Professional Documents
Culture Documents
The Feasibility of Artificial Consciousness Throug
The Feasibility of Artificial Consciousness Throug
1
Institute of Computer Science, University of Tartu, Estonia
2
Institute of Biology, Humboldt University Berlin, Germany
3
Brain and Mind Center, The University of Sydney, Sydney, Australia
Abstract
Interactions with large language models have led to the suggestion that these models
may be conscious. From the perspective of neuroscience, this position is difficult to
defend. For one, the architecture of large language models is missing key features of
the thalamocortical system that have been linked to conscious awareness in
mammals. Secondly, the inputs to large language models lack the embodied,
embedded information content characteristic of our sensory contact with the world
around us. Finally, while the previous two arguments can be overcome in future AI
systems, the third one might be harder to bridge in the near future. Namely, we
argue that consciousness might depend on having ‘skin in the game’, in that the
existence of the system depends on its actions, which is not true for present-day
artificial intelligence.
Large language models and consciousness
Over the course of documented human history (and probably long before), humans
have speculated about how and why we are conscious. Why is it that we experience
the world around us, as well as our own thoughts, memories and plans? And how
does the organisation of our brain, shaped as it is over evolutionary time [1] and
steeped in social and cultural factors [2], coordinate the activity of its constituent
neurons and glia to allow us to experience anything at all? Recent advances in
neuroimaging have enabled neuroscientists to speculate about how these
mechanisms might arise from the seemingly endless complexity of the nervous
system [3]. In the last few years, a new player has entered the arena – Large
Language Models (LLMs). Through their competence and ability to converse with
us, which in humans is indicative of being awake and conscious, LLMs have forced
us to refine our understanding of what it means to understand, to have agency and
even to be conscious.
LLMs are sophisticated, multi-layer artificial neural networks whose weights are
trained on hundreds of billions of words from natural language conversations
between awake, aware humans. Through text-based queries, users that interact with
LLMs are provided with a fascinating language-based simulation. If you take the
time to use these systems, it is hard not to be swayed by the apparent depth and
quality of the internal machinations in the network. Ask it a question, and it will
provide you with an answer that drips with the kinds of nuance we typically
associate with conscious thought. As a discerning, conscious agent yourself, it’s
tempting to conclude that the genesis of the response arose from a similarly
conscious being – one that thinks, feels, reasons and experiences. Using this type of a
“Turing test” as a benchmark, many users are tempted to conclude that LLMs are
conscious [4], which in turn raises a host of moral quandaries, such as whether it is
ethical to continue to develop LLMs that could be on the precipice of conscious
awareness.
This perspective is often bolstered by the fact that the architecture of LLMs is loosely
inspired by features of brains (Fig. 1) – the only objects to which we can currently
(and confidently) attribute consciousness. However, while early generations of
artificial neural networks were designed as a simplified version of the cerebral cortex
[5], modern LLMs have been highly engineered and fit to purpose in ways that do
not retain deep homology with the known structure of the brain. Indeed, many of
the circuit features that render LLMs computationally powerful have strikingly
different architectures from the systems to which we currently ascribe causal power
in the production and shaping of consciousness in mammals [3]. For instance, most
theories of consciousness would assign a central role in conscious processing to
thalamocortical [6–11] and arousal systems [12–14], both features that are
architecturally lacking in LLMs. It is in principle possible for future LLMs to
approximate the crucial computations of the brain, such as global broadcasting
[15,16] or context-dependent signal augmentation [6,7,17], however at this stage,
these features appear to be unrelated to the remarkable capacities of modern LLMs.
Figure 1 – Macroscopic topological differences between mammalian brains and large language
models. Left – a heuristic map of the major connections between macro-scale brain structures: dark
blue – cerebral cortex; light blue – thalamus; purple – basal ganglia; orange – cerebellum; red =
ascending arousal system (colours in the diagram are recreated in the cartoon within the insert). Right
– a schematic depicting the basic architecture of a large language model.
One might ask why it is so important for the architecture of LLMs to mimic features
of the brain. The primary reason is that the only version of consciousness that we can
currently be absolutely sure of arises from brains embedded within complex bodies.
This could be further collapsed to humans, though many of the systems-level
features thought to be important for subjective consciousness are pervasive across
phylogeny, stretching back to mammals [7,18,19], and even to invertebrates [20]. We
will return to this point, but we will start with the question about precisely what we
mean by the term ‘consciousness’.
There are different meanings associated with the word “consciousness”: neurologists
often refer to levels of consciousness (e.g., the fact that you are [or are not] conscious;
i.e., state of consciousness; Fig. 2), whereas psychologists often interrogate the
contents of consciousness (e.g., consciousness of something; i.e., content of
consciousness) [11,15]. Furthermore, there is a distinction between different contents
of consciousness: our experiences can be described as primarily phenomenal [22]
(e.g., experiential, the sight/smell of an apple; or the feel of your arm) or more
cognitive (i.e., how we access and report conscious experiences [22]; or how we
manipulate abstract ideas and concepts, such as our sense of self, or ephemeral ideas
such as “justice” or “freedom”; Fig. 2).
In DIT, it is proposed that the levels and contents of consciousness arise at the level
of deep pyramidal neurons, which are large excitatory neurons that hold a central
position in both thalamocortical and corticocortical loops [6]. Unlike other theories of
consciousness, DIT focusses on a key physiological characteristic of thick-tufted,
deep (i.e., the bodies of these sit in layer 5 of the cerebral cortex; L5TT), pyramidal
cells – namely, that L5TT have two major compartments (Fig. 3, orange and red
cylinders) that process categorically distinct types of information: the basal
compartment (red) processes externally-grounded information whereas the apical
compartment (orange) processes internally-generated information. At rest, these two
compartments are separated from one another, allowing information to flow
through a network formed by the basal (red) compartments, unencumbered by
modulatory influence of the apical (orange) compartments. Without this modulation,
this processing remains unconscious. Crucially, the thalamus – a deep yet
highly-connected subcortical hub [27] – controls a biochemical switch that allows the
apical network to interact with the basal network, integrating contextually
modulated cortical processing into conscious experience [6,11]. In DIT, the level of
consciousness is determined by whether the process as a whole is active (or silent),
whereas the contents of consciousness relate to whether a particular pyramidal
neuron (or set of neurons) become active, relative to some other set of neurons.
Figure 3 – The neural mechanisms of consciousness. Dendritic Integration Theory (DIT) associates
consciousness with the subset of thick-tufted layer 5 pyramidal neurons (L5TT; right) that are
burst-firing, which occurs when depolarisation of the cell body via basal dendrites (red) coincides
temporally with descending cortical inputs to apical dendrites (orange), particularly in the presence
of gating inputs from higher-order, matrix-type thalamus (blue). Worked example (left): when inputs
hit the retina, they contain information content that will drive feed-forward basal activity in the
ventral visual stream, however only the patterns that are coherent with the information content will
be augmented (e.g., bilateral arrows and simultaneously active apical/basal compartments in ventral
stream). The top-down prediction of these features from frontal or parietal cortex can augment certain
features of the input stream, causing them to stand out from the background, leaving others inactive
(e.g., the light red basal compartment in primary visual area). Note that some predicted features may
not be augmented, and the primary stimulus can suggest potential associations that are not realised
(e.g., light red in basal compartment of a parietal neuron). The thalamus (light blue; dotted line) plays
a crucial role in shaping/gating the contents of consciousness. Note that other areas, such as the
hypothalamus, ascending arousal system and superior colliculus play crucial roles in this process, but
are omitted for clarity.
DIT makes a compelling (and now confirmed) prediction: artificially severing access
to the model (i.e., the top compartment), which is apparently achieved by various
different anaesthetic agents [6,28], should prevent conscious experience, despite
ongoing neural spiking activity in sensorimotor cortical areas. In other words,
activity in sensorimotor areas that leads to conscious perception is in part
determined by only the “right” activity generated in the top-down stream of
information that projects to the sensorimotor cortex. This aligns closely with the
view that conscious perception is a kind of “controlled hallucination” [29], in which
the hallucination in question would correspond to the top-down information that
predicts (and must match to) the bottom-up sensory information stream.
Crucially, our Umwelt defines our modes of experience – it provides the filter that
shapes the information that can impact our phenomenal conscious awareness. James
Gibson referred to this information – that which is pragmatically available to us – as a
set of “affordances” [19,37–39]. Our unique set of affordances has been shaped over
evolutionary time to contain echoes of previously useful shortcuts and tricks that are
directly yoked to our perceptual experience of our own Umwelt, which in turn
directly constrains our phenomenal capacities. For instance, bright lights in the
visible spectrum trigger cascades that innervate the spinothalamic system which we
in turn experience as pain [40], however similar bursts in the infrared or ultraviolet
spectrum fail to evoke the same discomforting sequelae.
If anything at all, what is the Umwelt of an LLM? In other words, what kinds of
affordances does an LLM have access to? By the very nature of its design, an LLM is
only ever presented with binary-coded patterns fed to the network algorithms
inherent within the complex transformer architectures that comprise the inner
workings of the artificial neural networks [41,42]. The Umwelt of an LLM – the
information afforded to it – is written in a completely different language to the
information that bombards our brain from the moment we open our eyes. This
informational stream has a highly non-linear, complex structure that LLMs are quite
obviously capable of parsing. However, crucially, the information stream does not
itself make robust contact with the world as it is in any way. They are infinitely more
different from humans in this regard than we are from bats, and we’re pretty darned
different from bats [43]. Our affordances not only constitute access to direct
experience of the physical world, they also inextricably link us to the consequences
so that we, like bats, pay a price for misinterpreting real-world signals.
While it has been argued that perceptual and cognitive aspects of consciousness are
inextricably linked [44,45], individuals with aphantasia provide an existence
disproof of this argument [46]. Aphantasia is a subtle idiosyncratic perceptual trait
wherein individuals are incapable of voluntarily inducing contents of mental
imagery – for instance, when asked to imagine a purple dinosaur balancing on a
yellow beach-ball with their eyes closed, they simply experience the black colour of
the back of their eyelids (note: author JMS is a card-carrying aphantasic).
Fascinatingly, despite this impaired capacity for mental imagery, aphantasics are
perfectly capable of conducting tasks consistent with cognitive consciousness (e.g.,
spatial working memory) without a modicum of phenomenal experience (e.g., they
don’t experience this phenomenally). Although individuals with aphantasia are
phenomenally conscious when they (unsuccessfully) attempt to imagine an item “in
their mind’s eye” (i.e., they are conscious of the backs of their eyelids), they do not
integrate the cognitive and experiential features of their consciousness into a unified
whole. In other words, they are living proof that phenomenal and cognitive
consciousness are dissociable entities.
Figure 4 – The curious case of aphantasia. When an individual perceives an apple (left), they
experience both the primary features of the apple, as while also anticipating the crunch the apple will
make when they bite into it. In the case of imagination (middle), an individual can perceive a weaker
visual experience of an apple, and also anticipate its crunchiness. In the case of aphantasia (right), the
individual has no phenomenal visual experience of the apple, yet is still capable of anticipating the
crunchiness.
Our main goal is not to convince readers that consciousness can only arise in living
systems, but rather to draw attention to the fact that the common assumption that
consciousness can be captured within computer software might be misleading. For
instance, real biological neurons may have properties that cannot be reinstated
within simulations. In particular, a clear but neglected difference between AI
systems (such as LLMs) and brains is that biological neurons exist as real physical
biological entities, whereas ‘neurons’ in deep neural networks and LLMs are just
pieces of code. One cannot take an artificial neuron out of the software – it exists
only within a simulation. In contrast, when a biological neuron is studied under in
vitro conditions, it still exhibits plasticity and can interact with other neurons
through neurotransmission. Furthermore, inside the cells, there is also no code, but a
further cascade of real physical intracellular complexity. For instance, consider the
Krebs cycle that underlies cellular respiration, a key process in maintaining cellular
homeostasis. Cellular respiration is a crucial biological process that enables cells to
convert the energy stored in organic molecules into a form of energy that can be
utilized by the cell, however this process is not compressible into software. In other
words, capturing cellular respiration in a computer simulation will not keep the cell
alive: processes like cellular respiration need to happen with real physical
molecules.
Another way of stating this is that living systems differ from software and machines
[47,48,51]. Sometimes it is assumed that the difference is mainly in embodiment –
LLMs do not have a body and they do not interact with the world through their
body – but the argument that living systems are different from software and
machines is not so much about embodiment, but rather about having “skin in the
game” [52]. Having skin in the game means, in simple terms, that the organism has
something to lose. An LLM could claim in a conversation that it does not want to be
shut down, but an LLM does not have skin in the game as there is no real
consequence to the software when it is actually shut down. In contrast, in biology the
system has something to lose on several levels [53]. It cannot stop living, as
otherwise it will die. As the philosopher Hans Jonas has said: “The organism has to
keep going, because to be going is its very existence” [54]. If cellular respiration stops, the
cell dies; if cells die, organs fail; if organs fail, the organism will soon die. The system
has skin in the game across levels of processing [53] and these levels are not
independent from each other. There is complex causality between real physical
events at the microscale and consciousness. As of now, there is no scientifically
validated argument for why only living systems should have consciousness (but see
[47-50]), however, given the above, we have to at least entertain the possibility that
consciousness is linked to this complex “skin in the game” process underlying life.
Concluding remarks
Here, we have attempted to provide a systems neuroscientists perspective on LLMs.
We conclude that, while fascinating and alluring, LLMs are not conscious, and also
not likely to be capable as such in their current form. Firstly, we argued that the
topological architecture of LLMs, while highly sophisticated, is sufficiently different
from the neurobiological details of circuits empirically linked to consciousness in
mammals that there is no a priori reason to conclude that they are even capable of
phenomenal conscious awareness (Fig. 1). Secondly, we detailed the vast differences
between the Umwelt of mammals – the ‘slice’ of the external world that they can
perceive – and the Umwelt of LLMs, with the latter being highly impoverished and
limited to keystrokes, rather than the electromagnetic spectrum. Importantly, we
have major reasons to doubt that LLMs would be conscious if we did feed them
visual, auditory or even somatosensory information – both because their
organisation is not consistent with the known mechanisms of consciousness in the
brain, but also for the fact that they don’t have ‘skin in the game’. That is, LLMs are
not biological agents, and hence don’t have any reason to care about the implications
of their actions. In toto, we believe that these three arguments make it extremely
unlikely that LLMs, in their current form, have the capacity for phenomenal
consciousness, but rather represent sophisticated simulacra that echo the signatures
of cognitive aspects of consciousness (Fig. 4), filtered as it is through the language we
use to communicate with one another [21].
The issue of the moral competence of LLMs is already a thorny issue given that they
can in principle produce statements that can aid and abet in malevolent purposes on
the part of the human user. Although this is a large and interesting topic, it is not the
main question we would like to address. Rather, in this essay we want to ask, “but
does it really matter if LLMs are conscious”? If LLMs can match and even exceed our
expectations in terms of getting superficially human-like responses that are useful
and informative, is there any need to speculate about what an LLM experiences?
Some argue forcefully, yes from a moral perspective [56]. According to this view, we
should carefully consider the ethical implications for any conscious entity, including
artificial intelligence, principally because it is assumed that some experiences could
be negative and that in this case the AI could also suffer. At this point, it is claimed,
we should care or at least be cautious about whether an AI really does suffer.
To this extent, we predict that LLMs do not (and will not) have experiences that can
be considered suffering in any sense that should matter to human society. One way
of arguing this point is analogous to the notion of “skin in the game” [52] which
emphasizes the importance of personal investment and engagement in moral
decision-making, and suggests that those who have a personal stake in an issue are
more competent to make ethical judgments than those who do not [52]. Here, we
would argue that not having the capacity for phenomenal consciousness would
preclude suffering and therefore personal investment. This reasoning also extends to
the common disincentives that are used for legal issues that LLMs could become
entangled with such as contracts, libel, etc., which are commonly penalized with
actions ranging from monetary compensation to incarceration. Without personal
investment on the part of the LLM, these disincentives will not be taken seriously by
injured parties and would therefore likely destabilize the rule of law.
Acknowledgements
We would like to thank Jakob Howhy, Kadi Tulver, Christopher Whyte and Gabriel
Wainstein for their helpful comments on the manuscript.
References