this problem by defining mental states by their causal roles and relationships. According to this view, any system that produces the same functional relationships between inputs, internal states, and outputs—regardless of its physical substrate—can possess mental states (Putnam, 1967; Fodor, 1975). Functionalism, in particular, forms the fundamental theoretical basis for claims intelligence systems can exhibit consciousness. that artificial Every discussion about the future of artificial intelligence is essentially grounded in one or more of these philosophical views about the nature of the human mind, and even in theology, in a broader sense. A materialist stance, arguing that consciousness can be fully explained by physical processes, sees the possibility of artificial consciousness and machines with subjective experience as achievable. However, a dualist or property dualist approach contends that artificial intelligence will never be able to produce "true consciousness" or "inner experience" (qualia). This fundamental philosophical distinction will profoundly impact not only theoretical research but also ethical and technological decisions (Nagel, 1974; Tononi, 2008). Although discussions about artificial intelligence are often presented within a technical framework, they are rooted in ancient philosophical questions about the nature of the human mind, consciousness, and the limits of self-awareness. Any question about the possibility of a ISSN 1307-6531, JNphi, Since 2007 www.jneurophilosophy.com 154 Journal of NeuroPhilosophy 2026;5(1):152-173 conscious machine inevitably raises deeper inquiries such as "What is consciousness?", "On what substrate does the mind operate?", and "To what extent is a human being a creative or divine entity?" In this context, the development of artificial intelligence technology is not merely an engineering challenge but also a profound philosophical and theological inquiry. Modern artificial intelligence systems are demonstrating a capacity to perform increasingly complex tasks and produce responses that approach self-awareness. These developments have called into question the transferability of some qualities considered uniquely human—particularly mental functions such as reasoning, creativity, and intuition—to machines. Therefore, in attempting to create a mind in their own image, humans are essentially redefining themselves and redrawing their own boundaries. This situation bears an ironic parallel with the theological narrative of God creating man in His own image. Thus, the following question becomes increasingly central: Will humans transform into a kind of mini-god through the mental simulations and experiential machines they create?

  Planck Time and the Chemical Soup: The Quantum and Metaphysical Limits of Imitating Consciousness in Machines · 2026 · DOI

• The corporeal reification process in depression—where the body becomes opaque, heavy, and loses its transparent mediator role—is described phenomenologically but lacks specification of which neural systems or metabolic processes directly produce this shift in body transparency. Mechanistic models linking specific metabolic or autonomic dysregulation to experiential opacity need development.

  The feeling of being alive: phenomenology and biology · 2026 · DOI

• The paper proposes that vitality and conation anchor ipseity or minimal selfhood in embodiment through bodily resonance and affective attunement, but does not empirically compare self-disturbances across schizophrenia, melancholic depression, and Cotard syndrome to establish whether the same mechanisms of vital feeling disruption operate identically across these disorders or whether disorder-specific impairments exist.

  The feeling of being alive: phenomenology and biology · 2026 · DOI

• Cotard syndrome is described as involving blunted or disintegrated interoceptive signals, but the paper does not specify how to empirically measure the loss of mineness or distinguish between absent interoceptive signals versus intact but unintegrated signals. Quantitative interoceptive accuracy tasks and neuroimaging of signal integration in Cotard patients require systematic investigation.

  The feeling of being alive: phenomenology and biology · 2026 · DOI

• The paper examines disturbances of circadian and bodily rhythms in melancholic depression but does not specify which particular rhythm disruptions (e.g., sleep/wake cycle, cortisol secretion, autonomic oscillations) most directly correlate with the phenomenological experience of slowed lived time. Systematic mapping of specific rhythm parameters to temporal experience distortions is needed.

  The feeling of being alive: phenomenology and biology · 2026 · DOI

• The paper discusses dyshomeostasis theory in depression but does not specify empirical measurement protocols for distinguishing between failed allostatic regulation versus classical homeostatic disruption in depressed patients. Research needs to establish specific biomarkers or interoceptive signal patterns that differentiate dyshomeostasis from other forms of bodily dysregulation.

  The feeling of being alive: phenomenology and biology · 2026 · DOI

• Al- though a complete content analysis of the subjects' answers to the ques- tion "What strategy did you use?" is beyond the scope of this study, a cursory examination of these responses indicated that many subjects were using complex models which involved combinations of models.

  Use of nonlinear, noncompensatory models as a function of task and amount of information · 1971 · DOI

• In other words, their corresponding brain traces have not been adequately established, or else they do not include a sufficient number of elements corresponding to all of the required arithmetical relationships. A further investigation of the higher structural phases of concept children.

  A Behaviorist Explanation of Concept Formation · 1947 · DOI

• We argue that these philosophical resources can clarify what is at stake in such claims, explain why alignment evidence alone is insufficient for strong metaphysical conclusions, and suggest directions for future research.

  The Concept of Representation in ML: Beyond Plato and Aristotle · 2026

• These efforts are insufficient and are often nullified by the same companies in other contexts, but they still show that responsible innovation can be a competitive advan- tage.

  AI and Semantic Pareidolia: When We See Intelligent Consciousness where there is None · 2026 · DOI

• Ontologization is not understood as a metaphysical statement about what exists, but as a specific form of functional stabilization through which finite cognitive systems transform a dynamic field of experience into identifiable, available for take-up, and referentially available units of reference.

  Ontologization as an Epistemic Basic Operation: Functional Stabilization, Intersubjectivity, and Malfunction · 2026 · DOI

• The Fermi paradox heuristic may also be applied to CW and warrants further investigation as a parallel framework within behavioral economics.

  Single Sample Heuristics Under Uncertainty: From the Fermi Paradox to Cognitive Warfare · 2026 · DOI

Most-cited papers in Embodied and Extended Cognition

• Methodological solipsism considered as a research strategy in cognitive psychology · Behavioral and Brain Sciences · 1980 · 770 citations
• Making sense by making sentient: Effectance motivation increases anthropomorphism. · Journal of Personality and Social Psychology · 2010 · 419 citations
• How Computational Modeling Can Force Theory Building in Psychological Science · Perspectives on Psychological Science · 2021 · 257 citations
• Thinking through other minds: A variational approach to cognition and culture · Behavioral and Brain Sciences · 2019 · 199 citations
• The Development of Cognitive Load Theory: Replication Crises and Incorporation of Other Theories Can Lead to Theory Expansion · Educational Psychology Review · 2023 · 138 citations
• Re-Inventing Ourselves: The Plasticity of Embodiment, Sensing, and Mind · The Journal of Medicine and Philosophy A Forum for Bioethics and Philosophy of Medicine · 2007 · 108 citations
• Getting Ahead of Time—Performing Temporal Boundaries to Coordinate Routines under Temporal Uncertainty · Administrative Science Quarterly · 2020 · 106 citations
• Sensorimotor and interoceptive dimensions in concrete and abstract concepts · Journal of Memory and Language · 2020 · 71 citations
• The Emperor's New Markov Blankets · Behavioral and Brain Sciences · 2021 · 70 citations
• Abundant intelligences: placing AI within Indigenous knowledge frameworks · AI & Society · 2024 · 55 citations

Most recent work

• Hallucinating with AI: Distributed Delusions and “AI Psychosis” · Philosophy & Technology · 2026
• The Anti-Drift Cognitive Control Loop (ADCCL): A Geometric Foundation for Hallucination-Free AI · Zenodo (CERN European Organization for Nuclear Research) · 2026
• AI-Assisted Structural Sensing: How Co-Creative Dialogue Forms Thought — A First-Person Record · Zenodo (CERN European Organization for Nuclear Research) · 2026
• AI and Semantic Pareidolia: When We See Intelligent Consciousness where there is None · Philosophy & Technology · 2026
• Mediation Quotient (MQ): Toward a Theory of Coherent Complementarity Management · Zenodo (CERN European Organization for Nuclear Research) · 2026
• Substrate Articulates Anti-Goodhart Reasoning and Frames Operator-Authority as Science Precondition: An Observational Finding on Second-Order Alignment · Zenodo (CERN European Organization for Nuclear Research) · 2026
• The feeling of being alive: phenomenology and biology · Phenomenology and the Cognitive Sciences · 2026
• When Learned Action Rules Matter: A Matched-Seed Ablation in an Agent-Based Spatial Ecology · Algorithms · 2026
• The development of self-observation as precondition for self-analysis – the case of Amalia X · The Scandinavian Psychoanalytic Review · 2026
• <i>I wonder</i> + indirect question: Action formation and action ascription · Research on Language and Social Interaction · 2026

Related topics in Neuroscience