An Objection to Chalmers's Fading Qualia Argument

Would a neuron-for-neuron silicon isomorph of you have conscious experiences? Or is there something special about the biology of neurons, so that no brain made of silicon, no matter how sophisticated and similar to yours, could actually have conscious experiences?

In his 1996 book and a related 1995 article, David Chalmers offers what he calls the "fading qualia" argument that there's nothing in principle special about neurons (see also Cuda 1985). The basic idea is that, in principle, scientists could swap your neurons out one by one, and you'd never notice the difference. But if your consciousness were to disappear during this process, you would notice the difference. Therefore, your consciousness would not disappear. A similar idea underlies Susan Schneider's "Chip Test" for silicon consciousness: To check whether some proposed cognitive substrate really supports consciousness, slowly swap out your neurons for that substrate, a piece at a time, checking for losses of consciousness along the way.

In a recent article critical of Schneider, David Udell and I have criticized her version of the swapping test. Our argument can be adapted to Chalmers's fading qualia argument, which is my project today.

First, a bit more on how the gradual replacement is supposed to work. Suppose you have a hundred billion neurons. Imagine replacing just one of those neurons with a silicon chip. The chemical and electrical signals that serve as inputs to that neuron are registered by detectors connected to the chip. The chip calculates the effects that those inputs would have had on the neuron's behavior -- specifically, what chemical and electrical signals the neuron, had it remained in place, would have given as outputs to other neurons connected to it -- and then delivers those same outputs to those same neurons by effectors attached to the silicon chip on one end and the target neurons at the other end. No doubt this would be complicated, expensive, and bulky; but all that matters to the thought experiment is that it would be possible in principle. A silicon chip could be made to perfectly imitate the behavior of a neuron, taking whatever inputs the neuron would take and converting them into whatever outputs the neuron would emit given those inputs. Given this perfect imitation, no other neurons in the brain would behave differently as a result of the swap: They would all be getting the same inputs from the silicon replacement that they would have received from the original neuron.

So far, we have replaced only a single neuron, and presumably nothing much has changed. Next, we swap another. Then another. Then another, until eventually all one hundred billion have been replaced, and your "neural" structure is now entirely constituted by silicon chips. (If glial cells matter to consciousness, we can extend the swapping process to them also.) The resulting entity will have a mind that is functionally identical to your own at the level of neural structure. This implies that it will have exactly the same behavioral reactions to any external stimuli that you would have. For example, if it is asked, "Are you conscious?" it will say, "Definitely, yes!" (or whatever you would have said), since all the efferent outputs to your muscles will be exactly the same as they would have been had your brain not been replaced. The question is whether the silicon-chipped entity might actually lack conscious experiences despite this behavioral similarity, that is, whether it might be a "zombie" that is behaviorally indistinguishable from you despite having nothing going on experientially inside.

Chalmers's argument is a reductio. Assume for the sake of the reductio that the final silicon-brained you entirely lacks conscious experience. If so, then sometime during the swapping procedure consciousness must either have gradually faded away or suddenly winked out. It's implausible, Chalmers suggests, that consciousness would suddenly wink out with the replacement of a single neuron. (I'm inclined to agree.) If so, then there must be intermediate versions of you with substantially faded consciousness. However, the entity will not report having faded consciousness. Since (ex hypothesi) the silicon chips are functionally identical with the neurons, all the intermediate versions of you will behave exactly the same as they would have behaved if no neurons had been replaced. Nor will there be other neural activity constitutive of believing that your consciousness is fading away: Your unreplaced neurons will keep firing as usual, as if there had been no replacement at all.

However, Chalmers argues, if your consciousness were fading away, you would notice it. It's implausible that the dramatic changes of consciousness that would have to be involved when your consciousness is fading away would go entirely undetected during the gradual replacement process. That would be a catastrophic failure of introspection, which is normally a reliable or even infallible process. Furthermore, it would be a catastrophic failure that occurs while the cognitive (neural/silicon) systems are functioning normally. This completes the reductio. Restated in modus tollens form: If consciousness would disappear during gradual replacement, you'd notice it; but you wouldn't notice it; therefore consciousness would not disappear during gradual replacement.

As Udell and I frame it in our discussion of Schneider, this argument has an audience problem. Its target audience is someone who is worried that despite in-principle functional identicality at the neuronal level, silicon might just not be the right kind of stuff to host consciousness. Someone who has this worry presumably does not trust the introspective reports, or the seemingly-introspective reports, of the silicon-brained entity. The silicon-brained entity might say "Yes, of course I'm conscious! I'm experiencing right now visual sensations of your face, auditory sensations of my voice, and a rising feeling of annoyance at your failure to believe me!" The intended audience remains unconvinced by this apparent introspective testimony. They need an argument to be convinced otherwise -- the Fading Qualia argument.

Let's call the entity (the person) before any replacement surgery r0, and the entity after all their neurons are replaced rn, where n is the total number of neurons replaced. During replacement, this entity passes through stages r1, r2, r3, ... ri, ... rn. By stipulation, our audience doesn't trust the introspective or seemingly introspective judgments of rn. This is the worry that motivates the need for the Fading Qualia argument. In order for the argument to work, there must be some advantage that the intermediate ri entities systematically possess over rn, such that we have reason to trust their introspective reports despite distrusting rn's report.

Seemingly introspective reports about conscious experience may or may not be trustworthy in the normal human case (Schwitzgebel 2011; Irvine 2013). But even if they're trustworthy in the normal human case, they might not be trustworthy in the unusual case of having pieces of one's brain swapped out. One might hold that introspective judgments are always trustworthy (absent a certain range of known defeaters, which we can stipulate are absent), in other words, that unless a process accurately represents a target conscious experience it is not a genuinely introspective process. This is true, for example on containment views of introspection, according to which properly formed introspective judgments contain the target experiences as a part (e.g., "I'm experiencing [this]"). Infallibilist views of introspection of that sort contrast with functionalist views of introspection, on which introspection is a fallible functional process that garners information about a distinct target mental state.

A skeptic about silicon consciousness might either accept or reject an infallibilist view of introspection. The Fading Qualia argument will face trouble either way.

[A Trilemma for the Fading Qualia Argument (click to enlarge and clarify figure): Optimists about silicon chip consciousness have no need for an argument in favor of rn consciousness, because they are already convinced of its possibility. On the other hand, skeptics about silicon consciousness are led to doubt either the presence or the reliability of ri's introspection (depending on their view of introspection) for the same reason they are led to doubt rn's consciousness in the first place.]

If a silicon chip skeptic holds that genuine introspection requires and thus implies genuine consciousness, then they will want to say that a "zombie" rn, despite emitting what looks from the outside like an introspective report of conscious experience, does not in fact genuinely introspect. With no genuine conscious experience for introspection to target, the report must issue, on this view, from some non-introspective process. This raises the natural question of why they should feel confident that the intermediate ris are genuinely introspecting, instead of merely engaging in a non-introspective process similar to rn's. After all, there is substantial architectural similarity between rn at at least the late-stage ris. The skeptic needs, but Chalmers does not provide, some principled reason to think that entities in the ri phases would in fact introspect despite rn's possible failure to do so -- or at least good reason to believe that the ris would successfully introspect their fading consciousness during the most crucial stages of fade-out. Absent this, reasonable doubt about rn introspection naturally extends into reasonable doubt about introspection in the ri cases as well. The infallibilist skeptic about silicon-based consciousness needs their skepticism about introspection to be assuaged for at least those critical transition points before they can accept the Fading Qualia argument as informative about rn's consciousness.

If a skeptic about silicon-based consciousness believes that genuine introspection can occur without delivering accurate judgments about consciousness, analogous difficulties arise. Either rn does not successfully introspect, merely seeming to do so, in which case the argument of the previous paragraph applies, or rn does introspect and concludes that consciousness has not disappeared or changed in any radical way. The functionalist or fallibilist skeptic about silicon-based consciousness does not trust that rn has introspected accurately. On their view, rn might in fact be a zombie, despite introspectively-based claims otherwise. Absent any reason for the fallibilist skeptic about silicon-based consciousness to trust rn's introspective judgments, why should they trust the judgments of the ris -- especially the late-stage ris? If rn can mistakenly judge itself conscious, on the basis of its introspection, might someone undergoing the gradual replacement procedure also erroneously judge its consciousness not to be fading away? Gradualness is no assurance against error. Indeed, error is sometimes easier if we (or "we") slowly slide into it.

This concern might be mitigated if loss of consciousness is sure to occur early in the replacement process, when the entity is much closer to r0 than rn, but I see no good reason to make that assumption. And even if we were to assume that phenomenal alterations would occur early in the replacement process, it's not clear why the fallibilist should regard those changes as the sort that introspection would likely detect rather than miss.

The Fading Qualia argument awkwardly pairs skepticism about rn's introspective judgments with unexplained confidence in the ri's introspective judgments, and this pairing isn't theoretically stable on any view of introspection.

The objection can be made vivid with a toy case: Suppose that we have an introspection module in the brain. When the module is involved in introspecting a conscious mental state, it will send query signals to other regions of the brain. Getting the right signals back from those other regions -- call them regions A, B, and C -- is part of the process driving the judgment that experiential changes are present or absent. Now suppose that all the neurons in region B have been replaced with silicon chips. Silicon region B will receive input signals from other regions of the brain, just as neural region B would have, and silicon region B will then send output signals to other brain regions that normally interface with neural region B. Among those output signals will be signals to the introspection module.

When the introspection module sends its query signal to region B, what signal will it receive in return? Ex hypothesi, the silicon chips perfectly functionally emulate the full range of neural processes of the neurons they have replaced; that's just the set-up of the Fading Qualia argument. Given this, the introspection module would of course receive exactly the same signal it would have received from region B had region B not been replaced. If so, then entity ri will presumably infer that activity in region B is conscious. Maybe region B normally hosts conscious experiences of thirst. The entity might then say to itself (or aloud), "Yes, I'm still feeling thirsty. I really am having that conscious experience, just as vividly, with no fading, despite the replacement of that region of my brain by silicon chips." This would be, as far as the entity could tell, a careful and accurate first-person introspective judgment.

(If, on the other hand, the brain region containing the introspection module is the region being replaced, then maybe introspection isn't occurring at all -- at least in any sense of introspection that is committed to the idea that introspection is a conscious process.)

A silicon-chip consciousness optimist who does not share the skeptical worries that motivate the need for the Fading Qualia argument might be satisfied with that demonstration. But the motivating concern, the reason we need the argument, is that some people doubt that silicon chips could host consciousness even if they can behave functionally identically with neurons. Those theorists, the target audience of the Fading Qualia argument, should remain doubtful. They ought to worry that the silicon chips replacing brain region B don't genuinely host consciousness, despite feeding output to the introspection module that leads ri to conclude that consciousness has not faded at all. They ought to worry, in other words, that the introspective process has gone awry. This needn't be a matter of "sham" chips intentionally designed to fool users. It seems to be just a straightforward engineering consequence of designing chips to exactly mimic the inputs and outputs of neurons.

This story relies on a cartoon model of introspection that is unlikely to closely resemble the process of introspection as it actually occurs. However, the present argument doesn't require the existence of an actual introspection module or query process much like the toy case above. An analogous story holds for more complex and realistic models. If silicon chips functionally emulate neurons, there is good reason for someone with the types of skeptical worries about silicon-based consciousness that the Fading Qualia argument is designed to address to similarly worry that replacing neurons with functionally perfect silicon substitutes would either create inaccuracies of introspection or replace the introspective process with whatever non-introspective process even zombies engage in.

The Fading Qualia argument thus, seemingly implausibly, combines distrust of the putative introspective judgments of rn with credulousness about the putative introspective judgments of the series of ris between r0 and rn. An adequate defense of the Fading Qualia argument will require careful justification of why someone skeptical about the seemingly introspective judgments of an entity whose brain is entirely silicon should not be similarly skeptical about similar seemingly introspective judgments that occur throughout the gradual replacement process. As it stands, the argument lacks the necessary resources legitimately to assuage the doubts of those who enter it uncertain about whether consciousness would be present in a neuron-for-neuron silicon isomorph.

----------------------------------------

Related:

"Chalmers's Fading/Dancing Qualia and Self-Knowledge" (Apr 22, 2010)

"How to Accidentally Become a Zombie Robot" (Jun 23, 2016)

Much of the text above is adapted with revisions from:

"Susan Schneider's Proposed Tests for AI Consciousness: Promising but Flawed" (with David Billy Udell), Journal of Consciousness Studies, 28 (5-6), 121-144.