Recently I asked a state-of-the-art AI model to explain the Four-Model Theory of Consciousness and compare it to the leading theories in the field. It did a fluent job — confident, well-structured, and wrong in a very particular way. Not randomly wrong. It made the same cluster of misreadings I’ve seen from hurried human readers, almost point for point.

That was oddly reassuring. If the strongest pattern-matcher we have defaults to these errors, they aren’t careless mistakes — they’re the natural readings, the grooves the mind slides into. Which means they’re worth mapping. Here are the three biggest, plus a bonus, each with why the mistake is reasonable and what the theory actually says.

Misreading 1: “Four models” means four boxes in the brain

The theory is called Four-Model Theory, and it presents a 2×2 table. A 2×2 table looks like four cells, and four cells look like four modules — separate systems wired together, box-and-arrow style. Every cognitive-architecture diagram since the 1980s has trained us to read it that way.

But the four are not things. They are two distinctions crossed: what a model is about (the world, or the self) and how it runs (implicitly, baked into the machinery, or explicitly, as a running simulation). The brain doesn’t implement four models — it implements an effectively uncountable number of overlapping ones. The four named in the theory are the extremal corners of that continuous space, the minimum sufficient set. Four is the floor, not the ceiling.

And the two axes aren’t even symmetric. The self-models are nested inside the world-models — your explicit self is a region of your explicit world, not a neighbor of it. Four independent boxes gets it wrong twice.

Misreading 2: “Criticality” is the requirement

The theory says conscious systems operate near criticality — the edge-of-chaos regime where a system is neither frozen into order nor dissolved into noise. Readers reasonably conclude: criticality is the requirement. It’s the most measurable thing in the whole theory, so of course everyone reaches for it.

But criticality is the symptom, not the disease. The actual requirement is what I call free compute: the capacity for open-ended, universal computation, actually deployed — turned inward, running unconstrained models of the system itself. Criticality is what that condition looks like in certain substrates, the observable signature some systems show while doing it. Measuring the signature is how you test the theory; it isn’t what the theory is about.

The distinction has teeth. Your laptop is a universal computer and is not conscious — nothing in its situation forces that capability into autonomous self-modeling. Capability, freely instantiated, evolutionarily driven: all three, not just a tuning parameter sitting at its critical value.

Misreading 3: The implicit “feeds data” to the explicit

The theory distinguishes implicit models (stored in the structure — connectivity, weights) from explicit ones (the running simulation, the inner movie). The obvious picture: data flows from the implicit side to the explicit side, across some boundary, through some pipe. My AI interlocutor immediately demanded the mathematics of that transfer.

There is no pipe. The explicit simulation is made from the substrate the way running software is constituted by hardware. When your laptop runs a program, nothing is being shipped from the hardware into the software — the software is what the hardware is doing. Same relation here: the implicit state is the machine, the explicit state is the computation it’s executing. Asking for the transfer step is asking to formalize something that doesn’t exist.

(There is a real, separate question about which stored content surfaces into the running simulation at any moment — but that’s a question about the simulation’s contents, not about how the simulation comes to exist.)

Bonus misreading: “Sounds architectural — so build it”

Because the theory is unusually concrete about architecture, technically minded readers jump straight to: implement it in code. The AI model did too — its competence gradient runs toward writing software, so that’s where it drifted.

But FMT is a theory of what consciousness is — the architecture any substrate, biological or not, must satisfy to have it. That a conscious machine is possible in principle is a downstream implication, not the content of the theory, and the engineering doesn’t exist yet. An LLM agent wrapped in a memory loop mimics the shape of the architecture without the self-referential closure that does the actual work. Mimicry of the diagram is not instantiation of the thing.

None of these misreadings are foolish. Each follows the local slope of how we’ve learned to read theories: tables become modules, measurables become requirements, boundaries become channels, blueprints become build orders. The theory just happens to sit off those slopes — which is, I’d argue, part of what makes it worth the second look.

The full theory is openly available: the FMT paper on Zenodo (DOI 10.5281/zenodo.18669891), and fmt.matthiasgruber.com.