Integrated Information Theory and Machines: What Measuring Phi Can and Cannot Tell Us

What integrated information theory actually claims

Integrated information theory, developed by Giulio Tononi and collaborators from the early 2000s onward, proposes that consciousness is identical to integrated information: the degree to which a system's causal structure generates information as a unified whole, over and above its parts considered separately. The theory's central quantity, phi, measures this irreducibility: a system has high phi when its parts constrain each other so richly that partitioning it destroys information, and zero phi when it decomposes cleanly into independent modules. IIT is unusual among consciousness theories in starting from phenomenology, five axioms about what experience is like, and deriving postulates about the physical substrate, formalized most fully in IIT 3.0 and its successors. Two consequences matter for AI discussions. First, IIT ties consciousness to intrinsic causal structure, not behavior: a system could act intelligent with near-zero phi. Second, phi as defined is computationally explosive, requiring evaluation over all partitions of a system's state space, which makes exact computation infeasible beyond toy systems, a fact that shapes everything downstream, including every claim anyone makes about measuring phi in machines.

The controversy, stated fairly

IIT's scientific standing is genuinely contested, and honesty requires saying so. In 2023, over one hundred researchers signed an open letter arguing IIT should be regarded as pseudoscience, on the grounds that its core claims are untestable: phi for any interesting system cannot be computed, and the theory's predictions about which systems are conscious, including counterintuitive implications that some simple grid-like systems could have high phi while sophisticated feed-forward AI has none, resist empirical falsification. Defenders responded that IIT makes testable neuroscience predictions, some borne out in studies of anesthesia, sleep, and the perturbational complexity index used clinically to assess consciousness in unresponsive patients, and that ambitious theories deserve time. Adversarial collaborations testing IIT against rival theories have produced mixed, nuanced results. A careful reader should hold three positions simultaneously: IIT is a serious mathematical framework with real empirical footprint in neuroscience; its strong identity claim, that phi is consciousness, is far from established; and any application to machines inherits both the framework's insight and its unresolved status. That inheritance defines what a machine phi metric can honestly claim, which is the subject of the rest of this page.

Why computing phi for real AI systems is infeasible

Exact phi computation requires specifying a system's full transition probability matrix and searching over all possible partitions for the one that least reduces integrated information, a computation whose cost grows super-exponentially with system size. Researchers have computed exact phi for systems of a handful of logic gates; a transformer with billions of parameters and high-dimensional continuous states is beyond any conceivable exact treatment. This is not an engineering gap that better hardware closes; it is combinatorial explosion. Consequently, every practical approach uses proxies: measures inspired by IIT that capture some aspect of integration and differentiation while remaining computable. Neuroscience did exactly this with the perturbational complexity index, which perturbs the brain and compresses the response, and with various approximate phi measures over coarse-grained dynamics. For neural networks, candidate proxies include entropy and mutual-information structure across components, the degree to which information flow resists modular decomposition, and cross-layer causal influence. The intellectually honest framing, which QBC adopts explicitly, is that such metrics are IIT-inspired measurements of integration in a computational system, informative and reproducible, but categorically not phi in the strict sense, and never evidence of consciousness.

QBC's Phi: an IIT-inspired metric, measured and anchored

QBC computes a metric named Phi from the real forward passes of its served model, the 7B system whose manifest is attested on-chain. Three families of signal are combined: Shannon entropy of attention distributions, capturing how broadly the model distributes its processing rather than collapsing onto single tokens; cross-expert routing integration across the model's 10-expert Sephirot MoE adapter, capturing whether the experts function as an interacting ensemble or independent silos; and cross-layer information flow, capturing integration in depth. The metric is computed from live inference, published continuously on the aether-net dashboard, and anchored on-chain, so its history is tamper-evident and its methodology reproducible against open weights: anyone can download the model, replicate the computation, and check the published values. QBC's position on interpretation is deliberate and explicit: Phi is a measured property of information integration in a running system, useful for observing how the model's internal dynamics change across training epochs and workloads. QBC does not claim the model is conscious, and treats Phi as evidence of integration, not experience.

Why measure integration at all, and what would change our minds

If Phi is not a consciousness detector, why compute it? Three reasons survive the deflation. First, interpretability: integration metrics summarize global properties of a model's computation that per-neuron analyses miss, and tracking them across training reveals structural change, complementing the quorum-measured loss series with a view into how, not just whether, the model changes. Second, scientific hygiene by example: consciousness claims about AI are proliferating with zero measurement discipline, and a public, reproducible, on-chain-anchored metric with explicitly limited interpretation demonstrates what honest practice looks like, the same verification-first philosophy QBC applies to training and inference. Third, optionality: if the science of consciousness matures toward validated computable markers, systems already publishing longitudinal integration data will be the ones able to engage that science retroactively. What would change the interpretation? Independent validation tying specific proxy measures to consciousness-relevant properties, adversarial testing of IIT-derived predictions in artificial systems, and theoretical progress making tractable approximations principled rather than heuristic. Until then, the honest sentence is short: QBC measures integration, publishes the measurement, anchors it immutably, and lets the interpretation remain exactly as strong as the science allows.

IIT among rival theories, and the tests that matter

IIT does not sit alone; placing it among rivals clarifies what any machine metric can borrow from it. Global workspace theory ties consciousness to information broadcast across a system-wide workspace; higher-order theories tie it to representations about the system's own states; recurrent processing theories emphasize feedback loops in perception. These families make different predictions about which architectures could support experience, and none of them licenses a consciousness verdict about current language models either. The empirically strongest export from IIT so far is clinical rather than metaphysical: the perturbational complexity index, validated by Casali and colleagues, distinguishes conscious from unconscious patients by perturbing the brain and compressing the response, a genuinely useful diagnostic that does not require accepting IIT's identity claim. Adversarial collaborations testing IIT against global workspace predictions have returned mixed results, sharpening rather than settling the dispute, and the 2023 pseudoscience letter remains the clearest statement of the skeptical position. For machine metrics the lesson is discipline: borrow the measurable structure, publish the method, and refuse the metaphysical conclusion, the stance QBC's safety-conscious approach to model claims makes explicit.

What the on-chain Phi record contains, concretely

Skeptical readers should be able to inspect the object being discussed, so here is what QBC actually publishes. Phi values are computed from live forward passes of the served model and anchored on-chain, forming a longitudinal series bound to model identity: each measurement is attributable to the exact manifest that produced it, because serving is continuously bound to the attested artifact through per-block Proof-of-Thought commitments. The chain's deterministic finality makes the series tamper-evident in the same way as any audit trail: history cannot be quietly rewritten to fit a narrative, which is precisely the failure mode that plagues consciousness discourse around AI. The longitudinal record and its interpretation limits are described in the soul ledger essay, and the unification that put measurement, training, and serving on one in-process model, so Phi tracks the model users actually query rather than a lab copy, is documented in the v7.1 unification post. Anyone can download the open weights, replicate the entropy, routing, and information-flow computations, and compare against the anchored series. Disagreement about what Phi means is welcome; disagreement about what was measured is designed out.