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
