Models of Consciousness 7 Copenhagen Recap and the Consensus Project

Models of Consciousness 7 (MoC7), the flagship conference of the Association for Mathematical Consciousness Science (AMCS), concluded yesterday, Friday October 16, at the University of Copenhagen’s North Campus (HCØ and the Niels Bohr Building). Across five full days from October 12 to 16, the official MoC7 schedule and book of abstracts catalogued 120 numbered talks, including seven invited keynotes, six showcase addresses, two plenary lectures, 104 contributed talks across 28 parallel sessions, and 68 poster presentations.

Where this site’s earlier June MoC7 Copenhagen announcement, July consensus programme overview, and August keynote speaker preview mapped the initial speaker list, the completed programme reveals how mathematical consciousness science has matured over the past year. Running in the exact same October week as The Science of Consciousness 2026 in San Diego, MoC7 served as the field’s formal and structural counterweight, replacing metaphysical standoffs with natural-kind clustering, category-theoretic compositionality, and explicit indicator audits for artificial systems.

The Seven Keynotes, Two Plenaries, and Six Showcases

The backbone of the Copenhagen timetable was a sequence of morning and early-afternoon plenary blocks in Auditorium 1 (HCØ), paired with a Wednesday evening public prize lecture at the Louisiana Museum of Modern Art and a Friday afternoon closing keynote in the Niels Bohr Building’s Margrethe Bohr Salen. Note that Lauren Ross of UC Irvine, who appeared on early summer speaker announcements, was not on the final October timetable; Mirja Hartimo of the University of Helsinki and Guillaume Dumas of the University of Montreal anchored the Thursday and Friday morning keynote slots.

Natural Kinds, Prediction Maps, and the Transdisciplinary Consensus Track

Three talks in the opening forty-eight hours defined the methodological agenda of MoC7. Megan Peters opened the conference on Monday morning (#1) by attacking the circularity problem in consciousness measurement. Instead of assuming we already possess a gold-standard operational definition of consciousness and calibrating neural or machine markers against it, Peters proposed a proximity-guided iterative “natural kinds” strategy that clusters functional, physiological, and metacognitive observables and refines the taxonomy as clusters stabilize. That approach complements the formal indicator work covered in our analysis of the soft INUS analogy framework and van Rooyen’s Cross-Substrate Access Assay.

On Tuesday morning (#21), Liad Mudrik followed with her work on building quantitative “prediction maps” of theories of consciousness, extending the adversarial collaboration discipline of the Templeton COGITATE consortium into a fine-grained matrix of where Integrated Information Theory, Global Neuronal Workspace Theory, Recurrent Processing Theory, and Higher-Order Theories actually make divergent anatomical and temporal predictions. Right after lunch on Tuesday (#34), Jakob Hohwy presented From self-evidencing to consciousness, tracing how a predictive-processing agent that minimizes variational free energy inside a Markov blanket generates perspectival awareness, a theme that pairs directly with Karl Friston and Thomas Parr’s September 2026 paper on the brain as a phantastic organ.

Between those sessions, organizing chair Camilo Miguel Signorelli launched #20, The MoC transdisciplinary consensus project. Unlike conventional conference panels that end after ninety minutes, the consensus project ran as a five-day structured working track across classrooms A101 to A112 every afternoon from Monday through Thursday, culminating in the Friday 15:00 closing plenary discussion. By forcing theorists to separate mathematical definitions from empirical bridge laws in small working groups, the track sought to map exactly which axioms are shared across information-theoretic, predictive, and workspace models.

The Six Dedicated AI and Substrate Tracks at MoC7

Artificial consciousness and substrate independence occupied six dedicated parallel sessions across the five days, making MoC7 the densest technical gathering on formal machine consciousness criteria in 2026.

1. Language Models: Agency, Interpretability and Risk (Monday, Auditorium 3, talks #9–#12). Leonard Dung (#9, Why AI Welfare Risk is Robust) argued that moral risk under uncertainty holds across a wide range of functionalist priors, while Francis Rhys Ward (#10, Goal-Directed Agency in Language Model Agents) and Hoagy Cunningham (#11, Looking Inside Models for Evidence of Welfare-Relevant States) connected mechanistic interpretability directly to welfare assessment.
2. The Prospects of Machine Consciousness (Tuesday, Auditorium 3, talks #26–#28). GiulioStar (#26) audited computational functionalism under structural constraints, and Błażej Skrzypulec (#28) analyzed what formal criteria separate simulated behavioral reports from genuine internal state organization.
3. Confidence and Self-Monitoring in Neural and Artificial Systems (Wednesday, Auditorium 3, talks #57–#60). Building on Showcase #102 by Medha Shekhar on metacognitive dissociations for valenced awareness, this track compared biological prefrontal confidence signals with self-monitoring calibration in deep neural networks.
4. Mind The Substrate: On Substrate Dependence and Organisational Invariance (Thursday, Auditorium 2, talks #72–#74). Speakers dissected the mortal computation and organizational invariance debates, asking when low-level physical dynamics leak through a computational abstraction boundary.
5. Indicators of AI Consciousness (Thursday, Auditorium 5, talks #80–#83). Extending the lineage of Patrick Butlin, Robert Long, and colleagues’ indicator-property rubric covered in our flagship survey of AI consciousness definitions, contributors tested how architectural indicators behave when applied to transformer agents and recurrent world models.
6. Enactive and Developmental Routes to Consciousness (Friday, Auditorium 4, talks #111–#114). Echoing Evan Thompson’s Wednesday evening Mind-Matter Prize lecture at the Louisiana Museum (#70, Consciousness in Life), this closing track argued that genuine agency requires developmental sensorimotor coupling and autopoietic self-maintenance rather than static pretraining.

Wednesday morning’s AMCS Presidential Address by Lenore Blum (#51, World Models and the Emergence of a “First-Person” Perspective in an AI) and Friday morning’s keynote by Guillaume Dumas (#101, Formalizing Generative Neurophenomenology: Computational Foundations of Self and Others) provided the constructive synthesis across those tracks. In Blum’s Conscious Turing Machine (CTM) lineage, a first-person perspective emerges when an architecture builds a multimodal internal world model that includes a self-node tracking its own Broadcast actions. Dumas extended that formalization to multi-agent social coupling, modeling how two generative agents co-construct representations of self and other across a shared interaction boundary.

Comparison to The Consciousness AI

The debate running through MoC7’s Wednesday and Friday sessions, pairing Lenore Blum’s internal world model (#51), Jakob Hohwy’s self-evidencing (#34), and Evan Thompson’s insistence on embodied life-regulation (#70), maps directly onto the verification milestones we track in The Consciousness AI repository.

A purely disembodied world model can predict next tokens without having any internal stake in what happens next, and a policy network can appear to route information through a workspace layer that actually does no causal work. Earlier this month, our own architectural audit (tcai-research #57 and #63) uncovered exactly that failure mode in a baseline policy configuration, finding that the final thalamic_relay layer received zero gradient from any active loss term across 172 of 172 runs, leaving its output as fixed drift plus noise. Fixing that silent bypass required closing two loops that MoC7’s speakers emphasized from the theory side. First, the tectal motor map gate (tcai-research #58) verified that the embodied agent physically moves toward the target its tectal map selects (95 and 102 of 120 trials). Second, the value gate (tcai-research #62) verified across three random seeds that the agent learns from homeostatic reward which object actually feeds it (56/60 vs 28/60 with weights frozen).

MoC7’s mathematical consensus track and our engineering verification gates point to the same conclusion. Formal elegance on a whiteboard or a diagram is never enough; every claimed workspace broadcast, world-model self-node, or homeostatic drive must prove its causal necessity under ablation.

Sources

• Models of Consciousness 7 (MoC7) Official Schedule and Programme Information. Association for Mathematical Consciousness Science (AMCS), University of Copenhagen (HCØ and Niels Bohr Building), October 12–16, 2026. Checked October 17, 2026.
• Models of Consciousness 7 Conference Portal. Organized by Camilo Miguel Signorelli and the AMCS Board.
• The Consciousness AI Project Repository. Verification gate records #57, #58, #62, and #63.

Researchers covered here

• MP Megan PetersUniversity of California, IrvineMetacognitive access in AI, consciousness test validity, computational requirements for self-awareness 2 articles
• JO John O'KeefeSainsbury Wellcome Centre / Department of Cell and Developmental Biology, UCLNobel Prize in Physiology or Medicine (2014), Discovery of Hippocampal Place Cells (1971), The Hippocampus as a Cognitive Map (1978, with Lynn Nadel) 3 articles
• LM Liad MudrikSchool of Psychological Sciences / Sagol School of Neuroscience, Tel Aviv UniversityAdversarial collaboration methodology, high-level unconscious processing limits, neural basis of free will and conscious volition, Cogitate Consortium co-lead 4 articles
• ET Evan ThompsonDepartment of Philosophy, University of British ColumbiaThe Embodied Mind (1991, with Varela and Rosch), Mind in Life, Biology, Phenomenology, and the Sciences of Mind (2007), Waking, Dreaming, Being (2014), Enactivism 4 articles
• HA Harald AtmanspacherCollegium Helveticum / ETH ZurichDual-Aspect Monism (Pauli-Jung conjecture), Quantum Interaction, contextual emergence, Models of Consciousness (MOC) co-organizer 3 articles
• LR Lauren RossUniversity of California, IrvineScientific explanation, causation in neuroscience, metascientific evaluation of consciousness measures 2 articles
• GD Guillaume DumasUniversite de MontrealSocial neuro-AI, inter-brain synchronization during social interaction, computational psychiatry 2 articles
• MH Mirja Helena HartimoUniversity of HelsinkiHusserlian phenomenology, philosophy of mathematics, phenomenological approaches to consciousness 2 articles
• LB Lenore BlumDepartment of Computer Science, Carnegie Mellon UniversityConscious Turing Machine (CTM) formal architecture (with Manuel Blum), Blum-Shub-Smale (BSS) model of real computation, theoretical computer science 2 articles
• JK Johannes KleinerUniversity of Bamberg, Munich Center for Mathematical Philosophy, Association for Mathematical Consciousness ScienceMathematical consciousness science, the Kleiner-Hoel dilemma, Defining Biological Naturalism 2 articles
• CS Camilo Miguel SignorelliDepartment of Computer Science, University of Copenhagen. Department of Computer Science, University of OxfordThe Models of Consciousness transdisciplinary consensus project, compositionality and biological constraints on conscious computation, organizing chair of MoC7 Copenhagen 1 article
• JH Jakob HohwyMonash Centre for Consciousness and Contemplative Studies, Monash UniversityThe Predictive Mind (2013), predictive processing and self-evidencing accounts of conscious perception, adversarial testing of consciousness theories 2 articles
• PB Patrick ButlinEleos AI Research. Previously the Global Priorities Institute, OxfordLead author of the indicator properties framework for assessing consciousness in AI systems 24 articles