NOTE-340: List — What is it like to be a group agent?

Contribution

The paper separates three things that the group-mind debate tends to run together: agency, awareness and phenomenal experience. It grants Schwitzgebel (LIT-159) the first two for organised groups, and locates the whole disagreement in the third, which only a psycho-physical bridge principle can settle. It then gives the first IIT-based case, in print, that actual group agents have little or no phenomenal consciousness, and connects that to a normative claim from List and Pettit’s Group Agency (LIT-391): because groups are not conscious, their rights are only derivative.

Key insight

Schwitzgebel’s marks of consciousness (information processing, self-monitoring, goal-directed responsiveness) are all functional, and List accepts that a group has every one of them, as awareness. A bridge principle that ties experience to function alone will then either credit the group or be ad hoc. IIT is the one candidate that ties experience to a physical, hardware-neutral quantity that does not track function: Φ, the irreducibility of the system’s causal dynamics to a partition. Function can be organised in a way that is easy to partition. A society of self-contained units, whose steering board still works when part of it is randomised and whose members do the heavy computing, can be a competent agent and still score low.

Assumptions

• Functionalism about agency (§2). An agent is any system with representational states (“beliefs”), motivational states (“desires”) and a capacity to act on them (pp. 5–6). A thermostat qualifies “in a very basic sense” (p. 6).
• Naturalistic indispensability argument for group agents (pp. 6–7). Premise 1: our best social-scientific theories (the theory of the firm) attribute belief–desire agency to collectives. With a thin Quinean criterion of ontological commitment, they are committed to group agents. Premise 2: take that commitment at face value, defeasibly (Fine’s “natural ontological attitude”). The conclusion is a “cautious form of realism about group agents” (p. 7).
• Group agency is not joint agency (pp. 8–9). A joint action needs only individual intentions suitably directed; a group agent needs a single “centre” of belief, desire and agency. Joint agency is “neither sufficient, nor … even necessary” for group agency.
• The awareness/experience distinction (§3), following Chalmers 1995. Awareness is functional and can be studied third-personally; experience is not functional, as the conceivability of zombies is taken to show (pp. 11–12).
• Eliminativism (type-A materialism) is set aside, “on the grounds that it does not do justice to the ‘what is it like’ question” (p. 14). Type-B materialism and Chalmers’s naturalistic dualism are both kept, and on both the question becomes which bridge principle holds (pp. 14–16).
• IIT is assumed as the bridge principle, in Aaronson’s simplified exposition of Φ (§6.1), explicitly not IIT 3.0 (n. 17). For the argument it suffices that a high Φ is necessary for a high level of phenomenal consciousness (p. 34).

Key results

All are informal arguments. There is no theorem, no data, and no Φ computed for any system.

• Group agents have consciousness as awareness (§3, pp. 12–13). The FBI is aware of information in an investigation (Goldman 2004); an organisation hit by scandal attends to it; group perception is mediated by members and procedures. Even “wakefulness and sleep” fit: List’s university goes on vacation and “wake[s] up” only for an emergency (p. 13). This is offered as realism, “not just metaphorical”.
• Survey of bridge principles (§5), each with its group verdict.
  • 1a Agency or cognition implies consciousness: too many false positives (thermostats, robots, groups).
  • 1b Agency above a complexity threshold: ad hoc, and admits groups that “inherit the relevant cognitive capacities from their human members” (p. 18).
  • 1c Higher-order cognition: nothing bars groups, but it fails for non-human animals (Carruthers 2011).
  • 2a Mammalian brains: excludes groups, except brain-simulating populations (Block), but cannot tell sleep and coma from waking.
  • 2b Synchronised neural activity (Crick & Koch 1990; Clark 2009): implementation-specific, and consciousness is lost in synchronous seizures (Tononi & Koch 2015).
  • 3a Information processing (Chalmers’s double-aspect view): a panpsychism, too unrestricted, and silent on the unity of consciousness.
  • 3b Global workspace (Baars): permits group consciousness, but only “asserts functionalist conditions for access consciousness and then stipulates” that they are phenomenal ones (p. 24).
  • IIT: phenomenal consciousness rises with integrated information.
  • Desiderata for any principle (pp. 19–20): capture the paradigm cases (humans, chimpanzees, dogs yes; tables, combustion engines no), non-ad hoc, revisionary only in marginal cases and then for reason.
• Φ, as used (§6.1, pp. 26–29). Binary components, a state-transition function f on Sⁿ. For a bipartition {A, B}, EI(B→A) = H(Y_A | X_A) with X_B uniform. Φ is the minimum over bipartitions of [EI(B→A) + EI(A→B)] / min(|A|, |B|).
• Evidence for IIT (§6.2). It accounts for cortex versus cerebellum, and for loss of consciousness in sleep, anaesthesia and coma via the perturbational complexity index (quoted from Tononi & Koch 2015). Both are reported, not examined.
• Three reasons a group agent’s Φ is low (§6.3, pp. 32–33).
  • Reason 1. Large organisations “are decomposable into functionally relatively self-contained units, which are each internally more interdependent than they are dependent on other units”. So the whole cannot have high Φ; “at most, some smaller sub-units might”.
  • Reason 2. Even a board or assembly, the “cortex” analogue, “would still retain much of its functioning” with part of it replaced by a random process, because members respond rationally and procedures are robust. So EI across the cut stays low.
  • Reason 3. Much of the group’s processing is done by individuals; the “computational contribution made by cross-member connections, while non-negligible, is still moderate”. List says Schwitzgebel attributes a related point to Chalmers and quotes it.
  • Verdict: Φ is non-zero but “would not be even remotely close” to a rat’s or mouse’s (p. 33). A “biomorphic” science-fiction collective modelled on a neural network might score higher (p. 35).
• Against Schwitzgebel (§6.4, pp. 34–35). List agrees that the US is “a candidate” for psychological states and has non-zero Φ. He says Schwitzgebel’s “massively complex informational integration” is the informal functionalist notion, which the US undeniably has, and that Reasons 1–3 count against high Φ in the formal sense. “Glass largely empty” rather than “glass partly full”.
• Exclusion (§6.5, p. 36). Where systems nest, consciousness is at the Φ-maximal level only. In a group, Φ “presumably” peaks at the members, so their consciousness excludes the group’s even if group Φ were substantial. List records that Schwitzgebel rejects exclusion and does not answer him; he says the low-Φ argument suffices without it.
• Normative implication (§7, pp. 37–40). List and Pettit 2011 (chs 7–8) asserted that groups should be held responsible but not given individual-grade rights, and critics (Briggs 2012) said functionalism alone could not justify the asymmetry. If phenomenal consciousness is necessary for non-derivative moral significance, and groups lack it, then group rights must be justified by their service to individuals. Citizens United is the worked example. The final answer: “(close to) nothing” (p. 40).

Claims

id	claim	strength	support
C1	Organised collectives are agents in a thin functionalist sense, and realism about them is warranted	moderate	indispensability argument, pp. 6–7, defeasible by its own statement; the “black box” mechanisms are deferred to List & Pettit 2011
C2	Group agents have consciousness as awareness, literally	moderate	examples (FBI, scandal, vacation), p. 12–13; depends on the awareness/experience split
C3	Whether groups are phenomenally conscious depends wholly on which bridge principle is true	moderate (conditional on rejecting eliminativism)	conceptual, §4
C4	Agency-based, higher-order, biological and global-workspace principles are each inadequate as general bridge principles	weak–moderate	one or two objections each, §5; GWT is faulted for not explaining experience, a standard objection
C5	A typical group agent’s Φ is low, far below a rodent’s	weak	three structural reasons, §6.3; asserted (“I suggest”, “I believe”); no Φ estimated for any group or animal; uses the simplified Aaronson Φ
C6	High Φ need only be necessary for high consciousness, so Aaronson’s high-Φ unconscious systems do not touch the argument	strong as logic	§6.3, p. 34
C7	Exclusion independently rules out group consciousness if members have higher Φ than the group	conditional	§6.5; exclusion is assumed, and Schwitzgebel’s objections to it are not answered
C8	Lack of group consciousness justifies a normative asymmetry: groups have only derivative rights	weak–moderate	§7; needs the further premise that consciousness is necessary for non-derivative moral significance, which n. 19 says many theories share

Concepts

• Group agent — a collective that meets the general definition of an agent; how it comes to do so is a separate, practical question (p. 5).
• Joint agency — individuals acting together on shared or suitably directed intentions, with no group-level centre (p. 8).
• Consciousness as awareness / as experience — the functional, third-personal sense (access, attention, wakefulness) against the “what it is like” sense (§3).
• Psycho-physical bridge principle — a specification of the functional conditions under which a system is phenomenally conscious; a necessary truth for type-B materialism, a contingent law for naturalistic dualism (pp. 15–16).
• Low-entropy partition — a bipartition across which neither side’s next state depends much on the other’s; its existence means low Φ (p. 32).
• Normative individualism — only individuals (and perhaps individual animals) are ultimate units of moral significance (p. 37).

Connections

• Schwitzgebel (LIT-159; NOTE-131). The paper is a direct reply. It quotes the “liberal enough” conclusion and the Chalmers objection, and it agrees with Schwitzgebel against neural-synchrony criteria (n. 11) and on not ruling out aliens by stipulation (n. 10). Schwitzgebel’s later book (LIT-216; NOTE-188, its n. 56) answers it by attacking exclusion with the sleeping-citizen case, which does not touch List’s main, non-exclusion argument.
• Minsky (LIT-429, LIT-399, LIT-402; Singh, LIT-444). Not cited. Reason 1 is a near-decomposability argument: a whole built from self-contained specialist units is cheap to partition. The same reasoning bears on a society-of-agents mind. If Minsky’s agents are as loosely coupled as his “mindless” specialists are usually described, IIT would give such a mind little Φ for the same reason it gives a corporation little. List’s argument is indifferent to whether the parts are minded, which is what makes it a bridge between the decomposition and composition questions rather than an anti-nesting principle.
• IIT (LIT-185; NOTE-122). The paper relies on Tononi & Koch 2015 and Aaronson 2014, not on the IIT 4.0 that Barrett et al. assess. Their finding that the algorithm is undefined for non-Markovian grainings bears directly on applying Φ to a polity, and so on C5.
• Butlin et al. (LIT-056). List’s dismissal of GWT as a bridge principle for experience is a reason to doubt that LIT-056-style indicators can settle the group case either way.
• Collective intelligence (LIT-046) and “Real Patterns” (LIT-220). The first is a computational account of what groups can do; this paper explains why that leaves experience open. The second is Dennett’s stance, mentioned for thermostats (p. 6).
• Tomasello (LIT-125). The joint-versus-group agency distinction of §2 (after Pettit & Schweikard 2006) separates the kind of “we” that Tomasello’s account of obligation needs (joint commitment) from the corporate agent List is asking about.
• Kammerer (LIT-409). The other contemporary materialist reply to Schwitzgebel. Kammerer keeps a functionalist bridge principle and adds an exception for wholes whose function runs through parts representing them. List changes the bridge principle instead. The two are independent: List’s argument goes through if the members represent nothing, and Kammerer’s if Φ is irrelevant.
• Functionalism and anti-nesting. Block’s China-brain style population simulation is the one case List’s Principle 2a might still count conscious (p. 21). Putnam’s anti-nesting stipulation is not discussed. Both are being filed by another reader in this batch.
• Not held. Tollefsen’s Groups as Agents (2015), Theiner’s “A Beginner’s Guide to Group Minds” (2014), Huebner, Bruno & Sarkissian 2010 on Chinese intuitions, and Knobe & Prinz 2007. Each was searched for in literature.d by name.

Bearing on the record

• For the bridge THEORY. This is the strongest held source for the claim that agency can compose without experience composing with it. It should be cited for the agency/awareness/experience split and for the conditional (on IIT) verdict. It should not be cited as showing that a group’s Φ is low: that is asserted from three structural reasons, and nothing is computed.
• On anti-nesting. It is not an anti-nesting principle. It does lean on exclusion as a secondary route, but its main argument concerns how partitionable the whole is, whether or not the parts are conscious.
• On LIT-159’s open question (does a formal IIT calculation put a polity’s Φ above a brain’s?). This is the closest held attempt, and it does not compute anything.
• No ML instruction. Nothing belongs in the anthology.

Limitations

• No Φ is computed or estimated. Reasons 1–3 are qualitative. Reason 2 is about a board’s robustness to randomising some of its members, which is a claim about function, not about the conditional entropies in the definition.
• The simplified Φ. The argument uses Aaronson’s exposition, not IIT 3.0 (n. 17), and assumes binary components and a known wiring diagram. List himself notes that computing Φ is probably intractable (p. 31).
• Reason 1 proves too much, or too little. Brains are also modular. The paper relies on the cortex being hard to partition and does not show that a group’s units are more separable than cortical areas are.
• The Chalmers reply is not answered. Reason 3 is the objection Schwitzgebel had already met with global-workspace and “fame” models (NOTE-131). List does not address that reply.
• Exclusion is used, not defended. §6.5 records Schwitzgebel’s rejection and moves on.
• Conditional throughout. The author calls the argument “tentative and conditional” (pp. 4, 40) on IIT.

Open questions

• What is Φ, under any version of IIT, for even a small formal model of a committee or firm, compared with a cortical model of similar size? This is the computation the paper’s verdict assumes.
• Does Reason 1 apply to a Minsky-style society of mind? If the brain is such a society, what keeps its Φ high: dense recurrent coupling between agents, which Minsky’s K-line and c-line memos (LIT-402, LIT-399) would have to supply?
• Is a high-Φ “biomorphic” collective (p. 35) possible with human members, and would exclusion then strip the members of consciousness (the case Schwitzgebel’s one-ballot reductio targets)?

Corrections

• No dossier existed, so there is no seeded skim to correct.
• Date. The request gave 2018. That is the issue year (Noûs 52(2), June 2018). The article was published online on 28 July 2016 (Crossref), and List’s own manuscript cites itself as 2016. Volume, issue, pages (295–319) and DOI were confirmed on Crossref.
• A numbering slip in the manuscript. Both the global-workspace principle and the IIT principle are labelled “Principle 3b” (pp. 23 and 25). Whether the version of record fixes this was not checked.