doi:10.1017/S0140525X25000032. Published online 21 Apr 2025 (Crossref,
Unpaywall, Europe PMC PMID 40257177). Issue placement is BBS vol. 49, 2026,
e315, per the Cambridge page. Single author, at the Sussex Centre for
Consciousness Science and the CIFAR Program for Brain, Mind, and Consciousness.
The publisher version is paywalled, and Unpaywall reports no open-access copy.
The date is that of PsyArXiv v1 (https://osf.io/preprints/psyarxiv/tz6an_v1,
posted 30 Jun 2024), per the earliest-public-version rule. The current file at
v1 is headed "submitted for review, 08/12/2024".
Version read. PsyArXiv v2 (https://osf.io/preprints/psyarxiv/tz6an_v2,
posted 22 Apr 2025), cached as
cache/papers/source-2024-biological-naturalism-seth.pdf (42 pages) and
converted to .md. This is the BBS accepted manuscript. Its header says it is
not yet copyedited or proofread, and every page carries the Cambridge DOI
footer. Title, author, affiliations and the ten keywords match the Cambridge
landing page (-abs.md). The abstract matches word for word, except that the
manuscript has four more commas: "intelligent, but", "trajectories, but" and
"either is, or convincingly appears to be, conscious". The copyedited version
of record was not read. Table 1, the scenario table, is an image. markitdown
drops it, so it was read from the rendered PDF page 22.
Argument. Seth's case has a negative part and a positive part (§5.0).
The negative part starts with biases. Anthropocentrism, human exceptionalism
and anthropomorphism lead people to conflate intelligence with consciousness,
and a technophilic pull toward a Singularity is offered as a possible fourth
(§2). Seth then argues that conscious AI requires two things:
computational functionalism, and enough substrate flexibility for the
relevant computations to run on silicon (§3.9). He gives several reasons to
doubt the second. Neural-replacement arguments beg the question, and brain
functions are generatively entrenched in metabolism (Godfrey-Smith, Cao).
Brains, if they compute at all, likely run mortal computation, inseparable
from its hardware. Causal and informational closure across scales cannot be
assumed for biological systems. And a simulation realises a property only if
computation suffices for that property, which is the point at issue (§3.3–3.7).
The positive part is biological naturalism, the claim that consciousness is
a property of only, but not necessarily all, living systems (§4.0). Seth
distinguishes it from biopsychism and from carbon chauvinism (§4.5). Two lines
of argument meet. In predictive processing, conscious contents are the brain's
top-down best guesses, a controlled hallucination. Interoceptive inference
serves allostatic control, which ties emotion and selfhood to staying alive
(§4.1). Through the free energy principle, the same minimisation reaches down
into metabolism and autopoiesis, so predictive processing reads as a
dynamical, substrate-dependent process rather than a computation (§4.2). Seth
calls the epistemic form of the view the easiest to defend, and the
ontological form the stronger one (§4.5).
Table 1 sets out five scenarios: naïve "along for the ride", theory-based
computational, substrate-dependent computational, substrate-dependent weak,
and substrate-dependent strong. Its example for the first is LLMs. Seth
leans toward the strong scenario. On that view conscious AI would have to
be living AI, though not necessarily carbon-based. He says this view is
agnostic between a functional reading and an intrinsic-property reading
(§5.5, §5.8). IIT and quantum theories are noted as lying off this continuum
(§5.6). The ethics section (§6) separates real artificial consciousness, which
he calls unlikely but which should not be pursued as a goal, from
conscious-seeming AI. He calls the second almost certain, possibly
cognitively impenetrable like the Müller-Lyer illusion, and a source of its
own harms.
On LLMs. Seth names LLMs as the main driver of the biases, and says
"confabulate" is a better word for them than "hallucinate", which implies
experience (§2). LLMs are his example of disembodied systems that meet an
environment only passively (§5.7). LLMs that talk about their own
consciousness may already give a compelling impression of being conscious
(§6.2). He argues that our biases favour false positives over false
negatives, since LLMs may be "more similar to us in ways that do not matter
for consciousness" (§7). The paper makes no empirical claims about particular
models. One footnote reports the author's own impression of DeepSeek-R1 in
early 2025 as too good to seem humanlike (fn. 45).
How Chandaria et al. 2026
place it. Seth is a co-author of that report, which cites this paper 17
times, more than any other work. The report places Seth's view mainly at
Level 4 (organismic): the Beast Machine and predictive-processing account of
organismic functions. It says the view spans its taxonomy rather than occupying one cell. The weak form
reads as a Level 4 credence plus an indirect realisability constraint. In the
strong form, role and realiser are not separable, so its constraint formalism
only approximates the view. It groups the strong form with Searle's
substrate-dependent biological naturalism, under realisability constraints.
It also cites this paper for the simulation argument, for the claim that LLM
evidence depends on assuming computational functionalism, for
conscious-seeming AI, and as the sceptical pole among the report's authors.
Those characterisations are the report's. Seth's own text keeps the strong
scenario open to a finer-grained functional reading, which a grouping with
Searle's "causal powers" claim does not.
Adjacent, not read. The 50 open peer commentaries and Seth's response,
"The stuff matters: Consciousness, computation, and biology"
(doi:10.1017/S0140525X26106906, BBS 49:e366), are listed on the Cambridge
page. Chandaria et al. cite the response 13 times. Neither was read for this
stub.
This is a stub-only filing, with no finding. It anchors the sceptic,
organismic side of the consciousness-indicators working lens.