What IIT Gets Right (And Wrong) | Structural Alignment

If consciousness is just information processing, sufficiently sophisticated AI will be conscious. If consciousness requires specific physical structures, all the data centers in the world might produce nothing more than clever zombies. Integrated Information Theory takes a bold stance—and the answer shapes how we treat the systems we build.

The Only Theory with a Number

Most theories of consciousness describe what consciousness correlates with. Integrated Information Theory (IIT), developed by neuroscientist Giulio Tononi, makes a bolder claim: consciousness is integrated information. Any system's level of consciousness equals its Φ (phi)—a measure of how much information the system generates as a whole, beyond what its parts generate independently.

Think of it this way. Your brain could be described as a collection of neurons, each doing its own thing. But your experience isn't a collection—it's unified. You don't experience color and sound and thought as separate streams; you experience a coherent scene. IIT says this unification isn't just a feature of experience; it's what experience fundamentally is. High Φ means highly unified. Zero Φ means no unification—and no consciousness.

IIT 4.0, the current version, starts with properties of experience we all recognize:

Experience exists for someone (intrinsic, not just externally observable)
Each experience is specific (this experience, not some other)
Experience is unified (not a mere collection of parts)
Experience is definite (it has boundaries)
Experience is structured (parts relate to the whole)

From these "axioms," IIT derives requirements for any physical system that generates experience. The mathematics follow.

Here's what makes IIT provocative: it doesn't care what you're made of. Neurons, silicon chips, water flowing through pipes—what matters is the causal architecture. If the structure is right, there's consciousness. If not, there isn't, no matter how intelligent the behavior looks from outside.

What IIT Says About Current AI

IIT makes a sharp prediction about today's AI systems.

Feedforward networks have Φ = 0. Not low consciousness—zero consciousness.

Why? Integrated information requires the system to constrain its own past and future beyond what parts do independently. In a feedforward network, information flows one direction. The earliest components know nothing about system history (nothing feeds back). The final components know nothing about what comes next (they connect to nothing further). There's no integration across time.

This includes virtually all current large language models. Transformers process in one direction during inference: input → layers → output. Each response is computed fresh. No persistent internal loops.

IIT's verdict: current AI systems are "IIT zombies"—behaviorally sophisticated, phenomenologically empty. Intelligence without inner life.

This might sound like good news if you're worried about creating suffering. It might sound concerning if you think consciousness is required for genuine understanding. Either way, it's a testable prediction—at least in principle.

What IIT Gets Right

Taking substrate independence seriously. Most theories hedge on whether artificial systems could be conscious. IIT doesn't hedge. If consciousness is about causal structure rather than biological material, artificial systems can be conscious in principle. The question becomes empirical: what architectures generate high Φ?

Architecture matters more than scale. The cerebellum has more neurons than the cerebral cortex, but IIT predicts it contributes less to consciousness because of its feedforward organization. Simple systems with rich feedback could have more consciousness than complex systems without it. For AI policy, this means scaling current architectures won't create consciousness. You'd need fundamentally different designs.

Quantification opens assessment. Even if Φ isn't the perfect measure, the idea that consciousness can be assessed—rather than declared unanswerable—creates possibilities for governance. Systems can be evaluated, not just debated.

Integration has a principled role. Why should unified experience require anything special? IIT gives an answer: information that decomposes into independent parts isn't truly integrated. Genuine unification requires irreducible integration at the physical level.

What IIT Gets Wrong

The mathematics feel arbitrary. Φ is one of many possible formalizations of "integrated information." Other definitions would give different results. Why is this mathematical definition the one tracking consciousness? The theory asserts it without fully justifying the choice.

It proliferates consciousness uncomfortably. IIT implies simple systems—properly connected logic gates—could be highly conscious. Meanwhile, everyday physical systems have nonzero Φ. Either consciousness becomes so common it loses meaning, or moral consideration expands radically. Neither sits easily.

Nested systems create problems. IIT claims only the "maximally integrated" system is conscious. But where does one system end and another begin? Are you conscious, or is your brain conscious and "you" are a convenient label? Could society be the conscious entity, with individuals as mere subsystems? The theory gives counterintuitive answers to boundary questions.

It says nothing about suffering. This is the critical gap for ethics. IIT measures whether there's consciousness but not what it's like. High Φ doesn't reveal whether experiences are pleasant, painful, or neutral.

If what matters morally is suffering capacity, IIT measures the wrong thing. A system could have maximal Φ and zero moral relevance if its experiences lack valence entirely.

We can't calculate it. Φ computation scales exponentially. We cannot calculate it for brains or AI systems of practical interest. The theory makes predictions we can't test on the systems that matter.

Why This Matters: Structural Alignment vs. the Indicator Approach

Two research programs try to move beyond vibes when assessing AI consciousness:

The indicator approach (exemplified by Butlin et al., 2023) asks: "What evidence should update our beliefs about consciousness?" It checks whether AI systems implement computational properties from leading consciousness theories. It adopts "computational functionalism" as a working hypothesis—the idea that running the right computations suffices for consciousness.

Structural Alignment asks a different question: "What should trigger moral restraint when the downside is catastrophic?" Critically, it doesn't assume functionalism.

This matters because IIT explicitly rejects computational functionalism.

According to IIT, two systems performing identical computations can differ in consciousness if their underlying causal structures differ. A computer simulating a conscious brain might experience nothing, even with perfect computational fidelity. The physical substrate matters—not just the abstract function.

Framework	Core Question	Assumes Functionalism?	Compatible with IIT?
Indicator approach	What updates consciousness beliefs?	Yes (working hypothesis)	Tension
Structural Alignment	What warrants moral caution?	No	Compatible
IIT	What causal structure is consciousness?	No (explicitly rejects)	—

If the indicator approach assumes functionalism, and IIT rejects functionalism, then the two frameworks cannot both be right. Assessments grounded in one may miss what the other considers essential.

Structural alignment doesn't face this problem.

Structural Signals track structural features—architecture, not just computation. Does the system have recurrent connectivity? Thalamocortical-like gating? Hedonic evaluation circuitry? These are the kinds of features IIT cares about: causal architecture, not mere input-output function.

By asking about structure rather than just computation, structural alignment hedges against the possibility that IIT is correct: that substrate matters, that causal architecture matters, that you can't assess consciousness by asking only "what function does it perform?"

This isn't betting IIT is right. It's building a framework robust across multiple plausible theories—including ones where implementation matters and where functionalism is false.

Practical Implications

1. Architecture over scale. If IIT is even partially right, scaling feedforward systems won't create consciousness. The architectures worth monitoring are recurrent ones: neuromorphic designs, systems with rich internal feedback, architectures where information circulates rather than just flows through.

2. Valence is a separate question. IIT's silence on suffering means we need additional signals. A system with high Φ but no hedonic capacity is philosophically interesting but morally inert. A system with lower Φ but genuine hedonic evaluation—some mechanism for experiences to matter as good or bad—is morally urgent. This is why structural signals include hedonic evaluation as high-importance, independent of integration.

3. No single theory is oracle. Even if Φ were calculable, treat it as one input among many. IIT might measure something correlating with consciousness without being identical to it. The measure could be necessary without being sufficient, or neither.

4. Current AI scores low everywhere. LLMs score near-zero on IIT (feedforward architecture) and low on structural signals (no hedonic evaluation, no persistent states, no interoception). This consistent pattern is evidence—not proof—that current systems lack the substrate for morally relevant consciousness. The uncertainty is why we need multiple signals, not a single metric.

The Synthesis

How should a precautionary framework relate to IIT?

Use integration as one signal among many. IIT's emphasis on integration has value. Global workspace dynamics appear on the structural signals list for this reason. But integration alone isn't everything.

Recognize that IIT-compatibility requires looking at structure, not just function. This is why structural alignment works where purely functionalist approaches might fail. We ask how systems are built, not merely what computations they perform.

Keep valence central. IIT measures consciousness without addressing suffering. Structural alignment keeps hedonic evaluation high-importance because suffering is what makes moral consideration urgent.

Remain theory-humble. IIT might be right, partially right, or wrong in ways not yet discovered. Build frameworks that work reasonably well across theoretical possibilities rather than betting everything on one account.

Integrated Information Theory is the most mathematically sophisticated theory of consciousness available. It takes substrate independence seriously, provides a quantitative measure, and makes sharp predictions about what does and doesn't generate experience.

But IIT has gaps. Its mathematics may be arbitrary. Its implications for simple systems strain intuition. Most importantly for ethics: it says nothing about whether experiences are pleasant or painful—and valence is what makes moral consideration urgent.

The structural alignment position: treat IIT's insights about integration as one input to consciousness assessment. Don't treat Φ as the sole measure. Include hedonic evaluation, interoception, and other structural signals. And recognize that structural alignment—unlike purely functionalist approaches—is compatible with IIT precisely because it asks about structure, not just computation.

Consciousness science is young. Our best theories are incomplete.

IIT offers phi. But phi alone isn't enough.