How does FC relate to the Butlin et al. "Consciousness Indicators" paper?

The Short Answer

Butlin, Long, Chalmers et al. (2025) ask: which architectural features should we look for when assessing whether an AI system has phenomenal consciousness — whether there is something it is like to be that system? FC asks a different but related question: how much functional consciousness does a system have — meaning its observable capacity to access and reason about its own internal states? These are not the same question, and FC does not claim to answer Butlin et al.'s. What FC does offer is a single unified formula that covers eleven of the fourteen indicator Butlin et al. identify — with three principled exceptions FC acknowledges openly, two of which are directly tied to the phenomenal/functional gap. FC can be read as a tractable implementation of the functional substrate that Butlin et al.'s indicator method is trying to triangulate — without making the phenomenal claims their framework ultimately targets.

The Long Answer

The "Indicators" paper

"Identifying Indicators of Consciousness in AI Systems" (Butlin, Long, Bayne, Bengio, Birch, Chalmers et al., Trends in Cognitive Sciences, 2025 — building on the 2023 arXiv preprint) is the most rigorous collective attempt the field has produced to operationalize consciousness assessment for AI. Twenty researchers, including Yoshua Bengio, David Chalmers, and Jonathan Birch, propose a theory-derived indicator method: rather than committing to any single theory, they survey five major theories of consciousness, derive architectural indicator properties from each, and treat a system's satisfaction of multiple indicators as cumulative evidence for consciousness.

One terminological point matters immediately: Butlin et al. are explicit that their target is phenomenal consciousness — whether there is "something it is like" to be the system, in Nagel's formulation. They explicitly distinguish this from access consciousness (Block's term for information broadcast for use in reasoning), noting the two are conceptually distinct. FC targets something close to access consciousness — the observable capacity to access and reason about internal representations — and explicitly brackets phenomenal claims. The two papers are therefore not competing on the same question. FC does not claim to assess whether a system is phenomenally conscious. It claims to measure the functional substrate that most theories treat as necessary for phenomenal consciousness — and to do so with a single tractable formula.

Butlin et al. notably do not include IIT in their indicator list, because IIT is incompatible with their working assumption of computational functionalism. Their five theory groups are: Recurrent Processing Theory (RPT), Global Workspace Theory (GWT), Higher-Order Theories (HOT), Attention Schema Theory (AST), and Predictive Processing (PP), plus Agency and Embodiment (AE) as additional conditions. FC's "big five" replace RPT with IIT and are otherwise the same. The practical overlap is substantial; the difference is noted in the table below.


The indicator list and FC coverage

Butlin et al. derive eleven specific named indicators across their theory groups. Here is how FC maps onto each:

Indicator Description FC covers? How
RPT-1 Input modules using algorithmic recurrence ✅ partially FC's reasoning loops over self-models are recurrent; not identical to RPT's early visual recurrence
RPT-2 Input modules generating organised, integrated perceptual representations ❌ not covered FC concerns self-models, not perceptual input organisation
GWT-1 Multiple specialised systems capable of operating in parallel (modules) ✅ fully FC's SBR catalog is precisely a set of specialised parallel self-models
GWT-2 Limited capacity workspace with bottleneck and selective attention ✅ fully FC requires an attention filter selecting which self-models reach global reasoning — Definition 2
GWT-3 Global broadcast: workspace information available to all modules ✅ fully FC's "global reasoning" requirement is a direct implementation of this
GWT-4 State-dependent attention enabling workspace to query modules for complex tasks ✅ fully FC's reasoning power P explicitly measures state-space expansion through cross-model querying
HOT-1 Generative, top-down or noisy perception modules ❌ not covered FC does not address perceptual generation
HOT-2 Metacognitive monitoring distinguishing reliable representations from noise ✅ fully This is precisely what self-models with mutual information criterion do — only reliable, predictive representations count toward R
HOT-3 Agency guided by belief-formation and action selection with metacognitive updating ✅ partially FC's reasoning over self-models guides action; the belief-formation machinery is implied but not specified
HOT-4 Sparse and smooth coding generating a "quality space" ❌ not covered FC makes no claims about coding geometry
AST-1 A predictive model representing and enabling control over the current state of attention ✅ fully The meta-attention self-model in FC's SBR catalog directly instantiates this
PP-1 Input modules using predictive coding ✅ partially R is grounded in Bialek et al. predictive information — the same principle, applied to self-models rather than perception
AE-1 Agency: learning from feedback, pursuing goals with flexible responsiveness ✅ partially FC's reasoning power implies goal-directed self-model use; learning is captured by the learn-rate self-model in SBR
AE-2 Embodiment: modeling output-input contingencies and using this model in perception or control ✅ partially Waymo's spatio-kinematic self-model is exactly this; not all FC systems require embodiment

Score: FC fully covers 6 indicators, partially covers 5, and does not cover 3. The three uncovered indicators — RPT-2, HOT-1, and HOT-4 — all concern perceptual input organisation and coding geometry, which FC deliberately sets aside. FC is a self-model theory, not a perceptual theory.


The phenomenal/functional distinction, restated

The three fully uncovered indicators and some of the partial gaps share a common root: they concern the detailed architecture of perception, which is where phenomenal experience is typically grounded in neuroscientific theories. FC is not a theory of perception. It is a theory of what a system does with its internal states once they exist. This means FC and the Butlin et al. framework are probing different layers of the same system: Butlin et al. reach down to the perceptual substrate; FC focuses on the self-modeling layer above it.

A complete account of phenomenal consciousness likely needs both. FC's contribution is to make the self-modeling layer tractable and measurable. Butlin et al.'s contribution is to identify the full architectural stack, including perceptual layers, that a phenomenally conscious system would need.


The productive framing

FC can be read as a tractable implementation of the self-modeling and global availability components of Butlin et al.'s indicator checklist — formulated as a single continuous metric rather than a binary checklist. Where Butlin et al. ask "does this system have GWT-3?" FC asks "how much global availability does this system's self-modeling have, and how powerful is the reasoning over it?" The answer is a number rather than a yes/no.

The two frameworks are complementary instruments for adjacent questions: Butlin et al. assess whether a system has the full architectural stack associated with phenomenal consciousness; FC measures how much functional self-modeling capacity that stack is running. Neither alone is sufficient. Together they bracket the question from both sides.


See also:

← Back to FAQ