Masafumi Oizumi, Larissa Albantakis, Giulio Tononi (2014), PLoS Computational Biology 10(5), e1003588 (published 8 May 2014) — DOI-10.1371/journal.pcbi.1003588, open access (CC BY)
Key takeaways
- Axioms to postulates. Five axioms are “self-evident truths about consciousness”. Five postulates are “assumptions about the physical world”, which the authors call “unproven”.
- Exclusion, stated twice.
- Mechanism level: a mechanism specifies one cause and one effect, the maximally irreducible ones. This is defended as “a causal version of Occam’s razor … causes should not be multiplied beyond necessity”.
- System level: “Of all overlapping sets of elements, only one set can be conscious – the one whose mechanisms specify a conceptual structure that is maximally irreducible.” A complex is a local maximum of ΦMax “as compared to all overlapping sets of elements”. Complexes “cannot overlap”, and “no subset or superset of ABC can form another complex”.
- Identity. An experience is the maximally irreducible conceptual structure (MICS) of a complex. ΦMax is its quantity, and the shape of the constellation of concepts is its quality.
- Many complexes, never nested. A system “may condense into a major complex and non-overlapping minor complexes”. Minor complexes are “paraconscious”, and a split brain is two main complexes. Whether an octopus, an ant colony or a computer forms one complex or many “is not practically feasible” to determine, but “in principle testable”.
- Function does not fix consciousness. A feed-forward network can match a conscious network’s input–output behaviour and have no complex: “true ‘zombies’”.
- Unfeasible beyond toy systems. The analysis is exhaustive over partitions and is “unfeasible for systems of more than a dozen elements or so”.
Standing in the record
Filed on 2026-10-02 for the owner’s bridge between Minsky’s society of mind and Schwitzgebel’s conscious United States. NOTE-365 is the close reading, and it placed the work: Active. It is the anti-nesting principle itself, in IIT’s formal statement.
Is this the statement Schwitzgebel attacks? Partly. NOTE-131, which read Schwitzgebel’s 2014 manuscript of LIT-159, records that its §2 engages Tononi 2008–2012, this paper and Tononi & Koch 2014. The postulate it targets, that where integrated systems nest consciousness occurs only at the most-integrating level, is, in substance, this paper’s system-level exclusion. The strict “local maximum” comparison is also what makes Schwitzgebel’s one-ballot threshold possible. But the defence he attacks is not this paper’s. He has exclusion defended by Occam’s razor and the absurdity of a two-person group consciousness. Here Occam’s razor supports exclusion only within a mechanism: one cause, not “strong synapses” and “all synapses” at once. Exclusion between systems rests on the phenomenological axiom alone (“each experience has definite borders”). Group consciousness is never mentioned, and ant colonies are left open. NOTE-131 also records that the published version of LIT-159 drops this paper from its references. The later statements are Tononi & Koch 2015, to which the book version of the argument (LIT-216, read in NOTE-188) replies, and IIT 4.0 (2023), which LIT-185 (read in NOTE-122) takes as the current theory and which keeps exclusion as an axiom.
On NOTE-131’s numbering, this paper contests C3 for any whole that overlaps a conscious part with higher Φ. It is the target of C5, and nothing in it answers C5’s reductio. It is neutral on C4, since Putnam is not cited, although its system-level exclusion is the formal descendant of Putnam’s stipulation. Two things in it cut the other way and a bridge should keep them:
- exclusion is to be evaluated “not only over elements, but also over spatial and temporal grains”, and “the macro may emerge over the micro, just like the whole may emerge above the parts”. So the theory as stated does not privilege the smaller system. A whole at a coarse grain can win, which is the opening the book version of Schwitzgebel’s argument uses (via Hoel, Albantakis & Tononi 2013, this paper’s reference 20);
- exclusion forbids overlap, not plurality. A brain may host one major and several minor conscious complexes side by side.
For the decomposition side, the zombie result is the sharpest thing here. On IIT, organising mindless parts so that the whole behaves like an agent does not bring experience with it. Whether it does depends on the causal architecture, recurrent or feed-forward, not on the behaviour. Minsky’s agents (LIT-429), Brooks’s behaviour layers (LIT-436) and Levin’s goal-pursuing Selves (LIT-439) all sit under that verdict, and none of the three engages it.
It carries no instruction for machine-learning practice. Its remark that feed-forward classifiers would be unconscious is a claim about ML systems under IIT, not advice, and the anthology does not hold the paper.