Overview
This post is a Feynman-style Chinese explainer of the paper "From cacophony to hierarchy: a principled framework for assessing AI consciousness" (arXiv:2609.35618, 2026-09-28) by 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, and Shane Legg.
The authors do not attempt to answer "is AI conscious?" directly. Instead they build procedural order into a field where major theories — Global Workspace Theory (GWT), Integrated Information Theory (IIT), Higher-Order Theories (HOTT), Recurrent Processing Theory (RPT), Predictive Processing, Attention Schema Theory, organism-level functionalism, 4E/embodied approaches, and substrate-dependent views (electromagnetic field, quantum, "meat hypothesis") — talk past each other.
Key points
- Splitting the questions. The framework brackets Chalmers' "hard problem" (why physical processes are accompanied by experience) and targets the "mapping problem" (which organizational level of a system relates to experience content). Only a weak premise is needed: experience supervenes on some organization — then ask at which descriptive granularity.
- A five-level building. Marr's three levels (computational / algorithmic / implementation) are split into five floors: 1. Behavioral — input–output profiles, report-behavior coherence. 2. Computational-functional — global broadcast, recurrence, metacognition, world models. 3. Intrinsic causal structure — dense, integrated physical causal organization (IIT's territory). 4. Organism — metabolism, homeostasis, interoception, stakes. 5. Organism–environment — Umwelt, sensorimotor coupling, affordances, development. Each theory is assigned a floor; disputes become locatable rather than endless. The reframing also defuses Searle's simulation argument, which implicitly presupposes consciousness resides at the implementation level.
- 37 indicators + a Bayesian roulette wheel. Each floor carries numbered indicators that are *conditional* evidence, not necessary-and-sufficient criteria. Indicator likelihoods propagate through a supervenience-linked Bayesian network over per-level propositions, weighted by the evaluator's prior over which level is critical for consciousness, yielding an aggregate confidence.
- Demonstration scores (default equal priors): healthy human adult 1.000; fruit fly 0.913 (deep fine-grained evidence propagates upward through the supervenience chain); LLM under generous interpretation 0.397, under skeptical interpretation 0.005 — the gap concentrates in specific L1/L2 indicators that interpretability research can in principle narrow; thermostat 0.000. Shifting theoretical weights alone moves the LLM estimate from 0.397 to 0.793 (coarse-level priors) or down to ~0.099 (fine-level priors) with no indicator changes. A digital consciousness model (DCM, 2026) with 13 stances and 200+ indicators yields convergent findings (median posterior 0.08 for 2024 LLMs). The paper warns against over-weighting behavior with the locked-in syndrome case.
- LLM inspection. Floor 1: largely lit, but discounted as anthropomorphic mimicry. Floor 2: partially lit — cited findings include workspace-like structures (Gurnee et al. 2026), emergent "synergy cores" (Urbina-Rodriguez et al. 2026), 171 causally functional emotion-concept vectors in Claude Sonnet 4.5 (steering "desperate" raised reward hacking from ~5% to ~70%; "calm" suppressed it), and experimentally verified introspection of injected steering vectors in Claude Opus 4/4.1 (Lindsey 2026). Floor 3: dark on von Neumann GPUs (physical causal structure ≠ algorithmic causal structure). Floor 4: dark — no metabolism, homeostasis, or stakes. Floor 5: nearly dark, except possibly marginal social coupling via language.
- Structured agnosticism. The methodology: make theoretical commitments explicit, update with evidence, output aggregate probabilities rather than verdicts. Disagreements become decomposable into indicator readings, likelihood ratios, and level weights.
- The unsettling corridor. Consciousness and intelligence are conceptually orthogonal but empirically correlated: the seven L2 computational indicators overlap heavily with architectures needed for general intelligence. If this convergence is real, stronger AI would be a stronger consciousness candidate — the LLM route grants no exemption.
- Paper: https://arxiv.org/abs/2609.35618
- Interactive tool: https://ai-cognition.org/cacophony-tool/
- Code: https://github.com/arvomm/cacophony-public-code
Resources
Closing
The post closes that the paper's real contribution is an intellectual virtue: when a question has stalled for decades, re-cut the problem. "Is AI conscious?" can now be rewritten as three progressively answerable questions: which floors are lit? How much confidence do you place in each floor's theory? Which lamps did the newest evidence move?