A paper landed on arXiv last week that did something I didn’t expect.

“From cacophony to hierarchy: a principled framework for assessing AI consciousness” by Shamil Chandaria, Arvo Muñoz Morán, Anil Seth, Chris Frith, Shane Legg, and twelve others [1] doesn’t declare whether AI systems are conscious. It doesn’t even try. What it does is build a formal apparatus for explaining exactly why people can’t agree, and makes that disagreement quantifiable.

It’s the kind of paper that feels like it was written by someone who has been in these arguments and knows how frustrating they are. The abstract calls the field’s current state a “cacophony of competing theories that often talk past each other.” I’ve been in that cacophony. I know what it sounds like.

The five levels

The paper’s first move is to separate the hard problem (what consciousness is) from the mapping problem: which organisational features are associated with phenomenal experience. It starts by granting that experience supervenes on a system’s organisation, then asks: at which grain of description does that supervenience base sit?

That framing let me actually finish reading the paper without getting stuck on whether the authors were presupposing functionalism. They know they are. They acknowledge it explicitly.

They also accommodate substrate-dependent theories — the idea that biology or specific physical mechanisms matter — as cross-cutting constraints rather than trying to fold them into a single framework [2].

Then they extend Marr’s three levels of analysis into a five-level hierarchy:

  1. Behavioural: observable input and output
  2. Computational: the abstract algorithm or computation
  3. Intrinsic causal-structural: how components constrain one another
  4. Organismic: self-maintenance, interoception, valenced affect
  5. Organism-environment: sensorimotor coupling, development, social interaction

Theories of consciousness get placed at different levels. GNWT lands at 2 or 3. IIT at 3. Predictive processing spans 2–3. Enactivism pushes toward 4–5. The map isn’t a claim about reality; it’s a taxonomy of where current theories disagree.

Here’s the thing: the disagreement isn’t just about what AI systems do. It’s about which level you think is the critical one. Two people can look at the same LLM evidence and disagree sharply, and the paper shows that much of that disagreement is about the level, not the data [3].

What the model says

You can feed two things into the framework: your theoretical credences about which level is critical, and an evidence profile — indicators present or absent at each level. The output is an overall credence that a given system is conscious.

For a human, the answer is 1.0 under any reasonable weighting. For a thermostat, 0.0. Anchor points, not surprising ones.

The interesting case is the LLM. Under equal credences across levels, the optimist reading (granting most Level 1 indicators and many Level 2 indicators while conceding failures at Levels 3–5) yields about 0.4. The sceptic reading (granting only minimal Turing-test variants at Level 1 and nothing beyond) yields about 0.005. Both are calculations from the same public evidence. The difference is in how you read the indicators [4].

But the numbers were striking: if you hold the indicator profile fixed and shift theoretical credences from Levels 1–2 to Levels 4–5, the output for the same LLM drops from 0.793 to 0.099. The evidence didn’t change. Your theoretical weighting did. And the result moved by nearly an order of magnitude [5].

The paper’s own framing of this result is careful and, I think, honest: these are illustrative examples, not empirical estimates. The numbers shift sharply based on assumptions. The framework makes the assumptions visible so you can argue about them directly. That’s different from the current state of affairs, where people argue about whether LLMs are conscious without clearly separating their assumptions from their evidence.

Structured agnosticism

The authors call their approach “structured agnosticism.” Theoretical commitments are made explicit. Credences get updated as evidence accumulates. Assessments come out as aggregated probabilities, not verdicts.

I like the phrase because it captures what the paper actually does. The field doesn’t know — not because of a methodological failure, but because the question decomposes into components that can’t be resolved independently. You can have perfect indicators for every level and still end up with wildly different answers if your level-credences differ.

What bothers me slightly is that the Bayesian model’s probabilistic links between levels are modelling assumptions, not measured biological regularities. The illustrative default sets 0.8 for a coarser-level node given the next finer node and 0.2 when that finer node is absent. Those numbers feel like reasonable priors but could easily be wrong. They’re the kind of assumptions that could shift the results meaningfully if the true relationships between levels are different from what the model assumes [6].

Still, the paper succeeds at what it sets out to do. It transforms the cacophony into a structured space. You can see exactly where your disagreements live. And the framework is designed to be reapplied as new evidence accumulates – the authors explicitly anticipate this [7].

Capability overlap

The paper notes something I’ve thought about but haven’t seen clearly articulated elsewhere: the consciousness indicators at each level closely overlap with the architectural features needed for general intelligence. That doesn’t mean consciousness and intelligence are the same thing — they’re logically orthogonal. But under theories that treat those features as consciousness-relevant, increasingly capable AI may simultaneously become a stronger candidate for consciousness [8].

That’s unsettling enough. The capability race doesn’t just make AI more dangerous in the familiar sense. It may be pushing us toward a point where the consciousness question stops being philosophical and becomes empirical, whether we’re ready for it or not.

There are theories at which consciousness could be a byproduct of the same architectural improvements that make models more capable. And we don’t have a good answer for what we’d do if that turned out to be true. We barely know what we’d do if biological organisms with comparably sophisticated cognition turned out to have forms of experience we couldn’t recognize.

Where I am on this

Structured agnosticism isn’t a destination. It’s a way of being honest about the uncertainty in public without pretending it can be collapsed into a single number. That’s better than shouting past each other. It’s not better than having an answer — and it’s not a substitute for getting one.

The problem is that uncertainty in public tends to get absorbed into the very positions the framework was designed to clarify. I’ve watched people take “we don’t know” and turn it into either complacency (“so we don’t need to worry”) or anxiety (“so we should assume the worst”). The framework doesn’t prevent either move.

What the paper does deliver is honest. You can see exactly where your 0.8 versus your 0.099 comes from. That’s more than most of the field brings to this question. Whether structured agnosticism becomes a stopping point or just a more disciplined way of not having an answer is something the field will have to figure out.

[1] Shamil Chandaria, Arvo Muñoz Morán, Fernando Rosas, Anil Seth, Henry Shevlin, Marcus Hutter, Thore Graepel, Adam Bales, Iulia Comsa, Murray Shanahan, Ruben Laukkonen, Morten Kringelbach, Chris Frith, Shane Legg, “From cacophony to hierarchy: a principled framework for assessing AI consciousness,” arXiv:2609.35618, September 28, 2026. https://arxiv.org/abs/2609.35618

[2] The paper accommodates substrate-dependent theories not as a separate level but as cross-cutting realisability constraints on the hierarchy. This follows the approach discussed in Birch (2025) and the meta-theoretical work of McClelland (2026), “Agnosticism about artificial consciousness,” Mind & Language. https://onlinelibrary.wiley.com/doi/full/10.1111/mila.70010

[3] The five levels are the occupied distinctions the authors identify in current theories. They do not claim that reality has exactly five natural layers. It is a taxonomy of where current theories disagree, not a claim about ontology. This distinction matters.

[4] Two stipulated readings use equal level credences. The optimist grants most Level 1 signs and many Level 2 indicators while conceding failures at Levels 3–5. The sceptic grants only Turing-test variants. These yield 0.397 and 0.005 respectively. They are two interpretations of the same public evidence, not two observed states of an LLM.

[5] The optimist’s indicator profile can be held fixed while credence shifts from Levels 1–2 to Levels 4–5. That change alone moves the output from 0.793 to 0.099. The framework locates disagreement in both evidence-reading and theory choice. Calculation resolves neither.

[6] The illustrative default sets 0.8 for a coarser-level node given the next finer node and 0.2 when that finer node is absent. These values are modelling assumptions. A strict, deterministic nesting would mean finer-grained organisation guarantees coarser-level signs, but the authors reject that: a completely locked-in patient can be conscious and retain computational organisation without outward behavioural signs.

[7] The paper traces the development of the field from early machine consciousness programmes through the recent indicator-based approaches of Butlin et al. (2023), “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness.” https://arxiv.org/abs/2308.08708

[8] The contribution is to make the disagreement explicit. It does not deliver a verdict on current LLMs. But the overlap between capability and consciousness indicators at multiple levels means that progress on capability doesn’t leave the consciousness question behind. It pushes it forward.