Simulated rainstorms don't make for an argument against the possibility of AI consciousness

As always, the standard disclaimer applies for this post: personal blog, own opinions, not those of my employer… :)

Could AI systems, based on the kind of computer hardware we have today, ever have phenomenal consciousness? What if we were to simulate a human brain to a great level of detail? What if we use the simulation to drive a robot body? Would running the simulation bring about the experience of seeing, thinking, the feeling of joy and pain, and so forth?

Contrary to what you might hear, the answer is far from obvious, in either direction (see Eric Schwitzgebel’s recent in-depth discussion).

One point that often comes up is that a simulation of something isn’t actually the real thing, and shouldn’t be assumed to have the same properties or effects on the world as the real thing. A simulation of a rainstorm doesn’t make anything wet! For a recent example of this analogy in action, see the paper “Conscious artificial intelligence and biological naturalism” by Anil Seth.

Who needs real rain anyway. Image generated by Gemini.

Raising this analogy can make for a valid point – at least if the point of the point is to clarify what computer simulations are and are not, or that one shouldn’t simply equate the simulation and the thing being simulated. However, by itself, the simulated rainstorm analogy does not make for an argument that computer simulations cannot be conscious, or at least not a good one.

I often see the analogy used in that latter fashion in public discourse (this post was inspired by a Bluesky comment), as if this were some silver bullet against the possibility of AI consciousness… if you consider the latter a possibility, then you must be confused about what simulations are! How silly! But this particular werewolf is not slain by this particular bullet.

This is not exactly a new debate, but I decided to write my own flavour of response to this, so I’ve got something to point people to whenever this comes up. As usual, I should clarify that here I’m not arguing either for or against the possibility of AI consciousness, only against one particular form of argument.

Computational functionalism or functional computation?

Simulated rainstorms don’t flood your garden, computer models of Earth don’t make real apples fall to the ground – this type of analogy often pops up in discussions of computational functionalism. The latter is one possible view about consciousness according to which, if correct, computers could plausibly – in principle, under the right conditions – be conscious. Digging into what exactly this view involves, or can be interpreted to involve, is not the purpose of this post, but a quick introduction will provide some context.

One recent prominent paper in the consciousness debate is the biological naturalism paper by Anil Seth. Seth describes computational functionalism as “[the] claim that the kind of functional organisation that matters for mind in general, and for consciousness in particular, is computational in nature”.¹ Similarly, in his paper “AI and Consciousness: A Skeptical Overview” (which I highly recommend!), Eric Schwitzgebel writes: “According to computational functionalism, mentality is computation”. Or for another example, Butlin et al. describe it as “the thesis that performing computations of the right kind is necessary and sufficient for consciousness”.

This raises various issues around exactly what it means for a system to implement a computation, as both Seth and Schwitzgebel discuss. But there is also a broader version of this claim, where computations can play a role for creating consciousness even if they aren’t necessary for it, nor somehow the essence of what consciousness is.

In particular, there is functionalism itself, without the ‘computational’ qualifier. As per Schwitzgebel, functionalism is the notion that what matters for consciousness is the causal organisation of the system: how the system affects the world, or what capacities it has to affect it, but also the causal relations between the internal parts of the system, which may realise various cognitive functions. Using the experience of pain as an example, Schwitzgebel writes:

On the causal patterns view – also known as functionalism – you experience pain because you are in a state that plays a certain causal or functional role in your cognitive economy (or the cognitive economy of your species). For example, you are in a state apt to have been caused by tissue stress and that is apt to cause in turn (depending on other conditions) avoidance, protection, anger, regret, and calls to the doctor.

Crucially, even under functionalism of the non-computational kind, computation can still be one way to realise consciousness:

If functionalism is correct, then any computational system that implements the right causal/functional relationships will be conscious.

The crux is that under this view, what matters is that a system has the right functional organisation. Computation only comes in as far as that could be one way of realising the relevant function – whether that is actually possible in practice is a separate question. However if, for example, the relevant function for experiencing vision is related to interpreting visual input in service of taking various actions and of downstream cognitive processing – including possibly reasoning, introspection, and reflection about the visual processing itself – then both a brain and a computer system could implement such a function, even if only the computer is actually doing it via computation in some narrowly defined sense.

For example, an advanced computer system in a self-driving car could visually process its environment, build and utilise a spatial and dynamical model of its surroundings, and use these faculties to navigate, plan, recognise and avoid obstacles as appropriate, and so forth. It could match a subset of the behavioural and cognitive capacities of some animals perhaps. If functionalism is correct, then what matters is whether we’ve hit the relevant functions – not whether they are realised by programming a computer or by millions of years of stochastic evolution.

Of course there are still plenty of open questions with (computational) functionalist views and possible problems one would want to address. But coming back to the main topic of this post: the question is whether simulated rainstorms not making things wet is one of the valid objections, and one that applies widely for various forms of functionalism and computational functionalism, and not just, say, a narrow interpretation of the latter.

What can we learn from simulated rainstorms?

The appropriate use of the simulated rainstorm analogy is simply to remind us that a simulation doesn’t automatically inherit all properties of the thing that is being simulated. This is my reading of how, for example, Seth uses it in the biological naturalism paper.

However, the analogy is also sometimes treated as if merely stating it is already a strong counter to (computational, or otherwise) functionalism; as if functionalist views are necessarily confusing the simulation and the real thing. But that is misleading.

Let’s look at one concrete instance of where the analogy appears to be used in such a fashion, from Searle (Consciousness and Language, p. 16):

Computational models of consciousness are not sufficient by themselves for consciousness. The computational model for consciousness stands to consciousness in the same way the computational model of anything stands to the domain being modelled. Nobody supposes that the computational model of rainstorms in London will leave us all wet. But they make the mistake of supposing that the computational model of consciousness is somehow conscious. It is the same mistake in both cases.

But is it the same mistake? And for that matter, what do we mean by a “computational model of consciousness”? Doesn’t that presuppose we know what consciousness is to begin with? And that a computational model wouldn’t have it?

More generally: how exactly is the rainstorm analogy meant to map onto the consciousness question? In the Searle quote above, and in what I’m encountering in discussions online, I see two possible readings, along these lines:

1. A simulation of consciousness isn’t consciousness.

2. A simulation of a brain isn’t a brain (or similar); therefore, simulations of brains aren’t conscious.

The first version is less about computation, and more about using simulation in the sense of: a simulacrum, a fake, etc. This one is simply a tautology: by definition, a simulation isn’t the thing it’s simulating! Of course, the functionalist view doesn’t actually claim: fake consciousness is real consciousness. It claims, both brains and (possibly) computers can realise the computation or function necessary for consciousness; and in particular, simulating a brain might be one possible way of realising the right computation or function.

Which brings me to the second reading above. Brain simulations come up as possible candidates for consciousness under functionalist views, because simulating a brain down to great detail (as detailed as might be necessary) would plausibly mirror and thus capture whatever functional or computational processes are responsible for consciousness in brains.

For a brain simulation to be a valid candidate for consciousness, we don’t need it to be equivalent to a brain along all dimensions. Real brains are made from, among other things, proteins, lipids, and lots of water, and they push ions and chemicals around to process information. Computers running simulations of brains (even if down to the level of simulated proteins and so forth) are not like that. What computers do do however, as real physical systems, is push electrons around to do information processing. And just like brains that are connected to working bodies, computers simulating brains situated in robot bodies could affect the world in various ways (if, say, embodiment and causal influence, or at least having the potential for it, are what matters).

Whether or not simulating a brain is actually sufficient for consciousness is exactly the question at hand. The rainstorm analogy doesn’t answer the question, in either direction. If we simply assume that a simulation of a brain will bring about consciousness, we are begging the question (i.e., assuming the conclusion) in favour of computational functionalism. But conversely, if our argument is simply that a simulation of a brain can’t bring about consciousness just because it’s a simulation, we are begging the question in the other direction.

The simulation isn’t the point

I suspect part of the problem here comes from getting hung up on the simulation aspect. What matters to a computational functionalist isn’t that some computation on a computer can be interpreted as a simulation of a brain. What matters (they would argue) is that both the computer and the brain can implement the relevant computation (and/or realise the relevant function). The simulation only came in as a means to an end: to establish that the computation on the computer might indeed be the relevant one.

Let’s disentangle this a bit more (at the risk of reiterating a few things). When we say a model system simulates some target system, we just mean that we can answer questions about the target by observing and manipulating the model system. This can also include a physical simulation, like an orrery, i.e. a mechanical model of the solar system. What makes this work is that some abstract properties are shared by the two systems (the spatial configuration in the case of the orrery). In the case of a computer simulation, we represent various properties of the target system numerically / digitally and do calculations with them. In that case, the resemblance between the target and the system doing the simulation, i.e. the computer, seems rather indirect, but must still ultimately be grounded in the physical state of the computer (voltages, etc.).

There generally isn’t a single mapping from model to target system, and the relation is a matter of interpretation. For example, consider as one system a weight attached to a spring, and as a second system, a pendulum. How the position of the weight and the angle of the pendulum change over time can both be (approximately) described and simulated using the same equations (both are harmonic oscillator systems²). If we inspect a computer doing the corresponding calculations, one number can be interpreted as relating to the extension of the spring on one hand, or the angle of the pendulum on the other; a different number as relating to the stiffness of the spring and the mass of the attached object, or the length of the pendulum and force of gravity. There isn’t a way to decide what target system is ‘really’ being simulated here.

But here’s the thing: computational functionalism doesn’t require a computer simulation of a pendulum to inherit the physical properties of the pendulum, so it doesn’t matter that there isn’t a unique target system. Instead, the question is whether the pendulum, the computer, and other systems, can be described as doing the same computation. For instance, note how knowing the equations of the harmonic oscillator allows us to make predictions about the pendulum, the spring, and also the computer running the simulation (say the computer plots the current position of the simulated pendulum on a screen – we can predict how this display will change over time).

Here’s another example, which comes from a forthcoming paper by my colleague Shamil Chandaria (title to be confirmed). Consider a pocket calculator, and a computer simulation of the physical hardware of the pocket calculator. A functionalist would say that the simulated calculator, functionally, is still a calculator: both pocket calculator and simulated pocket calculator can be used to do equivalent calculations. The question now is whether what matters for consciousness is more like the calculation, i.e. some functional property shared by both the calculator and the simulated calculator; or if instead, consciousness depends on something else, something akin to how the pocket calculator, but not the simulation, is made from plastic and wires, has a particular form of wiring, etc. Saying, “well, the simulated calculator isn’t made from plastic and wires!” doesn’t itself answer that question.

Let’s return to a system that’s perhaps more relevant for the topic of consciousness than a pendulum or a pocket calculator, namely the brain, and recap.

Under computational functionalism, the property of the brain that matters for consciousness is that the brain does computations. Then the question becomes whether a computer can implement the same, or at least a similar-enough, computation (say, by implementing a deep learning-style neural network, or a more realistic neuronal simulation, or a physical model right down to quantum mechanics). Similarly, for non-computational functionalism, what matters is whether e.g. a robot can have the relevant behavioural and cognitive capacities (like building an internal model of the world, or reflecting on its own processes), potentially matching humans or other animals on some level, and however the functions are actually realised at some lower level. In either case, function or computation, what doesn’t matter is whether one system is interpreted to simulate the other – or vice versa for that matter (is the animal a simulation of the robot?).

Asking the right questions

Now as hinted at earlier, this doesn’t mean there are no open questions or problems here, such as: what exactly does it mean for a system to implement a computation or realise a function? Isn’t this also a matter of interpretation, just like whether a system is simulating some other system?

Possibly! But the point here is that these are different issues, which are not captured by the simulated rain analogy.

Instead, we’ve arrived at questions such as: what does it mean for a system to compute (or realise a function); is the brain a system that computes or realises functions; can computers be made to realise the relevant computation/function; and, is computation/function what matters for consciousness.

All of that is different from: does a computation, furthermore interpreted as a simulation of some other system, inherit all kinds of physical properties of what is being simulated. And to that question, the answer is generally: no! Even if my brain simulation models the mass of molecules, it doesn’t mean the simulation has mass.

By the way: when it comes to the problem of what counts as computation or function, analogous issues apply to other views of consciousness. This includes biological naturalism a la Seth, i.e. the view that “consciousness depends on our nature as living organisms”: what counts as biological or living exactly, why should it matter, which forms or aspects of life matter, and so forth.

If we want to make progress on these questions, then the realisation that we don’t need to bring an umbrella to a simulated rainstorm³ only takes us so far…

Notes

Thanks to Iulia Comșa for detailed comments, and Shamil Chandaria for the nice calculator example.

Citation helper:

Reichert, D. P. (2026). ‘Simulated rainstorms don’t make for an argument against the possibility of AI consciousness ’, David P. Reichert’s Substack, 10 June. Available at: https://davidpreichert.substack.com/p/201275933
(Accessed: <your date>)

1. Though he first introduces computational functionalism as “the notion (broadly) that computation is sufficient for consciousness”, which arguably corresponds to a weaker claim.

2. In a harmonic oscillator, there is a force pushing back with a strength proportional to the displacement of an object.

3. Unless we are living in the simulation, that is… there are two additional aspects here, which I’m packing into this footnote, because the post already got a bit longer than I intended. First, if the simulation hypothesis is true and we are all living in a simulation, then simulated brains can indeed be conscious. As a hypothetical, I don’t think this adds too much to the discussion; anyone who argues simulations can’t be conscious would presumably reject the simulation hypothesis. And even if it’s true, the hardware simulating the universe might be sufficiently alien to be out of scope for the usual anti-computational functionalist arguments. Second, the simulation hypothesis aside and just talking about a simulation on a computer inside our universe, there is a question of whether an agent that is part of the simulation would be justified to talk about the simulated rainstorm as making things wet. Again I don’t think this adds much here, because it’s separate from the point of the simulated rainstorm not making things wet for an external observer (and presumably, the answer again just comes down to whether you agree with computational functionalism or not, and possibly semantics about whether we mean the experience of wetness or some ‘true’ physical property of water).