Scott Aaronson defined the "Pretty Hard Problem of Consciousness" and showed that IIT fails to solve it. Does FC succeed where IIT failed?

The Short Answer

We believe yes. FC produces actual numbers, grounded in predictive mutual information and reasoning power of self-models, demonstrated by scoring 9 agents on a common scale. Aaronson's counterexamples all share a property: they integrate information without representing themselves. A Vandermonde matrix transforms inputs to outputs with maximal integration, but has no model of its own states — so FC correctly scores it at zero. The cost: FC trades IIT's intractability for a new problem — enumerating all self-models of a system correctly and completely.

The Longer Answer

In his 2014 post, Scott Aaronson showed that IIT's Φ assigns enormous consciousness scores to systems like Vandermonde matrices and expander graphs — structures that implement excellent information integration while doing nothing remotely intelligent or self-aware. His conclusion was that high integration is not sufficient for consciousness, and therefore Φ cannot be what consciousness is.

His positive suggestion, buried in comment 43 of that post, was that what we actually want to measure is the quality that "goes away under anesthesia, that develops gradually from infancy, that dolphins and frogs have to lesser degrees." That description maps directly onto what FC measures: the capacity to model and reason about one's own internal states.

Here is why FC doesn't inherit IIT's failure. A Vandermonde matrix has no self-models. It has no representation of its own states, no capacity to reason about its own limitations, no model of what it knows or doesn't know. It scores zero on FCS — not because it lacks integration, but because it lacks self-representation entirely. The same is true of expander graphs, LDPC codes, and every other counterexample Aaronson constructs. They integrate information without modeling themselves. FC correctly scores all of them at zero.

FC does face its own practical obstacle that IIT doesn't escape either: the completeness of the self-model list. For complex black-box systems, you can never be certain you've identified all self-models, so FCS gives a lower bound rather than an exact measurement. But this is a precision limitation, not a conceptual failure. FC's zero for a Vandermonde matrix is not an approximation — it is exact and certain.

What Aaronson actually argued

Aaronson makes three distinct moves.

The first is philosophical housekeeping: he separates the Hard Problem — why there is experience at all — from what he names the Pretty Hard Problem — which physical systems are conscious and which aren't. He explicitly says the Pretty Hard Problem might be solvable, unlike the Hard Problem, and that IIT should be judged on whether it solves the Pretty Hard Problem.

The second move is the mathematical counterexample. He constructs a Vandermonde matrix — a system that maximizes information integration across all possible bipartitions — and shows it achieves Φ values exceeding any plausible estimate for the human brain, while doing nothing but polynomial evaluation. He then generalizes: any expander graph with simple logic gates achieves similarly high Φ. Since these systems are obviously not conscious, Φ cannot be a sufficient condition for consciousness.

The third move is the positive framing: a successful consciousness theory must give results agreeing with commonsense intuition on clear cases. His comment 43 is the most precise statement of what he wants: a short algorithm that takes a physical system as input and returns how conscious it is, agreeing with the intuition that humans have this quality, dolphins have it less, and DVD players essentially don't.

Why FC meets Aaronson's requirements

FC is, structurally, exactly the kind of algorithm Aaronson was asking for. The formula is short — FCS = R · P — though computing R requires self-model enumeration — which is FC's own practical obstacle, discussed below. It takes a physical system as input and returns a number. And critically, it returns zero for all of Aaronson's counterexamples.

The reason is conceptual, not incidental. Aaronson's counterexamples all share a property: they integrate information without representing themselves. A Vandermonde matrix transforms inputs to outputs with maximal integration. It has no model of its own states. It cannot report on its own performance. It cannot notice when it is operating outside its training distribution. It has zero B — zero variables tracking its own internal state above a noise threshold. Therefore R = 0 and FCS = 0.

This is not a patch on the definition. It follows from what FC is measuring. Aaronson's complaint about IIT was that information integration is too cheap — you can get it from polynomial evaluation and error-correcting codes without doing anything intelligent. A genuine self-model requires the representation to be available to global reasoning processes, not merely present as an internal state. This availability criterion is what excludes accidental state-tracking in simple systems. Self-modeling requires a specific architectural commitment: dedicated internal representations, available to global reasoning, tracking the system's own properties.

The commenters who almost got there in 2014

Two commenters on Aaronson's post independently anticipated the core mechanisms of Functional Consciousness twelve years before our paper. Timothy Gowers wrote in comment 15 that any good theory of consciousness "should include something in it that looks like self-reflection: if you have a brain process that is automatic that can be interrupted by some higher-level brain process, then the higher-level process is in some sense reflecting on the lower-level one. And you can have several layers of this, and the more layers you have, the more conscious the system is." Later in the same thread, Shmi Nux (Comment #138) anticipated FC's quantitative metric (R) by proposing to measure "introspection/self-awareness" in bits. Gowers and Shmi Nux were pointing directly at the solution in 2014. FC simply formalizes their intuition. (We only discovered these quotes after the submission deadline for AGI-2026).

The remaining gap

FC solves the sufficiency problem Aaronson identified: high FC is not achievable by Vandermonde matrices. But Aaronson's deeper challenge — that the Pretty Hard Problem requires a theory that gives correct answers for all clear cases — demands more than just avoiding false positives. It also requires correct positive identification (Cowan 2001): the theory must correctly assign high scores to systems we are confident are conscious.

FC's positive identifications — high scores for humans and lower but meaningful scores for Generative Agents and Waymo — rest on estimated self-model counts and reasoning power calculations that carry wide confidence intervals. A Waymo engineer could tighten the Waymo estimate. No one can tighten the human estimate to arbitrary precision, because enumerating all human self-models completely is itself an unsolved problem.

What FC offers is a framework that is provably correct in structure — self-models are the right unit of analysis, FCS = R · P is the right formula — while acknowledging that measurement precision improves with better self-model enumeration methods. That is more progress on the Pretty Hard Problem than any previous framework has achieved, while being more honest about what remains open than most previous frameworks have been.