Summary
Shiller, Duffy, Muñoz Morán, Moret, Percy and Clatterbuck build the Digital
Consciousness Model (DCM), a Bayesian hierarchical model that turns expert
credences about observable indicators into a posterior probability of
phenomenal consciousness under each of 13 stances, then averages the stances.
They run it on a class labelled "2024 LLMs" (examples given: Gemini 2.5 Pro,
GPT-4, Claude 3 Opus), on chickens, humans and ELIZA. From a prior with mean
1/6, the plausibility-weighted median posterior is 0.08 for LLMs, 0.49 for
chickens, 0.85 for humans and 0.006 for ELIZA. The authors say not to read
0.08 as an 8% chance. They endorse only the direction of each update and the
ordering of systems, which they report as stable across the prior settings tested (mean 10%, 16.7%,
50% weak, 50% strong, 90%).
This is the second entry whose object is the attribution of phenomenal
consciousness, after Chandaria et al.,
which shares an author (Muñoz Morán), borrows the DCM's parameter labels and
cites its LLM result. Unlike Chandaria's fabricated activations, the DCM's
inputs are real survey data, but the data are expert opinion about models,
not measurements of them. The entry files concept-less, within the
consciousness-indicators working lens.
Framework
Structure (§3–4). Indicators (206, such as Carbon-Based,
Self-Representations, ARC Performance, Activation Steering Effects,
Metacognition) feed 70 subfeatures, which feed 20 features. Each stance is a
separate model that says which features bear on consciousness and how
strongly. Strength is set by two nine-level labels per link: support
(the positive likelihood ratio) and demandingness (inversely, the
false-positive rate). Both are mapped to Beta parameters with concentration
κ=10 (Shiller et al. 2026,
Tables 2–3). Feature-to-indicator links are stance-independent. All variables
are binary and conditionally independent given their parent. The model is
built in PyMC.
What is measured and what is set. Only two inputs are elicited. For LLMs,
16 experts gave credences that each indicator is present (6 answered every
item, 10 a subset), and the model runs on the per-indicator mean. Thirteen
consciousness experts rated each stance's plausibility from 0 to 10, which
sets the stance weights (Appendix F prints the ratings). Chicken inputs come
from 2 surveys. Human and ELIZA inputs were supplied by the authors. Everything
else is an authorial choice from literature review and consultation: the
stance list, every support and demandingness label, the label-to-Beta
mapping, κ, the binary and independence assumptions, and the prior, which is
the same for every system so that only indicator evidence separates them (§3.5.1).
Stances. Attention Schema, Biological Analogy, Cognitive Complexity,
Computational Analogy, Embodied Agency, Field Mechanisms, Global Workspace,
Higher-Order Thought, Integrated Information, Person-like, Recurrent
Processing (Perceptual), Recurrent Processing (Pure) and Simple Valence. The
authors name free-energy, unlimited-associative-learning, panpsychist,
illusionist, quantum and sensorimotor views as missing. A weighted average
assumes the list is exhaustive, which they concede it is not (§9.4).
Key results
LLMs by stance (§5.2.1). Median posteriors range from 0.02 (Field
Mechanisms, Biological Analogy) to 0.57 (Cognitive Complexity). The evidence
raises the median above the prior on four stances: Cognitive Complexity,
Person-like, Recurrent Processing (Pure) and Simple Valence. It is most
disconfirming on Biological Analogy and Computational Analogy. Chickens score
higher than LLMs on every stance except Cognitive Complexity and Person-like.
Aggregates (§6). Equal weighting gives LLMs 0.08, chickens 0.47, humans
0.85 and ELIZA 0.006. Plausibility weighting barely changes this: 0.08, 0.49,
0.85, 0.006. Table 4 converts these to aggregate likelihood ratios of 28.33
(humans), 4.6 (chickens), 0.43 (LLMs) and 0.05 (ELIZA). The authors' summary:
the evidence is against 2024 LLM consciousness but not decisively, and much
weaker than the evidence against ELIZA.
Sensitivity (§7.1, Appendices D–E). Posteriors move a long way with the
prior. With a mean 0.1 prior for LLMs and ELIZA and 0.9 for chickens and
humans, the printed medians are 0.05, 0.00, 0.97 and 1.00 (Figure 12).
Across those prior settings, the authors report that the direction of update
and the ordering of systems hold, with one exception: Recurrent Processing
(Pure) confirms LLM consciousness under a weak 0.5 prior and disconfirms it
under a confident one. Collapsing the nine-level labels to five keeps the
direction but shrinks the larger updates.
Why it matters
Chandaria et al. cite this result twice: once as a median posterior of 0.08
from a median prior of 0.17 (§7.5), once as roughly 0.08 from roughly 0.17
(§11.7). The posterior matches the primary under both aggregations. The
prior matches as a rounding of 1/6, but the DCM reports it as the mean of a
Beta distribution, not as a median. The authors give the shape only as a 1:5
ratio, so its median cannot be recovered from the text.
On this entry's reading, the stance split is the DCM's most informative output
for the lens. LLMs gain on the stances that read behaviour and capability
(Cognitive Complexity, Person-like) and lose on the stances that need a
particular architecture or substrate (Computational Analogy, GWT, Biological
Analogy). Chickens show the reverse. The authors call these stances cruxes and
cite Seth there
(§8). Chandaria et al. read the same split as their level hierarchy at work.
The DCM derives indicators from theories, as Butlin et al.
2023 do, and
cites both that report and its 2025 update. On this entry's reading, it does
not hold computational functionalism as a working hypothesis the way the 2023
report does: biological-analogy and field stances sit beside the
computational theories as weighted options.
The Person-like stance scores LLMs on whether interacting with a system feels
like interacting with a person. Kirk et
al. measured a related quantity from
the user's side: steering a model raised users' perceived AI consciousness. The authors offer ELIZA's low score on
Person-like as evidence that the stance does not simply reward human-seeming
output (§8). On this entry's reading, ELIZA is a weak test of that, because the
anthropomimesis concern is about systems trained on human output, and ELIZA
was not.
The paper cites none of the wiki's introspection, emotion-vector or workspace findings, and it does
not report how experts scored indicators such as Metacognition,
Self-Representations or Spontaneous Expressions of Valence for LLMs. Filed
findings cannot yet be traced into its numbers.
interpretive tensions
Elicited opinion, not measurement. No model is run or probed. Each
indicator value is the mean credence of up to 16 experts, recruited by
targeted email as experts on system capacities rather than on consciousness. The authors
concede that indicators such as Consistent Preferences are theory-laden, so
an expert's view of machine consciousness may leak into their indicator
answers (§9.6). The posterior aggregates beliefs about LLMs under a stack of
authorial parameters. The paper does not claim more than this. On this
entry's reading, its empirical content is the survey data plus the structure
of the update, not the number.
The model's own checks are circular to a degree. With no ground truth, the
authors validate on whether humans, chickens and ELIZA come out plausibly
(§7.3). The human and ELIZA indicator values were supplied by the authors,
and for humans' 0.85 the authors allow that a low value may reflect missing
indicators or an error in the model or in a stance (§9.2).
Text against text. §8 says LLMs do worst on Embodied Agency and Biological
Analogy. §5.2.1 names Field Mechanisms and Biological Analogy as the lowest
medians and Biological and Computational Analogy as the most disconfirming.
The "2024 LLMs" class includes Gemini 2.5 Pro, which this entry understands
to be a 2025 release; the paper gives no release dates. Table 4 is
labelled approximate. On this entry's arithmetic from the §6.2 posteriors, it
reproduces LLMs (0.435) and humans (28.33) but not chickens (0.49 gives about
4.8, not 4.6) or ELIZA (0.006 gives about 0.03, not 0.05).
Independence cuts both ways. Conditional independence risks
double-counting correlated indicators, while sampling each indicator value
independently risks undercounting capabilities that co-occur. The authors
name both effects but measure neither (§4.5).
concepts
No concept instantiated. The paper's object is the attribution of
phenomenal consciousness from expert-scored indicators, and the wiki has no
concept for that. It reports no behavioural or mechanistic result about any
model.
This is the adjacent situation, matching the Chandaria
entry. Its indicators reach toward
introspection (Metacognition,
Self-Representations) and functional emotional
states (Spontaneous Expressions
of Valence, the Simple Valence stance). Without per-indicator LLM scores,
though, it bears on neither concept's evidence. With two filed entries now
taking consciousness attribution as their object, the question of a concept
for it is recorded here, not decided.
cross-references
- Chandaria et al. 2026: shares an
author and borrows the DCM's support and demandingness labels, converted
more polarised. v3 of the DCM adds a footnote noting that some
implementations reusing its labels state demandingness as a false-positive
rate. It does not name them. - Butlin et al. 2023
and Seth 2024:
both cited. The DCM brackets the two lens anchors by giving each a stance
family and a weight.
sources
- Shiller, D., Duffy, L., Muñoz Morán, A., Moret, A., Percy, C., &
Clatterbuck, H. (2026). Initial results of the Digital Consciousness
Model.
arXiv:2601.17060, v1 22 January 2026; v3 (read) 25 September 2026.