On Shiller's "Bomb Threats for Functionalists"

Possible solutions, collateral threats, and implications for AI consciousness

This post shares some highlights and my reactions to Derek Shiller’s interesting recent article “Bomb Threats for Functionalists”.

1. Background on functionalism and the bomb

Shiller understands functionalism about consciousness as the view that a system is conscious if and only if it has the right sort of dynamics and occupies the right sort of state within that dynamics. Equivalently, a system needs to have the right sort of functional organization and undergo the right sort of transitions between states within that organization.

Functionalist views vary on what the right sorts of functional organization and transitions are. But they agree that non-functional factors are irrelevant to consciousness except insofar as they’re needed for the right organization and transitions.

Shiller uses the following thought experiment to pose a problem for functionalists:

The Bystander and the Bomb. There is an ordinary human being standing next to a touchy bomb that is poised to explode. At every millisecond, the bomb has a high chance of detonating. Over any stretch of time, there is a small chance that it won’t. If the bomb explodes, the bystander would be vaporized and every part of their brain would immediately transition into vapor. Whether the bomb detonates is sensitive to minute vibrations in the air. This means both that the explosion is produced by physically chancy processes and that the fact that the bomb does or does not actually detonate says nothing about whether it would have detonated had the bystander acted in any other way. By sheer luck, the bomb never detonates.

The question is whether the ordinary human in this case is conscious. The intuitive answer is that they are. The bomb problem for functionalists is that of providing a functionalist basis for that answer.

The challenge is pressing because the bomb severely disrupts the counterfactuals and probabilities that constitute the bystander’s functional organization. Where they would normally be disposed toward various neural-state-to-neural-state transitions, the bomb’s presence would seem to replace those with highly likely transitions going from neural matter to vapor. So, why doesn’t the mere threat of the bomb wreak havoc on the transition probabilities in the bystander’s brain, thereby disrupting their functional organization and rendering them unconscious?

2. Some candidate solutions and their limitations

To solve this problem, Shiller thinks functionalists need to specify the functional organization on which consciousness depends in a manner that deems the bomb threat irrelevant. Shiller considers a range of potential solutions and argues that the most obvious candidates solutions to this problem for functionalism either don’t work or else raise further problems.

Let’s have a look at some of the candidate solutions and their liabilities.

Candidate solution #1: a system’s consciousness-relevant functional organization is defined independently of external forces.

Problem: our conscious brains rely on external forces such as pressure.

Candidate solution #2: a system’s consciousness-relevant functional organization is defined independently of abnormal external forces.

Problem: This proposal ties consciousness to normal conditions. But what’s normal depends on non-local factors that we’d expect to be irrelevant to consciousness. For instance, if normality is cashed out in terms of population-level statistics or evolutionary history, then whether an atom-for-atom duplicate of you would be conscious would depend on what happens to other individuals at distant places and times.

Candidate solution #3: a system’s consciousness-relevant functional organization is defined independently of destructive external forces

Problem: the bomb case could be varied so that the relevant functional organization is modified in a non-destructive manner. (Perhaps a chance device renders it highly probable that transcranial magnetic stimulation will be applied to your visual cortex in a way that disrupts your visual experience, but you’re in one of the unlikely cases in which the device remains inactive.)

Alternatively, if destructiveness is understood relative to a system’s normal functioning, then this proposal is an instance of the previous candidate solution and so faces the non-locality problem.

Candidate solution #4: a system’s consciousness-relevant functional organization is defined relative to the assumption that actual forces on the system at a time are held fixed. Since the forces the bomb would have exerted on the system if it had detonated are non-actual, they therefore don’t disrupt that organization.

Problem: If actual forces are held fixed, then this solution generates the counterintuitive prediction that systems whose functioning depends on variation in what forces are exerted upon them will generally not be conscious, even if they are otherwise functionally like some conscious system.

Candidate solution #5: the bomb doesn’t disrupt functional organization because it doesn’t remove dispositions.

Problem: It’s doubtful that there’s a general fact of the matter about, say, whether or not a particular person is disposed to anger quickly or whether various interventions would remove that disposition as opposed to preventing it from manifesting. What’s more likely is that we ascribe dispositions based on pragmatic groupings of counterfactuals that result in there being no fact of the matter about such things in many cases.

Comment: I’m sympathetic with this picture of dispositions that we ordinarily attribute and agree that it will result in cases in which it’s indeterminate whether certain dispositions have been removed. But I also think it’s open to functionalists to develop the solution under consideration by either appealing to dispositions about which there are generally facts of the matter or by allowing for consciousness to be indeterminate when it’s indeterminate whether consciousness-relevant dispositions have been removed.

Problem: in order for this solution to predict that the bystander is conscious, it needs to be supplemented with a theory—or at least an outline of a theory—that tells us when dispositions are removed and which predicts that the bomb doesn’t remove the consciousness-relevant dispositions. But how to formulate such a theory is just a variation of the problem that this response was introduced to solve.

Comment: I’d agree that more would need to be said in order to render this solution promising. On the other hand, that by itself is little reason to think this solution won’t work. A natural suggestion¹, which I’ll return to below, is that the functional organization that’s relevant to consciousness only features dispositions that just depend on internal features of the system.

3. Functionalism without dispositions

Shiller’s exploration of solutions to the bomb threat problem is largely premised on the idea that functionalists need to find a way of specifying consciousness-relevant functional organization that is impervious to the presence of the bomb.

This premise can be challenged. In particular, the functionalist could respond to the bomb problem by abandoning the appeal to functional organization. Would this not be to abandon functionalism itself? I don’t think so: they could retain a commitment to consciousness arising if and only certain functional conditions obtain. More specifically, they could hold that what matters for consciousness is just that certain causal relations between states of a system obtain, not the dispositional character of that system’s functional organization.²

Shiller anticipates and objects to a view along these lines, writing:

However we define functional roles, a system might fail to be conscious if the structures supporting counterfactual transitions are removed or suppressed. This feature generates some unintuitive verdicts but is a core commitment of the view needed to keep it from collapsing into absurdity (Chalmers, 1996).

Chalmers appeals to counterfactuals in the context of offering an account of implementing functional organizations that steers clear of triviality problems such as avoiding the result that a wall is implementing a particular computer program. The triviality problems arise because simple mapping accounts of what it takes to implement a particular computer program result in a very wide class of physical systems trivially implementing a very wide class of programs. When combined with functionalist or computational functionalist views of consciousness, these triviality problems threaten to lead to panpsychism, which isn’t supposed to be a consequence of such views.

By severely constraining what it takes to implement a functional organization, imposing counterfactual constraints offers a way to avoid such triviality problems. So far so good. However, there’s a further question as to whether counterfactual constraints are the only game in town when it comes to preventing panpsychism from trivially following from functionalist views. In my view, an important alternative can be found in manifest causal relations, that is actually holding causal relations that qualify their relata as actual causes and effects. These relations come in many varieties, and I take it there are obviously many substantive manifest-causal differences between, say, a wall and a laptop that’s running Microsoft Word. So, there’s at least a presumptive case for thinking that the basis of consciousness could consist in a certain sort of causal activity without panpsychism trivially ensuing.

It’s also true that functionalists have historically leaned on the intuition that mental states are associated with characteristic dispositions and that functionalism is well-positioned to account for that by explaining mental states in terms of dispositions. For instance, upon noticing that pain is associated with dispositions to avoid certain stimuli, it’s natural to suggest that functionalism can account for this association by explaining pain in terms of such dispositions. But heeding this suggestion isn’t mandatory. One could instead explain pain in terms of manifest causal relations defined over an actual sequence of states. One can take that line while holding that pain does tend to have the noted dispositions.

Does this suggestion require us to deny that “a system might fail to be conscious if the structures supporting counterfactual transitions are removed or suppressed”, thus risking the lapse into absurdity that Shiller warns against? No, the manifest causal relations between states presumably support counterfactuals, and the suggested view predicts that removing those relations could remove consciousness. What it denies is just that merely manipulating unmanifest dispositional features of a conscious system’s functional organization would leave its consciousness intact. But that result is independently plausible, as shown by Shiller’s bomb thought experiment.

4. Interpretationism

On behalf of functionalists, I’ll now offer a different response to Shiller’s challenge. The response appeals to interpretationism about consciousness, which is a version of functionalism on which individuals have whichever experiences they would be assigned by the interpretation of their functional states that strikes the best balance between simplicity and maximizing the rationality of individuals.

One motivation for interpretationism is that it provides a response to the psychophysical harmony problem: why do experiences tend to provide normative reasons for accompanying thoughts and behavior? For example, why does pain tend to be followed by avoidance behavior rather than seeking behavior? I take this to be a pressing explanatory problem that many existing theories of consciousness lack even a sketch of a solution to. So, I take this to be a fairly powerful motivation for interpretationism.

But interpretationism is also subject to distinctive problems of its own. One is that, although there is a striking correlation between the thoughts and behaviors that our experiences provide normative reasons for and the thoughts and behavior that in fact accompany our experiences, some aspects of our experiences seem not to play any sort of rationalizing role. For example, when a subject experiences one shade of blue that typically doesn’t make them more rational than if they had experienced another shade of blue. How can interpretationism account for such normatively inert phenomenology? Similarly, if subjects whose color experiences are inverted relative to ours would be just as rational as us, why do we have our actual color experiences rather than theirs?

One option for interpretationists, which goes most naturally with dualism, is to allow for stochastic selection among interpretive assignments, weighted by how well such assignments score on the criteria of simplicity and rationality maximization.

What does interpretationism predict about the bystander? Well, it seems that simple, rationalizing interpretations would deem the bystander conscious. After all, their behavior - and not their their non-dispositional internal processing will be just as if there were no bomb—unless they are made aware of their situation, in which case their functional states will presumably lend themselves to interpreting the bystander as having experiences that rationalize their reactions. Either way, interpretationism leads to the prediction that the bystander will be conscious.

5. Should we trust the intuition of bystander consciousness?

I share Shiller’s intuition that the bystander is conscious, notwithstanding the bomb’s interference with the transition probabilities in the bystander’s brain. But where does this intuition come from?

This question is worth asking because many anti-functionalist intuitions are suspect, due to their affinity with the intuition that consciousness couldn’t arise from a three-pound pile of cells. More generally, there is a lack of a priori connections between functional states and phenomenal states: for any given functional organization, we can both conceive of it giving rise to consciousness and of its failing to do so. To break this symmetry, we need empirical considerations. So, armchair pronouncements about which systems are or aren’t conscious due to functional facts call out for justification.

A related pitfall to avoid is that of trying to justify the intuition by holding that since we can tell simply from rational insight that consciousness depends only on internal physical and/or functional factors, the bomb is irrelevant because it’s an external factor. This is a mistake because we can’t tell from rational insight that consciousness depends only on internal factors: for any given internal physical or functional state that an individual is in, we can conceive of that state giving rise to consciousness and we can conceive of it failing to do so. Again, we need empirical symmetry breakers.

Yet I think the intuition that the bystander is conscious can be justified. The justification rests on the common sense observation that our experiences are remarkably resilient to environmental factors that merely affect transition probabilities in our brains without affecting the actual transitions. Think here of sporting events, driving, and combat. If we are not skeptics, then we trust our memories and the reports of others as grounds for taking subjects in such contexts to be conscious. From this, we can conclude that consciousness is robust to environmental interference with transition probabilities that doesn’t affect actual transitions.

I also think that the intuition of bystander consciousness can be supported using fading and dancing qualia arguments, notwithstanding the fact that the intuition is being wielded against functionalism while those arguments are usually offered as support for functionalism. But I’ll leave the exploration of that for another time.

6. Internalizing the problem

As a further wrinkle, Shiller proposes a variant of the problem that brings the bomb threat inside the skull, creating a further challenge for solutions that relied on the bomb being external, including my suggestion that the functional organization that’s relevant to consciousness only features dispositions that just depend on internal features of the system. In the variant, the individual

has had small bombs embedded at their neural synapses. At every millisecond, each bomb has a high chance of going off. If any bomb were to explode, every part of the victim would transition into vapor. By sheer luck, no bomb ever actually detonates.

Again, the intuitive verdict is supposed to be that the individual remains conscious, despite the non-exploding bomb disrupting the transition probabilities and hence the functional organization of the individual’s brain. Again, we should ask about the source of this intuition.

Like the intuition about the external version of the bomb case, the intuition about the internal version garners some support from the robustness of consciousness to environmental factors that influence transition probabilities in the brain. For we would expect some overlap in how environmental threats and internal threats affect those probabilities. So, our evidence that environmental threats are irrelevant to consciousness also provides reason to think the internal threats are irrelevant as well. In addition, internal variation across conscious subjects and within subjects over the course of their lives suggests that consciousness doesn’t depend very sensitively on internal factors that affect transition probabilities.

At the same time, the internal bomb case plausibly involves a kind severe disruption to transition probabilities that is unlike those we are familiar with. This distance from the cases that support the intuition should at least attenuate the intuition’s force. Since the intuition doesn’t wear its empirical basis on its sleeves, I worry that some people may take the intuition at face value, failing to duly discount it for being off-distribution relative to the data that generate it.

7. The scope of the problem

As the title of his paper indicates, Shiller conceives of the bomb cases as a challenge for functionalists. I view the scope of the challenge somewhat differently.

For one, I think bomb cases pose a challenge to various views beyond functionalism. That’s because many non-functionalist views of consciousness take it to have a partially functional basis, which is often enough to bring them within the scope of the challenge.

For instance, consider tracking theories. These theories explain why you experience colors and shapes in the environment in terms of you being in states that track those features. Tracking theorists often appeal to functional features in order to explain why some but not other states that track features of the environment are conscious. Similarly, scientific theories of consciousness such as the global workspace theory come in hybrid versions that combine the functionalist commitments of the scientific theory with a biological-substrate requirement on consciousness.

Even theories - such as dualism and Russellian monism - that do not explicitly harbor a commitment to consciousness depending partly on functional states are compatible with such dependence. Moreover, empirical correlations between functional states and experiences favor versions of such theories that take consciousness to have a partly functional basis.

I also think that the challenge is much more challenging for some versions of functionalism than it is for others. In particular, I see the challenge as a particularly daunting one for non-interpretationist versions of functionalism that take consciousness to depend at least partly on unmanifest dispositions. Likewise, I take the challenge to be particularly daunting for non-functionalist views that take consciousness to depend at least partly on unmanifest dispositions.

8. Implications for evaluating AI systems for consciousness

Suppose that the intuitions that generate the bomb problem are at least somewhat justified and hence that the bomb problem has at least some force. In that case, how does it bear on the evaluation of AI systems for consciousness?

While considering answers to this question, we should keep in mind that the bomb problem is just one datapoint among others that are relevant to AI consciousness evaluations. Accordingly, we should be wary of drawing overly strong conclusions from it. Probably, the best we can reasonably hope for is that it provides some non-decisive evidence for certain hypotheses of interest over others.

---

It might be thought that since the bomb problem provides evidence against functionalism and functionalism is friendly to the prospects for AI consciousness, the bomb problem provides evidence against the possibility of AI consciousness. But this is too quick: for the reasons given above, the problem challenges some non-functionalist views and even supports some forms of functionalism over some non-functionalist views. So, it’s at least not clear that the problem tells against functionalism.

---

Yet the bomb problem does provide evidence against computational functionalism, understood as the view that a system is conscious if and only if the system implements some member of a certain class of computer programs or algorithms. That’s because programs and algorithms characteristically take various inputs. So, whether a system is implementing a consciousness-realizing program or an algorithm will depend not just on its manifest causal activity but also on unmanifest dispositions that settle how it would respond to inputs other than the ones it actually received.

I happen to think computational functionalism is poorly motivated and (partly for reasons given here) less crucial for AI consciousness than it’s often taken to be. But for those who are drawn to computational functionalism and who think its truth matters a lot for whether AI consciousness is possible, the bomb problem probably provides evidence against the possibility of AI consciousness.

---

The bomb problem provides reason to prioritize the development and exploration of theories that tie consciousness to manifest causal activities rather than to dispositions. This observation holds on multiple levels.

First, the space of scientific theories of consciousness is far from fully explored. So, the bomb problem provides reason to seek new theories that don’t appeal to dispositions or which minimize such appeals.

Second, existing theories of consciousness vary in the extent to which they appeal to manifest causal activities vs. to dispositions. For instance, attentional and recurrent processing theories are naturally understood as appealing primarily to manifest causal activities. In contrast, some (though not all) higher-order and global workspace theories are thoroughly dispositionalist in character. The bomb problem provides a point in favor of putting more effort into theories that are closer to the non-dispositionalist end of the spectrum.

Third, theories from the science of consciousness are at an early stage and not currently tailored for application to AI systems. For some of these theories, it’s an open question to what extent they should be spelled out in terms of actual causal relations vs. in terms of dispositions. This is reflected in the division between different versions of global workspace and higher-order theories alluded to above. The bomb problem gives us reason to develop theories in non-dispositionalist directions.

---

Here’s a final connection. The bomb problem offers reason to be a bit more optimistic about the empirical tractability of AI consciousness evaluations. For, as we’ve seen, the problem favors views that account for consciousness in terms of manifest causal activities over views that account for consciousness in terms of dispositions. And manifest causal activities are easier to observe and study than dispositions.³

1. Shiller attributes this suggestion to Colin Klein (2008).

2. Even if this view wouldn’t in letter be a functionalist view according to Shiller’s terminology. I think it’s functionalist enough in spirit to merit the label.

3. For copy editing support and red teaming, I think instances of Claude Opus 4.6 and Claude Sonnet 4.6. The image for this post was generated by Nana Banana.