Questions for believers in AI consciousness
Debate over whether large language models could be conscious quickly runs into the problem that nobody can define or test consciousness in a way that works for both humans and machines. Commenters argue over whether physics and computability already rule out machine minds, whether internal representations or latent spaces count as an “internal world,” and how much weight to give analogies with animal consciousness. Many conclude the only defensible position is agnosticism for now, though some stress that if AI ever can suffer, questions of rights and moral status will become urgent rather than abstract.
Definitions and Tests of Consciousness
- Many argue the debate is muddled because “consciousness” lacks a precise, agreed definition.
- Suggested tests include:
- Ability to feel drives like hunger, thirst, pain, fear, desire, tiredness and exert agency or delay gratification.
- Dreaming as evidence of a world model and an internal “main thread”.
- Self-recognition (mirror test) and social awareness.
- Behavioral criteria (Turing-test style: if it can convincingly claim consciousness).
- Each test is heavily criticized: edge cases (people without pain/dreams), animals and Sims/NPCs passing, nations or organizations meeting criteria, and the ease of training models to say they “feel” things.
- Several note we cannot even prove other humans are conscious, so expecting a definitive AI test may be unrealistic.
Physics, Computation, and Metaphysics
- One camp insists any consciousness theory must respect thermodynamics, Turing computability, information theory, and current physics; claims of non-computable minds or “magic” are rejected unless empirically supported.
- Others reply these frameworks don’t directly address what consciousness is; they constrain theories but don’t explain subjective experience.
- There is debate over whether quantum randomness or wavefunction collapse could exceed Turing computability; no consensus in the thread.
- Positions span: strict physicalism, bio-essentialism, panpsychism, and dualism; some argue only agnosticism is defensible given current knowledge.
AI Architectures and “Internal Worlds”
- Many distinguish between:
- Internal representational spaces (latent spaces, model state, context) and
- Genuine “internal worlds” with awareness and valenced experience.
- Some say any complex NN/transformer has an internal world in a weak sense; others argue LLMs are static “frozen brains” without continual learning, persistent self, or spontaneous activity.
- A subthread debates distributed inference (multi-GPU, datacenter) and asks: what thing would be the subject of experience in such a system? Opinions diverge and remain unresolved.
- Several stress that current models don’t do anything when unprompted, unlike continuously sensing, updating biological brains.
Ethical Stakes and Practical Relevance
- One side says the question matters because conscious AIs (or ones with valenced states) would have moral weight; harming or enslaving them could be akin to torture or murder.
- Others see AI-consciousness talk as premature or a distraction from real, current harms (to humans/animals) and from concrete AI risks (e.g., agents “pantomiming” panic and causing damage).
- Some adopt a pragmatic stance: unless there’s a clear, testable difference that changes how we must act, the debate risks becoming empty wordplay.
Reactions to the Article and the Discourse
- Several find the article’s choose-your-own-adventure style confusing, “wishy-washy”, and overly focused on verbal distinctions like “internal world” and group minds.
- Others appreciate pressing strong AI-consciousness proponents to clarify what exactly they mean and what kind of “AI” they’re talking about.
- A recurring complaint is that many high-level philosophical claims about consciousness are floated without clear grounding in science or operationalizable criteria.