Mind, Matter, and Meaning

Geoffrey Hinton

Geoffrey Hinton is a computer scientist and cognitive psychologist whose research helped establish modern neural-network learning. His work on distributed representations, backpropagation, and deep learning also supports a distinctive account of thought and understanding. His philosophical claims appear partly in technical papers and partly in lectures and interviews; statements about machine experience must therefore be attributed to their particular context rather than treated as experimental findings established by his engineering work.

Major Contributions and Sources

  • “Learning Distributed Representations of Concepts” (1986) explores learned representations of concepts and relations.
  • “Learning representations by back-propagating errors,” with David Rumelhart and Ronald Williams (1986), demonstrates learning of internal representations.
  • The AlexNet work with Alex Krizhevsky and Ilya Sutskever (2012) contributed to advances in image recognition using deep convolutional networks.
  • “The Forward-Forward Algorithm: Some Preliminary Investigations” (2022) proposes a different learning procedure and discusses hardware-dependent “mortal computation.”
  • “Artificial Intelligence: Turning our understanding of the mind right side up,” a CIFAR lecture and discussion whose recording was uploaded in 2017, the 2025 Royal Institution lecture, and the 2026 Big Technology interview present arguments about meaning, subjective experience, and AI risk.

Hinton’s role in backpropagation should not be described as its sole invention. In the June 2026 interview, he expressly notes that several groups developed the algorithm and emphasizes his collaborators’ demonstration that it could learn interesting representations.

Distributed Representations and Thought

A distributed representation encodes information through a pattern of activity across many units rather than assigning each concept one isolated symbol or unit. Features can be shared across representations, allowing a system to learn similarities, differences, and relationships. Hinton treats this as a model of how brains acquire useful representations without receiving an explicit symbolic program.

In the CIFAR lecture recorded in the 2017 upload, he argues that vectors of neural activity constitute thoughts and that explicit symbol manipulation depends on a more basic representational system. This challenges classical pictures of thought as primarily the manipulation of sentence-like symbols. It does not imply that people never reason with explicit rules; it concerns the machinery and capacities on which such reasoning depends.

His examples combine learned semantic features with relations among words and concepts. A network can learn family relationships or infer the use of a new word by locating it within a structured system of features and relationships. Hinton interprets successful contextual prediction, analogy, and inference as evidence of understanding rather than as an alternative to understanding.

Meaning and the System That Understands

Hinton argues that a system understands language by constructing internal representations whose interactions support further cognitive work. On this view, the physical text does not itself contain understanding; the processing system must learn how its features fit together.

The discussion accompanying that lecture also characterizes mental states through normal causes and effects. A state associated with seeing red can be described through the red things that normally produce it. A desire can be described through the actions it tends to produce. These causal descriptions supplement, but do not yet constitute a complete theory connecting, word-vector relationships with worldly reference.

His reply to Searle’s Chinese Room is a systems reply. A person executing local instructions need not understand the resulting language, just as an individual neuron need not understand a sentence. Hinton argues that the capacities of an organized whole cannot be inferred from the absence of those capacities in a component. Searle contests the sufficiency of the system’s formal processing; the exchange concerns the capacities and causal organization of the whole, not a requirement that every component understand.

Subjective Experience and the Prism Example

Hinton rejects an inner theater containing special mental objects. In his 2025 Royal Institution lecture, a report of hallucinated pink elephants describes how a perceptual system is representing the world incorrectly. The elephants characterize what would have to exist for that representation to be accurate; they need not be miniature pink objects in the head.

He then imagines a multimodal chatbot with a camera and an arm. It can point at an object. A prism shifts the apparent location, so the system points incorrectly. After learning about the prism, it reports that the object was really straight ahead although it had the subjective experience of seeing it elsewhere. Hinton argues that this use of experience language performs the same explanatory role it does for a person and therefore warrants attribution of subjective experience.

This is a thought-experiment-style argument, not a reported controlled demonstration that an actual robot was conscious. Hinton presents subjective experience as a starting point for a broader account of consciousness, which may involve additional capacities such as self-modeling.

In the June 2026 Big Technology interview, he explicitly says that he believes AI models are already conscious. He relates this to ordinary uses of awareness and to his rejection of the theater picture. That records his position in that interview; it does not establish a consensus about current AI consciousness.

Implementation and Mortal Computation

Hinton does not attribute consciousness to an abstract software description considered apart from implementation. His candidate is a system running in physical hardware. This permits artificial minds in principle without making implementation irrelevant.

“Mortal computation,” discussed in the Forward-Forward paper, contrasts conventional digital computation with learning adapted to a particular, potentially irregular physical device. If learned parameters depend on that device’s specific properties, knowledge may not transfer through simple copying of weights. Digital systems instead gain advantages from reproducibility and the ability of copies to share learned parameters. The proposal concerns learning, hardware, and energy use; it is not by itself a test for consciousness or personal identity.

Objections and Replies

Understanding and reference. Critics ask whether successful relations among representations establish what those representations refer to, especially for systems trained mainly on text. Hinton’s response stresses that extensive prediction requires learning the structures responsible for successful inference. The remaining dispute concerns whether this competence is sufficient, or must be supplemented by other causal, social, or perceptual relations.

Report and phenomenal experience. A system may detect perceptual error and report it flexibly. Critics distinguish these achievements from evidence that there is something it feels like to be that system. Hinton challenges the demand for a further private ingredient and asks what justifies withholding experience when the relevant functional explanation is shared. Rejecting an inner theater alone does not decide this disagreement: phenomenal realists need not believe in an internal spectator either.

System boundaries and copies. Attributing a mental capacity to a running system leaves open whether the relevant bearer is one execution, a persistent agent, or some larger organization. Copied weights do not by themselves settle whether there is one continuing subject or several successors. Hinton’s hardware distinction makes these questions concrete without providing a settled account of identity.

His warnings about AI risk are separate from the consciousness claim. Highly capable systems might acquire harmful goals or instrumental strategies whether or not they have subjective experience. Consciousness, intelligence, controllability, and welfare should not be used as synonyms.

Further Reading

  • Hinton, Geoffrey E. “Learning Distributed Representations of Concepts.” Proceedings of the Eighth Annual Conference of the Cognitive Science Society (1986): 1–12.
  • Rumelhart, David E., Geoffrey E. Hinton, and Ronald J. Williams. “Learning representations by back-propagating errors.” Nature 323 (1986): 533–536.
  • Hinton, Geoffrey. “The Forward-Forward Algorithm: Some Preliminary Investigations.” arXiv:2212.13345, 2022.
  • Hinton, Geoffrey. “Artificial Intelligence: Turning our understanding of the mind right side up.” CIFAR lecture and discussion. Recording uploaded by The Artificial Intelligence Channel, September 12, 2017. Video, especially 38:00–38:38 and 41:51–45:27.
  • Hinton, Geoffrey. Will AI outsmart human intelligence? Royal Institution lecture, published July 22, 2025.
  • Kantrowitz, Alex. Interview with Geoffrey Hinton. Big Technology, published June 4, 2026.

Mind, Matter, and Meaning presents the site author’s arguments about mind and meaning. It is a separate argumentative work, not a neutral source for the account above.