Borrowing from Biology: Why AI Consciousness Research Has a Substrate Problem

This is the second article in a two-part series. The first piece laid out the two methodological frameworks I'll be using here. If you haven’t read it, it’s worth doing so before continuing (see here)

The Hidden Step in AI Consciousness Research

If you read last week’s article, you’ll know that I think there are at least two broad methodological approaches to studying the mind that came out of the cognitive revolution. I called these the cognitive approach, which treats the mind as a functional description of what a system does; and the methodological naturalist approach, which treats theories of the mind as theories of the human brain, and where having a mind is a property of people, not systems.

This week I want to show you what that distinction looks like in practice, using a real example from AI consciousness research.

What the Paper Does

The example is a 2025 paper by Butlin et al., ‘‘Identifying Indicators of Consciousness in AI Systems’’ published in Trends in Cognitive Sciences.¹ It’s a serious paper. The author list includes David Chalmers, Yoshua Bengio, and Eric Schwitzgebel — names that carry real weight in philosophy and AI.

The methodology runs roughly as follows.

There are several leading neuroscientific theories of consciousness. Butlin et al. focus on a few of the most prominent: Global Workspace Theory (GWT), Higher-Order Theories (HOT), and Recurrent Processing Theory (RPT). Each proposes a different account of what consciousness is, or what produces it.

GWT, developed by Bernard Baars and extended by Dehaene and others, proposes that consciousness arises when information is globally broadcast across a “workspace” in the brain — made available to a wide range of cognitive processes simultaneously. HOT, associated with philosophers like David Rosenthal, holds that a mental state is conscious when there is a higher-order representation of it — i.e., when the system represents itself as being in that state. RPT, developed by Victor Lamme, proposes that consciousness is constituted by recurrent processing — feedback loops between different areas of the brain, as opposed to purely feedforward processing.

Butlin et al. take each of these theories and extract from it a set of “functional indicator properties.” These are the properties that, according to the theory, a system must have in order to be conscious.

So for GWT, something like the following question needs to be asked: does the system have a mechanism that integrates and broadly broadcasts information? For HOT, we ask something like: does the system form higher-order representations of its own states? For RPT, we ask: does the system engage in recurrent rather than purely feedforward processing?

The paper then assesses current AI systems against these indicators.

The argument is that the more indicators a system satisfies, the more likely it is to be conscious. To be clear, the paper doesn’t conclude that current AI systems are conscious. But they argue there is no principled reason why future systems couldn’t be.

As I say, this is a careful paper. But there is one step in this methodology that does an enormous amount of work and is never argued for. My aim is to read the paper through the different lenses I laid out in last week’s article — the cognitive approach lens and the methodological naturalist approach lens — to see how much work this hidden assumption really doe.s

Reading It Through the Cognitive Approach Lens

The paper is essentially intended to be read through the lens of the cognitive approach, that is, the dominant approach in AI and cognitive science today.

This view has its origins in the idea that we can understand the brain as a kind of computer —of course, the view no longer models the brain as a digital computer as initially advocated at the beginnings of the cognitive revolution. But the general conceptual framework, what I’ve referred to in other work as a kind of ‘tacit dualism’ is still widely believed and upheld.

In other words, what’s unique to this position conceptually — as residue of the initial computer view — is that an alleged distinction exists in the brain between something like its ‘hardware’ and ‘software’.

This tacit dualism is crystallised on the cognitive approach in the idea there are two levels of description in the brain: the functional level and the physical level.

The functional description says what a system does. The physical description says how that function is realised in matter. It is also assumed that these two levels are, in principle, independent — because the same function can be realised in different physical substrates.

This is known as the principle of multiple realisability.

Granting this much, what matters for having a mind on this view is satisfying the right functional description. The physical realiser — whether biological or not — is secondary.

Now read Butlin et al. again with that in mind.

When they extract “functional indicator properties” from GWT, they’re asking: what is the functional property that GWT says is necessary for consciousness?

And the answer they give is something like global information broadcasting. Not “global information broadcasting in a specific cortical-thalamic network in a human brain.” This question generally gets bracketed or put aside.

But this done subtly, without comment, because within the cognitive approach framework, bracketing the substrate is the obvious and natural thing to do. The functional description is what matters. It is ‘the mind’

The physical realiser is secondary.

That is the hidden step.

It’s not hidden because Butlin et al are being evasive. It’s hidden because, within the framework they’re working in, it’s not a step at all. It’s just obvious.

What follows from this, is the paper’s core conclusion: that an AI system implementing the right functional organisation — global broadcasting, higher-order representations, recurrent processing — is, by that very fact, a legitimate candidate for consciousness.

The conclusion seems to follow naturally. And it does, given the framework.

Reading It Through a Methodological Naturalist Lens

Here is what the same paper looks like if you approach it the way a methodological naturalist does.

On the methodological naturalist approach, theories of the mind are theories of the human brain. There is no separate “functional level” that floats free of physical implementation. The brain is the system being studied, and the theories we build are theories of how that system works — where “that system” means a specific kind of biological organ in a specific organism.

So what are GWT, HOT, and RPT, on this reading?

They are theories about specific properties of biological brains. To flesh out briefly just one example, GWT is a theory about how information is integrated and made globally available in the specific neural architecture of human and animal brains. More specifically, this is a hypothesis about what is necessary for consciousness.

Now, you could argue — and Butlin et al. implicitly do — that what matters in these theories is its functional core, and that the biological implementation is incidental. But that argument needs to be made. Usually, the justification is simply spelling out the approach I did above (i.e., the cognitive approach).

But the real question is: why believe in this approach? That’s the argument we need. And it is exactly the argument that is never made.

At the very least, believing in this framework requires the additional claim that the relevant properties of these theories can be abstracted away from their biological context without loss. And that claim is not obvious.

Take GWT. The global workspace in Dehaene’s version of the theory isn’t just a loose functional metaphor for “information that’s widely available.” It refers to specific neural mechanisms — long-range cortical connections, ignition patterns in frontoparietal networks. When you abstract away from that and say “any system that broadly broadcasts information satisfies the GWT indicator,” you are making a theoretical choice. You’re deciding that the neural specifics are implementation details rather than constitutive features of the theory.

My point is that that decision is the cognitive approach’s decision. It’s not GWT’s decision. GWT is so steeped in this view that it never makes the choice itself.

The methodological naturalist point isn’t that these theories are wrong. It’s that even charitably read, they don’t establish that their functional properties can be abstracted from their biological context and applied to AI systems.

That move is imported from the cognitive approach. And it’s doing all the explanatory work in Butlin et al.

Without it, the methodology doesn’t seamlessly get us to considering AI consciousness claims.

You can’t take a theory developed to explain consciousness in biological organisms, extract a functional property from it, assess an AI against that property, and conclude the AI is a candidate for consciousness — unless you’ve already decided that biological implementation is irrelevant. And deciding that is not a neutral methodological choice. It is a theoretical commitment. One that, as I argued last week, was never established on scientific grounds.

Why It Matters

I want to be clear about what I am and am not claiming here.

I’m not claiming AI systems definitely aren’t conscious. I’m not claiming the neuroscientific theories are wrong. I’m not even claiming Butlin et al. are doing bad philosophy or bad science, given the framework they’re working in.

What I’m claiming is that the framework is doing far more work than the paper acknowledges— than any of these papers tend to acknowledge. The conclusion that AI systems could be conscious doesn’t follow from the neuroscience. It follows from a prior commitment to functionalism — to the view that minds are functional descriptions — that is built into the methodology from the start.

If you approach the same neuroscience from a methodological naturalist perspective, the picture looks very different. Consciousness isn’t a property a system has in virtue of implementing the right functional organisation. It’s a property of organisms with specific kinds of brains. The question of whether AI systems are conscious can’t be settled by showing they implement functional indicators derived from biological theories of consciousness. It requires a different kind of inquiry — one that the field hasn’t even started developing.

But were we to try to begin to have this inquiry, it would need to start with being honest about the theoretical commitments built into existing methodologies.

Butlin et al. is, in many ways, the best version of what the cognitive approach can produce on this question. Careful, rigorous, well-referenced. And yet the step that does all the work — bracketing the biological substrate, treating functional organisation as sufficient for consciousness — is never argued for. It’s assumed.

That’s not a criticism of the authors. It’s a criticism of the framework. And it matters, because that framework is currently shaping how the entire field thinks about AI consciousness.

What about if we take a step back, and ask, as a methodological naturalist, what would we say about this kind of work? The most neutral answer I think, is that independent theories, reasons, and inquiry needs to be done to even begin to seriously address whether AI is conscious, if that is a question we want to ask.

Right now, this simple fact is eclipsed due to the dominance of the cognitive approach.

Photo by Wiki Sinaloa on Unsplash

Thanks for reading! If you found this useful, please consider subscribing. I post every Wednesday on AI consciousness and philosophy of mind.

Want to support this work? Consider a paid subscription. Currently, all my posts are free because accessibility matters to me, but paid subscriptions allow me to keep writing rigorous, informed philosophy that cuts through AI hype. Every paid sub helps

1. Butlin, Patrick, Robert Long, Tim Bayne, Yoshua Bengio, Jonathan Birch, David Chalmers, Axel Constant, et al. 2025. “Identifying Indicators of Consciousness in AI Systems.” Trends in Cognitive Sciences. https://doi.org/10.1016/j.tics.2025.10.011.