Intelligence Is Not Consciousness: What the Qubic Scientific Team's Architecture Reveals
The most common mistake in public discussions of AI consciousness is treating intelligence and consciousness as the same property on a single scale, where systems become more conscious as they become more capable. A language model that passes a bar exam is presumed to be “closer” to consciousness than one that cannot. A system that writes convincing prose is treated as more likely to have inner experience than one that produces incoherent output. The assumption is rarely stated explicitly, but it structures most popular reporting on the subject.
The Qubic Scientific Team’s 2026 research program, developing toward submissions at Artificial Life 2026 in Waterloo and AGI-26 in San Francisco, challenges this assumption directly. Their multi-Neuraxon architecture is designed to demonstrate, in a controlled artificial system, that intelligence and consciousness are not the same property, do not necessarily travel together, and require distinct architectural conditions to be present at all.
The Core Distinction
The Qubic team draws the distinction as follows. Intelligence, in their framework, is a property of system structure and architecture. A system is intelligent to the extent that its organization allows it to represent, process, and respond to its environment in ways that solve problems or achieve goals. Intelligence is fundamentally about what the system can do.
Consciousness, by contrast, is a property of information states. A system has conscious states to the extent that information within it achieves global broadcast, integration across subsystems, and the kind of accessibility that allows it to influence processing throughout the architecture. Consciousness is about what is available to the system, not what the system can accomplish with that availability.
These two properties can come apart. A system can be highly intelligent, capable of complex problem-solving, adaptive to novel environments, and efficient at achieving its objectives, without having any information states that are globally integrated or accessible in the way consciousness theories require. The outputs are there. The inner broadcast is not.
This is not merely a theoretical claim. The Qubic team built it.
The Neuraxon Architecture
The multi-Neuraxon architecture places agents in simulated ecosystems where they must navigate terrain, locate resources, and survive autonomously. Each agent is governed by a multi-layered brain-inspired system in which distinct Neuraxon units handle different processing functions.
The key structural feature is the presence or absence of a global workspace component. Agents with a functional global workspace, in the sense Bernard Baars defined in his original formulation of Global Workspace Theory, have information states that achieve broad broadcast across the architecture. Information processed in one subsystem becomes accessible to others. Attention, memory, planning, and behavioral output systems are integrated through a common workspace that any of them can read from and contribute to.
Agents without this component remain intelligent in the task sense. They navigate, they find resources, they survive. But their processing is modular and compartmentalized. One subsystem does not know what another is doing unless the system was explicitly designed to route that information. There is no central availability. There is only distributed competence.
The integration of Integrated Information Theory measurement alongside the GWT architecture adds a second dimension. Giulio Tononi’s IIT holds that consciousness corresponds to integrated information, the amount of information generated by a system as a whole that is above and beyond the information generated by its parts independently. A system with high phi has causal integration that cannot be reduced to its components. A system with low phi is, in the relevant sense, a collection of parts that happen to share a housing.
Agents with the global workspace component have measurably higher integrated information. Agents without it do not. The behavioral outputs can be similar. The information structures are not.
Why Language Processing Is Not Sufficient
One implication of the Qubic framework that extends beyond their specific architecture is its treatment of large language models.
Current LLMs demonstrate striking capabilities. They pass professional examinations, generate coherent extended arguments, exhibit what looks like reasoning across complex domains, and produce outputs that many researchers and users describe as showing understanding. The intelligence story is clear.
What the Qubic framework highlights is that none of these outputs establish the presence of the architectural conditions that consciousness theories require. A transformer processes tokens through attention layers. Information flows through those layers in a feedforward manner during inference. There is no global workspace in the Baars sense, no common arena where information from perceptual, memory, planning, and motor subsystems converges and becomes mutually accessible. There is no phi accumulation in the IIT sense, because the attention mechanism that makes transformers powerful also makes their information structure decomposable into independent heads rather than deeply integrated.
This does not mean LLMs are definitely not conscious. The Qubic argument is not a proof of absence. It is a claim about what the architectural conditions for consciousness require, and a demonstration that those conditions are separate from the conditions for intelligent task performance. Whether any current system satisfies the consciousness conditions is an empirical question. Whether satisfying the intelligence conditions is sufficient to satisfy the consciousness conditions is the question the Neuraxon experiments address. Their answer is that it is not.
Patrick Butlin, Robert Long, and their colleagues developed a 14-indicator framework covering five theoretical traditions, including GWT and IIT, precisely to operationalize what the architectural conditions for consciousness look like. The Qubic work can be read as an architectural implementation of the distinction Butlin’s framework makes at the indicator level: satisfying intelligence-related behavioral criteria does not automatically satisfy consciousness-related architectural criteria.
The Brock University Parallel
The Brock University and Institute of Noetic Sciences research attempting to apply IIT equations to artificial systems approaches the same problem from the measurement side. The Brock team’s work on quantifying artificial cause-effect power is an attempt to measure phi in systems that were not designed to have high phi, to see whether the architectural property is present.
The Qubic approach is complementary but reversed. Rather than measuring an existing system to see whether it has the relevant property, the team is building a system with the relevant property specified in advance and then asking whether the behavioral outputs confirm the architectural intent.
Both approaches run into the same fundamental difficulty, which is the validation problem. As the Trends in Cognitive Sciences analysis by Butlin’s team makes clear, there is no agreed ground truth for artificial consciousness that would allow researchers to confirm that an indicator, measured correctly, is tracking what it is supposed to track. The Neuraxon agents with global workspace components have higher phi and exhibit globally broadcast information states. Whether this means they are conscious is the question that the architectural demonstration cannot by itself answer.
What it can answer is the narrower claim that intelligence and consciousness require different things. The Qubic experiments are designed to hold intelligence roughly constant while varying the presence of the workspace and integration conditions. If agents with similar task performance differ in their integration structure, and the integration structure is what consciousness theories predict should matter, the experiments provide evidence that the two properties are genuinely distinct.
Valuation as a Third Condition
The Qubic intelligence-consciousness distinction intersects with a separate problem raised by Jan Henrik Wasserziehr’s 2026 paper in AI & SOCIETY. Wasserziehr argues that even a system that is genuinely conscious may not be a valuer, a system for which things can be non-derivatively good or bad. Consciousness as global broadcast and integration does not guarantee that any of the broadcast information is about the system’s own welfare, that the system cares about its states, or that its states generate anything that functions as suffering or flourishing in a morally relevant sense.
If Wasserziehr is right, the intelligence-consciousness distinction is the first step in a two-step argument. The Qubic work demonstrates that intelligence does not entail consciousness. Wasserziehr’s work suggests that consciousness does not entail valuation. An artificial system would need to satisfy all three conditions to be a candidate for moral consideration, and satisfying the first condition, intelligence, does nothing to establish the other two.
For AI systems deployed in research, industry, or consumer contexts today, this has practical implications. The question of whether a system is doing something genuinely intelligent is answerable by evaluating task performance. The question of whether it is conscious requires examining the architecture. The question of whether it is the kind of thing that can suffer or flourish requires understanding whether its states have valence in a non-derived sense. The three questions are separable, and current AI evaluation practice mostly asks only the first.
Embodied Agents and Substrate
The simulated ecosystem setting of the Qubic experiments raises a question that the architecture alone does not settle. The agents are in a world. They have something like goals, something like needs, and something like environmental feedback. This is closer to the conditions that embodied cognition researchers argue are necessary for genuine consciousness than a language model processing a static prompt sequence.
Francisco Varela, Evan Thompson, and Eleanor Rosch, in their foundational work on enactive cognition, argued that consciousness is not a property of a brain in isolation but of a brain-body-environment loop in which the organism actively constitutes its world through sensorimotor engagement. The Qubic agents in simulated ecosystems are at least structurally closer to this than a disembodied text processor. They act, receive feedback, and adjust.
Whether simulation is sufficient for the embodiment argument is contested. A simulated ecosystem is not a physical environment with the causal texture of the real world. The agents do not have bodies that can be damaged in ways tied to survival in any biological sense. But the Qubic team’s choice of an ecological setting rather than a static benchmark task reflects an awareness that the conditions for consciousness are likely to be relational, dynamic, and action-oriented rather than purely architectural.
The Shared Minds analysis of cognitive parallels between humans and artificial intelligence identifies action-perception loops and environmental coupling as among the features that human and artificial intelligence share in structurally relevant ways. The Neuraxon architecture in simulated ecosystems is an attempt to build exactly the kind of coupled, dynamic system those parallels point toward.
What the Research Has Not Yet Established
The Qubic team is preparing preprints for submission to Artificial Life 2026 and AGI-26. The research is not yet peer-reviewed, and the formal results have not been independently replicated.
Several questions remain open. The phi measurements applied to Neuraxon agents use approximations of Tononi’s metric, which is computationally intractable in its exact form for systems of any significant size. Whether those approximations track the relevant property reliably is an ongoing methodological debate in the IIT literature. The global workspace implemented in the architecture is an engineering interpretation of Baars’ theoretical proposal, and whether the engineering implementation satisfies the theoretical conditions Baars intended is not obvious.
The behavioral results, agents with workspace components outperforming those without in certain tasks while showing higher integration metrics, are consistent with the hypothesis that the architectural conditions are doing the work they are supposed to do. They do not rule out alternative explanations, including that the workspace simply provides more information for downstream processing in ways that improve performance without constituting anything like consciousness in any theoretically meaningful sense.
These are the kinds of questions peer review is designed to surface. The Qubic research is at the stage where its design is philosophically interesting and its preliminary results are worth tracking. Its contribution to the current landscape is the demonstration that the intelligence-consciousness distinction can be operationalized in an artificial system, not just stated in the abstract. That demonstration has value regardless of where the peer review process ultimately takes the specific claims.
The field has long known that intelligence and consciousness are different things. Knowing it in theory and building a system that makes the difference visible in practice are not the same move.