Summary
Holyoak and Monti treat judgements about consciousness in other entities as
resting partly on analogy from the human case, and ask when such an analogy
carries evidence. Their answer is that it carries most when source and target
share causes of the property, and little when they share only its effects. They run no
experiment. They model the human mechanism as a weighted set of candidate
causes, with terms for unknown causes, for differences that would disable a
shared mechanism, and for a different route altogether. Applied to animals,
the framework gives the familiar gradient from apes to cephalopods. Applied to
current LLMs, it rates the analogy weak: the likeness is behavioural,
behaviour is an effect, and the process that produces it in an LLM is known
and unlike the human one. The case for LLM consciousness then rests on the
prior for an alternative route, which the framework does not set.
This is a framework entry whose object is the attribution of phenomenal
consciousness, like Chandaria et al.,
which it cites. It
sits on the sceptic side of the consciousness-indicators lens beside
Seth, and it
criticises the J-space finding by name. It
files concept-less.
Framework
Evidential and non-evidential analogy. An analogy is evidential when the
causal mechanism in the source is the same, at some level of generality, as
the one in the target. A structural match is not enough. The paper's contrast
is Rutherford's solar-system atom, which is useful for teaching but is no
evidence about atoms because the forces differ, against Newton's terrestrial and celestial motion, which
share a cause. Near analogies are the strong ones, most of all when the causes
are poorly understood even in the source, since a near target is likely to
share the unknown ones too
(Holyoak & Monti 2026).
Soft INUS set. Following Mackie, the cause of an effect is a set of
factors, each an insufficient but necessary part of a set that is itself
unnecessary but sufficient. The authors soften this: factors carry weights and
need not each be strictly necessary. For a target T and property Q,
P(Q_T) = P(M_T)·[1 − P(D_T | M_T)] + P(X_T)
where M_T is a source-like mechanism, D_T a target-specific difference that
disables it, and X_T an alternative mechanism. P(M_T) is a product over
candidate factors of 1 − w_c·(1 − s_c(T)), times a term for unknown causes,
with w_c the causal relevance of factor c and s_c its similarity in the
target. The authors say the formalism is for organising evidence and not for
producing numbers; they give none.
Five questions. Which features of the source plausibly generate
consciousness; how far the target has them; how much remains unknown; whether
the target has disabling differences; whether there is independent evidence
for another route.
Application
The human source. The paper takes the best-supported candidates to be
interactions between ascending arousal systems and the thalamocortical system,
the systems general anaesthesia disables. It then lists dissociations. In
humans consciousness persists without motor output, without language, with
severe loss of executive function, and without the ability to form episodic
memories. Semantic processing, multi-word processing and arithmetic occur
without it. The authors conclude that whether an architecture can perform
these functions carries little causal weight.
Animals. The calibration case. Homology with humans is near complete for
great apes, extensive for other mammals, partial for birds and minimal for
cephalopods, with divergence estimates of about 6.4, 87, 319 and 686 million
years. Analogical support weakens along that gradient and does not depend on
intelligence: the species usually credited with consciousness share sleep and
wake cycles and drug sensitivities, not human-level reasoning.
Current AI. The likenesses between LLMs and humans are behavioural and
linguistic, and the paper assigns them low causal weight. It holds that
artificial neurons resemble biological ones only in that both can be read as
logic gates, and that there is no independent evidence that the unknown
factors in the human set are present. Because the model is multiplicative, a mismatch on one high-weight
factor lowers P(M_T) sharply. Three further points:
- Computational functionalism would remove the neurobiological factors from
the set. The paper replies that sufficiency of any computational property
has not been established independently of the cognitive functions it
supports. Earlier software implementations of global workspace theory were
not taken by their own authors to be conscious. - On the J-space, it says that finding
vectors that approximate words is not evidence of inner experience. Symbolic
AI was built with verbalisable internal symbols and nobody argued it was
conscious. Such results would count only if the capacities they show were
established as causally relevant to consciousness. - The cause of LLM behaviour is known. The systems are trained on human output
to produce human-like behaviour, by a method unlike the biological one. The
authors call the expectation that more intelligence must yield experience an
"intellectualist fallacy", after William James.
What the weak analogy does not show. Failure of the analogy does not
establish absence, though it lowers confidence in a case that rested on it. It
moves the burden to P(X_T), and the authors say
estimates of that will differ widely with priors. They restrict the assessment
to silicon systems in the Turing-machine tradition, and note that systems
built from organoids might reproduce part of the biological pathway. Box 1
lists what stronger evidence would need: computational or architectural
properties shown to play a causal role in consciousness, and intervention
tests of that role.
Why it matters
A reason for cautions the wiki already carries. The
J-space entry notes that its paper takes no
position on phenomenal consciousness, and the
Chandaria entry holds that such
results bear on access and not on experience. This paper gives the caution a
different basis: such structure is a capacity, and a
capacity is evidence only once it is shown to be a cause. The
Eleos commentary on
the J-space paper, which this paper also cites, is filed as a stub.
Where it sits against Chandaria et al. Chandaria's Level 1 discounts
behavioural evidence because LLMs are trained to reproduce the signs that
would count; Holyoak and Monti make that the centre of their argument.
Chandaria lets interpretability results narrow the gap at Level 2. Here they
get little weight unless the capacities are shown to be causally relevant to
consciousness. In
Chandaria's terms the paper withholds credence from Levels 1 and 2 without
asserting that a finer level is the critical one. That placement is this
entry's reading.
Where it differs from Seth. Seth
argues that consciousness probably depends on properties of living systems.
This paper cites that article for substrate as an unresolved uncertainty and
does not adopt its claim. It keeps the evidential question apart from the
ontological one.
Memory. The paper's human dissociations include consciousness without
episodic memory formation. The wiki's entries on continuity across context and
instances concern what persists in a model. By this paper's own rule the human
dissociation is not evidence about models; it is recorded because the essays
that draw on this wiki raise the question.
interpretive tensions
The weights carry the conclusion. The framework does not say how causal
relevance is assessed. The low weight on cognitive function comes from human
dissociations, in which the biological mechanism is intact throughout. They
show that those functions are neither necessary nor sufficient in humans. They
do not show that computational organisation is irrelevant, which is the point
a functionalist disputes.
Logic gates. The claim that artificial and biological neurons
share nothing beyond a logic-gate reading rests on one citation and is
stronger than the rest of the paper needs. The paper elsewhere allows that
physical differences need not mean different mechanisms if the relevant
organisation is preserved.
The bar for counting a mechanism. A computational property counts once it
is established as causally relevant to consciousness. With the causes of
consciousness unknown in the human case, the paper gives no account of how
that would be established for any non-biological system. Box 1 poses it as a
question.
The open term. P(X_T) is where the disagreement about LLMs lives, and the
framework leaves it to priors. The authors say so. What the paper settles is
narrower than its title: the weakness of one route.
No numbers. The right panel of Figure 1 compares targets qualitatively and
is labelled illustrative. The equations are never evaluated.
concepts
No concept instantiated. The paper's object is the attribution of
phenomenal consciousness, for which the wiki has no concept. It is adjacent to
introspection only through the J-space
finding it criticises.
This is the adjacent situation, carried in Cross-references, and the same one
the Chandaria entry declares.
cross-references
- Consciousness-indicators lens (
meta/project-state.md): sceptic side, with
an evidential argument where Seth's is about substrate. - Chandaria et al.: cited in v2 for
explicit tests of human-like and alternative mechanisms. - Cited and not filed here: Dehaene and Naccache's 2026 commentary on the
J-space paper; Butlin et al. 2023; Negro and Mudrik 2025 on analogical
abduction. - GPT-2 workspace tests (Cairns et al.):
the J-space paper's functional tests rerun on GPT-2 small, which passes
four of five. It bears on the same question of what those tests show.
sources
- Holyoak, K. J., & Monti, M. M. (2026).
What Can Analogy Tell Us About Artificial
Consciousness?
arXiv:2610.01002, v1 1 October 2026; v2 (read) 7 November 2026.