NOTE-818: Gidon, Aru and Larkum — Does neural computation feel like something?

Contribution

The paper turns the authors’ earlier thought experiment (Gidon et al. 2022, “Does brain activity cause consciousness?”) into a simulation, and adds one case to it. A feedback replay, a voltage clamp whose command voltage is the recorded response, leaves the network to run unaided when the stimulus matches, and corrects it only when it would deviate. The case is used to show that, on computational functionalism, whether a system is conscious can turn on an intervention that never acts. The argument is Maudlin’s (1989) point about counterfactuals, run on a biophysical model rather than on a hypothetical machine.

Key insight

Computation is defined over the transitions a system would make on other inputs, not only the one it makes. A device that removes every other transition, and stays idle on the actual one, therefore changes the computation without changing anything that happens. If experience is fixed by computation, it changes too. Whoever finds that absurd must say experience depends on something other than the computation; whoever accepts it must say an idle device can alter what a subject feels.

Assumptions

• Computational functionalism as the working hypothesis. “Some computations constitute conscious experiences”; here, a subject that performs hue-computation is hue-conscious. The authors say they “initially accept this working hypothesis but shall reevaluate it”.
• Computation as Turing and Putnam define it. “A system receives various inputs that trigger state transitions according to a set of rules (or an algorithm), leading to an output”, and “the inputs, states, transitions, and outputs define the functional organization”.
• Core and auxiliary computations. By stipulation, altering a core computation’s input/output function alters experience. Hue-computation is assumed core. The reader is invited to substitute any computation they think is core, including a global workspace.
• Changing functional organization is changing consciousness. Cited to Putnam, Shagrir, Block and Lewis. This is the premise that carries the conclusion.
• Causation read mechanistically, not counterfactually (Ross and Bassett 2024 rather than Lewis 1973). That is why an idle clamp is said to leave “cause-and-effect relations between the neurons” intact.
• Consciousness is not defined. It is “the experience of oneself or one’s surroundings that fades during deep sleep or anesthesia”.
• Deterministic model. Biological noise is set aside: the replay “would undo the ongoing noise and reintroduce the recorded noise”. Theories that need “true stochasticity” are explicitly not addressed.

Key results

• The model (Methods, Fig. 1). Five single-compartment Hodgkin–Huxley-type neurons (Na⁺, K⁺, leak; R_in = 50 MΩ), each driven by the preceding three in cyclic order; synapses with 2 ms rise, τ_syn = 10 ms, g_syn = 5 nS, and short-term depression to 50% per spike recovering with τ = 1 s. Two LGN nodes are random spike trains. Green light turns on both and all five neurons fire; red or blue turns on one and only neuron a or b fires.
• Two replays (Fig. 2). A feedforward replay overwrites each neuron’s voltage with the recorded trace and so “severed the connection between the neurons”. A feedback replay is a simulated voltage clamp (NEURON’s SEClamp) that injects current only when the voltage departs from the command trace. With no stimulus, both recreate the green response; the clamp current equals the missing LGN current (Fig. 2A2).
• Incongruent input (Fig. 3A). Red light plus green replay: all neurons fire as for green, and the clamp supplies only the input from the second LGN node. The replay “remained equally effective whether the neurons … were disconnected or rewired, or a new input was introduced (not shown)”.
• The computation degenerates (Figs. 3D, 4). Under replay, the state transitions no longer depend on the input. In the toy algorithm every branch but the recorded one is pruned, leaving “a fixed sequence of operations independent of the input”. In the Turing-machine version the clamp becomes part of the transition function, and the result is “a trivial computation consisting of a single sequence of states (i.e., universally realizable)”.
• Nothing type I and type II (Fig. 3B–C). Green light plus green replay: the clamp injects no current, because “no intervention was needed”. With the amplifier off (type I), activity is identical. With it on and idle (type II), activity is again identical, but the counterfactual transitions are gone. On the working hypothesis, type I is hue-conscious and type II is not, or is less so. “If the replay did nothing, how could it change consciousness?”
• Counterfactual eraser. The idle clamp “can prohibit only counterfactuals without affecting anything else, i.e., brain activity and cause-and-effect between the neurons”. By contrast, “green glasses” that turn all input green leave the cortical computation intact; they give identical activity but, for the functionalist, do not alter experience.
• Is the clamp part of the system? (Discussion.) The functionalist’s question is whether the functional organization changed, and it did: states “that would have been reached … are unreachable”. So “the feedback replay device is conceptually part of (and alters) the Turing Machine”. How much it changes depends on “the proportion of states used in the congruent case compared to all other cases”.
• IIT (Discussion). For IIT the clamp degenerates the transition probability matrix, giving “a simpler cause-and-effect structure that dramatically reduces Φ”. This is “not … a formal challenge to IIT”, because IIT locates consciousness in what the system is (its cause–effect structure), not in what it does. The shared oddity is that, for both, “the feedback replay can degrade consciousness, despite doing nothing”.
• Frankfurt (Discussion). The eraser is a Frankfurt device that prescribes every neuron’s activity rather than an outcome. If counterfactuals matter to consciousness, Jones may act exactly as prescribed “but with altered or possibly diminished conscious experience”.
• Conclusion. Substrates realise more functional states than the computation ascribed to them. “Natural counterfactuals” are “embodied in the biological brain’s composition, structure, and dynamics and, therefore, cannot be erased without altering the brain’s physical structure and/or dynamics.” The biological brain is “the only known substrate capable of consciously experiencing the world”. The risk “lies in ascribing machines with consciousness when they are not”.

Claims

id	claim	strength	support
C1	A feedback voltage-clamp replay reproduces a recorded network response exactly, supplying only the input that is missing	strong	the simulation, Figs. 2–3 (trivially true of an ideal clamp)
C2	Under the replay, the network’s state transitions no longer depend on its input, so the computation it implements degenerates to a single sequence	strong	Figs. 3D, 4, given the Turing–Putnam definition of computation
C3	When the input matches, the clamp injects no current, so activity and actual neuron-to-neuron causation are unchanged	strong	Fig. 3B2, with causation read mechanistically
C4	So computational functionalism must hold that an intervention that does nothing can change conscious experience	moderate	C2–C3 plus the premise that changing functional organization changes consciousness; an argument, not a finding
C5	IIT gives the same verdict (Φ falls) without contradiction, because it rests on cause–effect structure	weak	asserted, Φ not computed
C6	Consciousness depends on “natural counterfactuals” fixed by the substrate, not on computation	weak	offered as an “alternative view” in the closing section, not argued
C7	Attributing consciousness to machines that lack it is the larger moral risk	weak	an opinion, with no argument from the simulation

Method

1. Simulate a five-neuron recurrent “visual cortex” with two LGN inputs in NEURON; record every neuron’s voltage for a green stimulus.
2. Replay the recordings by overwriting (feedforward) or by a voltage clamp with the recordings as command voltage (feedback).
3. Present no stimulus, an incongruent stimulus (red), or the congruent one (green), and record the clamp currents.
4. Map the result onto a toy branching algorithm and a Turing machine whose transition function absorbs the clamp.

Concepts

• Computation. Input-triggered state transitions by rule, ending in an output; inputs, states, transitions and outputs make up functional organization.
• Core vs auxiliary computation. Core: altering its input/output function alters experience. Auxiliary: does not.
• Feedback replay / counterfactual eraser. A monitor that holds each element to a prescribed trajectory and acts only on deviation. Its “key feature … is their potential intervention rather than their actual intervention”.
• Nothing type I / type II. The replay idle because it is switched off, versus idle because nothing deviates.
• Natural counterfactuals. The alternative states a substrate’s physical make-up allows, as distinct from those a computation ascribed to it allows.

Connections

The paper extends Gidon, Aru and Larkum (2022), which recorded and replayed a whole brain’s activity and ended in a reductio about scattered neurons. It re-runs Maudlin’s (1989) “Olympia” argument, that consciousness cannot depend on counterfactual dispositions that are never exercised, and cites Barnes (1991) against it, but engages neither in detail. It relies on the Putnam and Searle point that a computation without counterfactuals is “universally realizable”, which is what Chalmers’s rock paper (LIT-404, read in NOTE-342) answers by requiring “reliable, counterfactual-supporting transitions”: on that criterion the clamped network implements only a trivial, inputless automaton. It sits opposite Pylyshyn’s and Chalmers’s neuron-replacement arguments (LIT-392, NOTE-347), which keep the causal organization and swap the substrate; this keeps the substrate and activity and removes the organization’s unused branches. The Frankfurt section borrows the device of LIT-694.

Bearing on the record

• QUESTION-059 (Deferred). The LIT was filed as “a simulation-based test of whether meeting the functional role suffices for feeling”. That overstates it twice.
  • It is not a test. The simulation fixes nothing about experience; it only shows that an idle clamp leaves activity unchanged while fixing transitions, which is true of any ideal clamp. The philosophical work is an argument from the definition of computation. QUESTION-059 says another argument from a definition would not answer it.
  • It is about consciousness in general, “the experience … that fades during deep sleep”, and says nothing of affect, valence or feeling in the affective sense. The title’s “feel like something” is the phrase for phenomenal consciousness. What it does supply is a dilemma for the identity side. Veit’s identity of feeling with “occupy[ing] a coordinated regulatory state” (THEORY-101) is a functional identity, so it inherits the eraser: the same regulatory activity, with its unused alternatives removed, would have to be a different feeling or none. Veit would need to say either that the state is individuated by actual causal activity (then the identity is not with a functional role in the Turing–Putnam sense) or that the eraser changes feeling. Neither the paper nor THEORY-101 says which.
• THEORY-045, THEORY-093, THEORY-090. The paper’s “natural counterfactuals … embodied in the biological brain’s composition” points the same way as the bodily-mechanism side of QUESTION-059, but only as a closing suggestion; it gives no mechanism and mentions no body. It is not evidence for those accounts.
• THEORY-023. The paper sides with the biological naturalist on AI consciousness, and its “misconsciousing” section asserts that false attribution is the greater risk. It does not touch the evidential deadlock THEORY-023 describes; it is one more statement of one side.
• THEORY-092. Block’s lookup machine matches every actual and counterfactual behaviour and lacks intelligence; the clamped network matches every actual activity and lacks the counterfactuals. Both say the mental property is not fixed by what actually happens, from opposite directions.
• THEORY-119. Frankfurt holds that a device that “plays no part” in an action cannot change responsibility for it. The paper argues that, on functionalism, a device that plays no part can change the agent’s consciousness. The two are compatible, but together they mean a functionalist who accepts Frankfurt holds an agent responsible for an act she may not have experienced, which is the question the paper ends that section on.
• THEORY-043. Its anti-nesting principle is a constraint on what counts as the system; the paper’s claim that the clamp is “conceptually part of” the Turing machine is the same kind of boundary decision, taken without a principle.
• No instruction for machine-learning practice; nothing for the anthology.

Limitations

• The simulation carries no evidential weight on consciousness. Its outputs are voltages and clamp currents. The consciousness verdicts come from the working hypothesis, applied by definition.
• The key premise is cited, not defended. That changing functional organization by removing unexercised branches changes experience is exactly what a functionalist following Maudlin’s critics (Barnes 1991, cited) or Klein (2018, cited) would contest. Those replies are not engaged.
• The size of the change is not estimated. The authors concede it depends on the proportion of states used, and the paper says only “possibly diminished or even abolished”.
• The core/auxiliary distinction is stipulated and admitted to be only implicit in functionalist theories.
• The positive alternative is not developed. “Natural counterfactuals” is named in one paragraph; what fixes them, and why they would matter for experience, is not said.
• Noise and stochastic theories are set aside, so the result says nothing about accounts on which experience needs genuine randomness.

Open questions

• Does a functionalist who makes implementation turn on counterfactual-supporting transitions (LIT-404) accept the Nothing type II verdict, or deny that a clamp outside the neurons is part of the implementing system? A reading of a reply by a functionalist would settle which.
• Is there a principled boundary on the system, other than “whatever changes the transition function”, that would put the clamp outside it?
• What, concretely, are natural counterfactuals, and could a non-biological substrate have them?