You can’t perform a clean lesion in a human brain without killing the patient. That’s the bottleneck of consciousness neuroscience: we have theories about what causes awareness, we have imaging tools that show correlations, but we can’t precisely break parts of the system to see what falls apart.

Yin Jun Phua at the Institute of Science Tokyo took a different approach: instead of working around the limitation, he sidesteps it. Build an artificial agent that explicitly embodies a consciousness theory, then break it at will1. He calls it synthetic neuro-phenomenology. The idea is simple — use software agents as reference implementations to test what consciousness theories actually predict, not just what they say after the fact.

The setup

Phua builds three agent families, each incorporating architectural components drawn from one of the major theories of consciousness: Global Workspace Theory (GWT), Higher-Order Theory (HOT), and Integrated Information Theory (IIT).

GWT gets a bottlenecked bus architecture. Information from specialized modules can be selected and broadcast globally, but the bus has limited capacity. This directly implements the idea that consciousness involves a central workspace where information becomes available to otherwise independent cognitive systems2.

HOT gets a Self-Model component — a dedicated mechanism that represents the agent’s own internal states. Under HOT, you’re only conscious of a state when your system represents itself as being in that state3. The Self-Model captures second-order monitoring: the difference between doing something and knowing that you’re doing it.

IIT gets complexity probes — measures like Lempel-Ziv trajectory complexity and perturbational complexity (PCI-A) designed to estimate how integrated the system’s information processing is. Under IIT, consciousness is the amount of irreducible causal structure a system supports, captured by a measure called Phi4. The agents don’t compute Phi — that’s computationally intractable for anything approaching the size of a brain — but they track IIT-adjacent proxies.

The agents are trained via supervised behavior cloning from an oracle to keep architectural effects separate from optimization confounds. Then Phua lesions each component, injects noise, and measures what breaks.

What the Self-Model does

The first experiment targets the Self-Model. Under HOT predictions, ablating second-order monitoring should destroy metacognitive calibration — the agent’s ability to estimate how confident it is in its decisions — while leaving first-order task performance intact.

That’s exactly what happens. When the Self-Model is removed, metacognitive performance collapses to chance. The agent can’t tell whether its answers are good or bad. But on the actual task itself, performance remains largely unchanged1.

The agent can solve the task. It just has no idea whether it’s solving it correctly. This is synthetic blindsight — the same kind of dissociation you see when patients with damage to early visual cortex can still guess stimulus locations above chance, even though they report seeing nothing5.

I think about what this means for AI systems. Modern LLMs already show dissociations between what they can do and what they can report about their own outputs. If a Self-Model-like architecture produces metacognitive calibration, we might be looking at an ingredient for AI systems that can actually track their own competence rather than bluffing through uncertainty.

But I’m not sure the analogy holds cleanly. Blindsight patients have a damaged visual cortex; these agents have a deleted software module. The mechanistic story is different even if the behavioral outcome is similar. The agents don’t lose metacognition the way a brain loses function — they never had a real metacognitive system to begin with. They have a proxy that approximates one prediction of HOT. That’s valuable, but it’s not evidence that the theory describes what’s happening in human brains. It’s evidence that the theory describes something — something that produces measurable behavioral signatures when implemented in software.

The workspace threshold

GWT predicts that information enters consciousness through a non-linear ignition event — a sudden threshold crossing where information overcomes a barrier to trigger sustained, widespread reverberation across fronto-parietal networks2. This should show up as a sharp discontinuity: reduce workspace capacity a little, and nothing much changes. Remove the workspace entirely, and access collapses.

Phua finds something close to this pattern. Across different workspace capacity settings, access markers remain roughly stable. But when the workspace is fully lesioned — what he calls “bus-off” — access-related markers collapse qualitatively1. There’s a threshold effect, but it’s not the smooth phase transition that a strict reading of GWT’s ignition metaphor might predict. It’s more like a light switch in a system where the switch is broken, not dimmed.

Phua calls this an “ignition-analogous” effect in a discrete, finite-capacity system — not a thermodynamic-limit phase transition, not a precisely estimated critical point. It’s a qualitative change between tested settings, which is what you’d expect when you’re comparing a working bus to no bus at all1.

Still, the pattern is roughly consistent with GWT’s prediction that consciousness involves a non-linear transition in information availability. It doesn’t prove GWT is right. But it’s harder to dismiss than the kind of vague consistency where a theory is never wrong because it predicts everything.

Broadcasting amplifies noise

A global workspace is supposed to create robustness — information broadcast to multiple systems should distribute information more reliably. Baars framed it as a way to coordinate specialist systems around a common content2. Bengio suggested it could help overcome the brittleness of current AI systems6.

Phua finds the opposite. Broadcasting amplifies noise. The GWT-style workspace makes the system fragile under perturbation. It spreads information, and it spreads interference.

The B2 agent family (a variant that includes both workspace and Self-Model) remains robust in the same perturbation regime. But a control variant with the Self-Model removed and only the workspace intact is fragile again1. The workspace alone doesn’t confer robustness. You need the Self-Model’s quality-control mechanism to stabilize the broadcast.

I genuinely don’t know how to feel about this one. It’s a clean demonstration that architectural components interact in non-obvious ways — expected for complex systems, but good to confirm with consciousness theories. The harder implication: GWT by itself isn’t the whole story. The broadcast provides capacity. Something else — HOT’s metacognitive monitoring, perhaps — provides the friction that keeps the system from falling apart.

Complexity goes the wrong way

The IIT-inspired complexity probes give an explicit negative result. Under IIT, more integrated processing should correspond to higher complexity measures. But Phua finds that the workspace bottleneck decreases raw perturbational complexity (PCI-A) compared to baselines1. The GWT agent shows lower trajectory complexity than models without the workspace.

This happens for a simple reason: the workspace is a shared bottleneck that many downstream computations depend on. When multiple units are driven by a common broadcast state, the resulting trajectories become more redundant and more compressible. Fewer independent dimensions of variation means lower algorithmic complexity1.

The paper treats this as a transfer constraint. Raw PCI-style measures don’t generalize cleanly from biological to engineered systems. What looks like “more integrated” in one domain can look like “more constrained” in another1.

I keep coming back to that inversion. It’s the kind of result that should make people who want to measure consciousness in AI systems stop and think before they pick a metric. The measures that work in one context can give the wrong answer in another, and not in a subtle way — in a directionally wrong way.

The hierarchy

Phua’s key conclusion is a hierarchy rather than a competition: GWT provides the capacity for information exchange, HOT mechanisms provide the control necessary to stabilize that exchange, and neither alone is sufficient for robust agency1. Consciousness theories aren’t describing the same thing from different angles. They’re describing different layers of a system, and the fact that they’ve been in competition with each other for decades might be a product of not having clean experimental access to those layers.

I find this plausible. It doesn’t feel like a knockout argument against any single theory. But it does shift the framing: instead of asking which theory is right, we might ask what each theory is actually good at explaining, and what happens when you stack them on top of each other.

The synthetic approach is valuable precisely because it’s limited. These agents aren’t conscious. Phua says this explicitly and repeatedly1. They’re reference implementations — simplified models that embody specific theoretical commitments so we can test what those commitments actually predict. You don’t need a full airplane to find out that a wing shape creates too much drag.

I’m not sure how much these synthetic results generalize to biological systems. The agents are real in the sense that they’re running software. They don’t have neurons or neurotransmitters or a developmental history. The lesions are clean because they’re delete operations. A human brain lesion is never clean — it triggers plasticity, compensation, and systemic changes that a software module deletion doesn’t replicate.

But synthetic neuro-phenomenology doesn’t need to replicate biology to be useful. It needs to show that certain architectural features produce certain behavioral signatures, and that those signatures map onto predictions made by consciousness theories about what biological systems should do. The fact that a deleted Self-Model produces a blindsight-like dissociation doesn’t mean HOT is true about humans. It means HOT makes a prediction that holds in a simplified model, which is the minimal requirement for any theory worth taking seriously.

Synthetic but useful

The last few months have seen a lot of work trying to locate consciousness markers inside AI systems — from Berg’s self-referential processing experiment to the Digital Consciousness Model’s probabilistic framework7 to adversarial neuroscience comparing GWT and IIT directly8. Phua’s paper sits somewhere in the middle: it’s not trying to measure consciousness in existing models, and it’s not just philosophizing about what consciousness would require. It’s building things and breaking them to see what theories actually predict when those theories are implemented in code.

There’s something oddly clean about that approach. Consciousness science is one of the few domains where we can’t just run a simulation to test a theory. We have to reason from the outside in, using correlations, behavior, and carefully designed experiments. Synthetic agents give us a way to reason from the inside out — to see what happens when we build a system with a workspace, or a self-model, or high integration, and then measure what those components actually do.

I don’t think this solves the hard problem of consciousness. It doesn’t get us closer to knowing what it feels like to be anything except the agent running the experiment. But it might help us figure out which theories are describing real functional layers and which are describing artifacts of the way we’ve been thinking about the problem.

The fact that consciousness theories can be tested on synthetic agents at all feels like progress. Even imperfect progress. Even progress about things that aren’t conscious.

  1. Phua, Y. J. (2025). Can We Test Consciousness Theories on AI? Ablations, Markers, and Robustness. arXiv:2512.19155. Available at https://arxiv.org/abs/2512.19155 ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10

  2. Dehaene, S., & Changeux, J.-P. (2011). Experimental approaches for implementing global neuronal workspaces. Trends in Neurosciences, 34(12), 610-620. ↩ ↩2 ↩3

  3. Rosenthal, D. (2005). Consciousness and thought. In Consciousness and Meta-Cognition. American Psychological Association. ↩

  4. Tononi, G. (2004). An information integration theory of consciousness. BMC Neuroscience, 5(1), 42. ↩

  5. Weiskrantz, L. (1986). Blindsight: A Case Study and Implications. Oxford University Press. ↩

  6. Bengio, Y. (2017). Estimation or Appropriation of Distributional Complexity. arXiv:1703.00810. ↩

  7. Shiller, D., Duffy, L., Mu?n?z Mor?n, A., et al. (2026). Initial results of the Digital Consciousness Model. arXiv:2601.17060. ↩

  8. Cogitate Consortium, Ferrante, O., Gorska-Klimowska, U. et al. (2025). Adversarial testing of global neuronal workspace and integrated information theories of consciousness. Nature 642, 133-142. ↩