Source-linked AI summary
The feasibility of artificial consciousness through the lens of neuroscience
Jaan Aru, Matthew Larkum, James M. Shine
TL;DR
The paper asks whether LLMs are, or may soon become, conscious, given that conversational interactions can invite that inference. It assesses the question through neuroscience arguments about sensory grounding, neural architecture, and biological organization, concluding that current LLMs are extremely unlikely to have phenomenal consciousness.
Problem
LLM interactions have prompted renewed questions about attributing consciousness from interactions with other agents.
Source: Large language models and consciousness · p. 2, What is consciousness? · p. 4Method
The paper evaluates LLM consciousness through neuroscience arguments about sensory streams, neural architecture, and biological organization.
Source: Large language models and consciousness · p. 4, The neural architecture supporting… · p. 9, Concluding remarks · p. 14Results
The authors conclude that current LLMs are extremely unlikely to have phenomenal consciousness and instead mimic it.
Source: Concluding remarks · p. 14Takeaways & Limitations
The paper concludes that LLMs are not conscious, while emphasizing the need to refine understanding of similarities and differences between living systems and AI.
Source: Concluding remarks · p. 14The argument does not claim that mammalian brains are the only possible basis for consciousness; its comparison concerns present-day LLMs.
Source: The neural architecture supporting… · p. 9
Abstract
from arXiv · showhide
Interactions with large language models have led to the suggestion that these models may soon be conscious. From the perspective of neuroscience, this position is difficult to defend. For one, the inputs to large language models lack the embodied, embedded information content characteristic of our sensory contact with the world around us. Secondly, the architecture of large language models is missing key features of the thalamocortical system that have been linked to conscious awareness in mammals. Finally, the evolutionary and developmental trajectories that led to the emergence of living conscious organisms arguably have no parallels in artificial systems as envisioned today. The existence of living organisms depends on their actions, and their survival is intricately linked to multi-level cellular, inter-cellular, and organismal processes culminating in agency and consciousness.
Large language models and consciousness
The conversational fluency of LLMs raises questions about consciousness, but the paper argues that present-day systems differ from living conscious organisms in sensory grounding, neural architecture, and biological organization. It frames these differences as reasons to doubt that current AI systems have, or will soon have, phenomenal consciousness.
Large language models and consciousness
- When asked a question, a system can give an answer with nuance typically associated with conscious thought, making it tempting to infer that its response arose from a similarly conscious being.Source: Large language models and consciousness · p. 2
- Although early artificial neural networks simplified the cerebral cortex, modern LLMs are highly engineered and lack deep homology with known brain structure.The paper notes that some computationally powerful LLM circuit features differ from systems implicated in mammalian consciousness.Source: Large language models and consciousness · p. 3
- LLMs lack thalamocortical and arousal systems that many theories assign central roles in conscious processing.The paper contrasts those systems with LLM circuit architectures.Source: Large language models and consciousness · p. 3
- Consciousness is tied to sensory streams meaningful to an organism, unlike the text and keystrokes that provide LLM inputs.The paper presents this as its first argument against attributing phenomenal consciousness to present-day AI.Source: Large language models and consciousness · p. 4
- The paper argues that consciousness may depend on complex biological organization across living systems, which it says is strikingly absent in AI systems.It presents this as a third argument against current AI systems having phenomenal consciousness.Source: Large language models and consciousness · p. 4
What is consciousness?
The paper distinguishes being conscious from the contents of consciousness and separates phenomenal experience from cognitive access and reporting. It focuses primarily on phenomenal consciousness, arguing that conversational fluency alone is not objective evidence of conscious agency.
What is consciousness?
- LLM conversations are remarkable, but they are not formal objective measures of consciousness or prima facie evidence of conscious agency.The paper says their arrival has prompted reassessment of whether interaction can support inferences about other agents' consciousness.Source: What is consciousness? · p. 4
- Neurologists often discuss levels of consciousness, while psychologists often examine its contents.The passage describes a level as whether a person is conscious and content as consciousness of something.Source: What is consciousness? · p. 5
- Conscious contents may be phenomenal experiences or cognitive processes involving access, reporting, and abstract thought.Examples include seeing or smelling an apple and manipulating concepts such as justice or freedom.Source: What is consciousness? · p. 5
- The paper focuses mainly on phenomenal consciousness and asks whether machines can experience the world phenomenally.It distinguishes this focus from other possible aspects of machine consciousness.Source: What is consciousness? · p. 5
The Umwelt of an LLM
The paper argues that an LLM’s text-based inputs offer little robust contact with the natural world, but that richer inputs alone would not establish consciousness. Conscious experience also presupposes an agent with a subjective perspective, internal properties, motivations, and goals.
The Umwelt of an LLM
- LLMs receive highly abstract binary-coded inputs that do not make robust contact with the world, unlike the dynamic sensory information available to organisms.The paper compares the information stream to one more different from humans’ than humans’ is from bats.Source: The Umwelt of an LLM · p. 6
- Future AI could receive richer inputs that better match natural-world statistics than the narrow electromagnetic slice available to humans.The paper leaves open whether future systems’ Umwelt could extend beyond the human one.Source: The Umwelt of an LLM · p. 6
- An organism’s Umwelt and conscious experience are not determined by sensory input alone; having an Umwelt presupposes an agent with a subjective perspective.Affordances also depend on an agent’s internal properties, motivations, and goals.Source: The Umwelt of an LLM · p. 6
- Simply adding massive data streams to future AI systems will not lead to consciousness, because consciousness does not arise merely from data.The discussion emphasizes that agency and a subjective perspective are part of the issue, not just the amount of information supplied.Source: The Umwelt of an LLM · p. 6
The neural architecture supporting conscious integration
Neuroscience theories link conscious integration to different neural mechanisms, including thalamocortical and cellular architectures that current LLMs lack. The authors therefore hesitate to ascribe phenomenal consciousness to present-day LLMs, while allowing that artificial consciousness may be possible in principle at greater computational specificity.
The neural architecture supporting conscious integration
- Global Neuronal Workspace Theory proposes that consciousness depends on a distributed frontoparietal central workspace that integrates information from local cortical processors and globally broadcasts it, distinguishing conscious from non-conscious processes.Source: The neural architecture supporting… · pp. 7–8
- Dendritic Integration Theory proposes that global conscious integration also depends on local integration in single layer 5 neurons, whose basal compartment processes externally grounded information.Source: The neural architecture supporting… · p. 9
- Current LLMs lack features hypothesized to support consciousness, including dual-compartment pyramidal neurons, centralized thalamic architecture, global workspace, and ascending arousal system.The authors consequently hesitate to ascribe phenomenal consciousness to present-day LLMs, which they describe as topologically extremely simple in comparison.Source: The neural architecture supporting… · p. 9
- The authors consider consciousness implementable in principle, but potentially requiring computational specificity beyond present-day and perhaps future AI systems.They do not claim that the mammalian brain is the only architecture capable of conscious awareness.Source: The neural architecture supporting… · pp. 9–10
Consciousness as a complex biological process
The section argues that consciousness may depend on biological organization across levels that current computational accounts and AI abstractions do not capture, while emphasizing that the relevant mechanisms remain incompletely understood.
Consciousness as a complex biological process
- Conscious processing involves integration between pyramidal-cell compartments across the thalamocortical system, while anesthetics can decouple the compartments and disrupt dendritic integration.The account links this coupling to top-down feedback and suggests higher-order thalamic control.Source: Consciousness as a complex biological… · pp. 10–11
- Current computational accounts may miss key mechanisms of consciousness because biological computations are incompletely understood and existing tools may be inadequate.The authors suggest current explanations may not capture the complexity of biological processes.Source: Consciousness as a complex biological… · p. 12
- Living systems’ organization across levels may be essential to consciousness, which may not be abstractable from the biological machinery underlying life.The authors caution that this does not establish that consciousness cannot be captured in software.Source: Consciousness as a complex biological… · pp. 12–13
- Present-day AI can model neural-circuit and network computations but does not simulate neuronal processes, and it remains uncertain whether consciousness depends on interactions at the neuronal level.The authors note that treating neurons as the relevant level is an assumption whose correctness is unknown.Source: Consciousness as a complex biological… · p. 12
- Cellular respiration depends on physical molecules and cannot be compressed into software; the authors use it to illustrate why consciousness might not be separable from underlying machinery.They explicitly do not claim that consciousness requires the Krebs cycle.Source: Consciousness as a complex biological… · p. 13
- If consciousness depends on processing across biological levels, AI may reproduce its signatures without conscious experience, leaving conscious machines potentially far off.The authors suggest that network-level mimicry may omit other levels and their coherent interactions.Source: Consciousness as a complex biological… · pp. 13–14
Concluding remarks
The authors argue that current LLMs are extremely unlikely to have phenomenal consciousness, citing differences in sensory input, architecture, and biological organization. They conclude that architectural comparison could benefit machine learning and neuroscience.
Concluding remarks
- Mammalian Umwelt encompasses perceptible aspects of the external world, whereas LLM inputs are limited to keystrokes rather than the electromagnetic spectrum.The authors describe LLM input as highly impoverished relative to mammalian sensory contact.Source: Concluding remarks · p. 14
- LLM architecture differs from neurobiological circuits empirically linked to consciousness in mammals, giving no a priori reason to infer that LLMs can be conscious.The authors identify this as a second argument against consciousness in current LLMs.Source: Concluding remarks · p. 14
- The authors consider it extremely unlikely that current LLMs have phenomenal consciousness, arguing that they mimic signatures embedded in human language.They also conclude that LLMs are not conscious in their current form.Source: Concluding remarks · p. 14
- Comparing LLM and mature-brain architectures may create opportunities to advance both machine learning and neuroscience.The authors point to brain-organization features for machine learning and simple distributed systems for neuroscience.Source: Concluding remarks · pp. 14–15
Box 1: LLMs and skin-in-the-game - do we have a moral quandary?
The box considers whether LLMs’ apparent human-like responses imply experiences that matter morally. The authors argue that lacking phenomenal consciousness would preclude suffering and personal investment, unlike living systems whose survival depends on processes across levels.
Box 1: LLMs and skin-in-the-game - do we have a moral quandary?
- The moral question is whether useful, human-like LLM responses warrant speculation about what an LLM experiences.The box contrasts that question with arguments for caution about possible AI suffering.Source: Box 1: LLMs and skin-in-the-game - do… · p. 15
- An LLM may say it does not want to be shut down, but software faces no real consequence when shut down, unlike biological systems with something to lose.The authors describe living systems as having stakes across levels of processing.Source: Box 1: LLMs and skin-in-the-game - do… · pp. 15–16
- The authors argue that lacking phenomenal consciousness would preclude LLM suffering and personal investment.They connect the absence of personal investment to the limits of applying common legal disincentives to LLMs.Source: Box 1: LLMs and skin-in-the-game - do… · p. 16
- The authors suggest that applying legal disincentives such as compensation or incarceration to LLMs without personal investment could destabilize the rule of law.This is presented as a likely consequence, not a demonstrated outcome.Source: Box 1: LLMs and skin-in-the-game - do… · p. 16