Home / Theories of Consciousness

Integrated Information Theory and AI: What Phi Means

A tangled knot of glowing tendrils beside a line of separate motes

Integrated information theory, or IIT, is the leading theory of consciousness that points away from conscious AI. It says consciousness is not about what a system computes but about how its physical parts cause and constrain each other, a quantity it calls phi. On that view an ordinary computer could reproduce everything a brain does and still have little or no experience, because its hardware is not integrated in the right way.

That makes IIT central to the 2026 debate. A framework published in September 2026 found that much of the disagreement about today's language models comes down to how much weight researchers give theories like this one. The theory has also had a hard test: a large brain study published in Nature in 2025 challenged one of its key claims.

What integrated information theory says

The neuroscientist Giulio Tononi proposed IIT in 2004. In the Internet Encyclopedia of Philosophy's account, the theory holds that consciousness is a kind of information that requires physical integration and can be measured by phi.

IIT works in an unusual direction. Most theories start from the brain and ask what it does. IIT starts from experience and lists features it takes to be self-evident: experience exists, it is structured, it is specific, it is unified, and it has definite borders. For each of these "axioms" the theory states a "postulate", a property the physical system must have. The highest value of phi found within a system measures how conscious it is. The current version, IIT 4.0, restates the axioms and postulates mathematically, so that in principle they could be applied to any system.

A sphere of fused glowing cells lit as one

What integrated information theory says about AI

Because phi is a property of physical cause and effect, IIT judges the hardware, not the software:

  • Feed-forward systems are not conscious. A system in which signals only pass one way has no integrated information; one that perfectly mimicked a conscious being would only simulate consciousness.
  • Conventional processors score low. Ordinary chips have few connections per circuit, so their phi is limited.
  • Simulation is not experience. In Dissociating Artificial Intelligence from Artificial Consciousness (2024), Graham Findlay, Christof Koch, Tononi and colleagues argued that a computer simulating a system with full functional equivalence need not share its experience.

That is the opposite of global workspace theory and AI, where the right kind of information processing would be enough, even after the 2026 discovery of a workspace-like structure inside Claude.

A closer look: IIT under test in 2025 and 2026

In April 2025 the Cogitate Consortium published the results of a head-to-head test of IIT and global neuronal workspace theory in Nature. The results matched some predictions of both theories but substantially challenged key tenets of each. For IIT, the study found no sustained synchronisation within the posterior cortex, which, the authors write, contradicts the claim that network connectivity specifies consciousness.

The September 2026 framework From cacophony to hierarchy, by Shamil Chandaria, Anil Seth, Murray Shanahan, Shane Legg and colleagues, adds a twist. It notes that IIT can be read as a theory about computation or as a theory about intrinsic physical causal structure, and it found that the verdict on current language models depends as much on where researchers place their trust among such theories as on the evidence. Because IIT's own verdict on ordinary computers is strongly negative, how much weight it gets matters a great deal. The guide can AI be conscious explains the range.

Older criticisms still stand, as the encyclopedia records. Scott Aaronson argued that simple mathematical structures could have higher phi than a human brain, John Searle objected that information is relative to an observer, and exact phi cannot be calculated for complex systems, so the theory cannot yet be checked directly on a brain or a large AI model.

Checklist: reading an IIT claim about an AI system

  • Is it about the hardware? IIT's verdict concerns the physical chips running a model, not its behaviour.
  • Is phi being calculated or estimated? Exact values are out of reach for large systems; most claims are arguments, not measurements.
  • Is the theory being assumed? A low-phi verdict only rules out consciousness if IIT is true, and the 2025 results challenged part of it.
  • Which version? IIT 4.0 changed details of the earlier formulations.

More in Theories of Consciousness · Report a correction