Artificial intelligence is not conscious – Ted Chiang
An essay by science-fiction author Ted Chiang arguing that large language models like Claude are not conscious has reignited long‑running arguments about what consciousness is and whether machines could ever have it. Commenters debate definitions (subjective experience, embodiment, memory, agency), challenge analogies that reduce LLMs to “just autocomplete,” and draw on thought experiments such as the Chinese Room and philosophical zombies. Many worry that anthropomorphizing chatbots—especially in corporate marketing that hints at “well‑being” or moral status—both muddies public understanding and distracts from more concrete issues like model misuse, bias, and power concentration.
Scope of Disagreement
- Strong disagreement on whether we can confidently say current LLMs are not conscious.
- Many argue we lack a clear, testable definition of consciousness; others say we don’t need a full theory to rule some things out (e.g., rocks, Word documents, current models).
- Several see “consciousness” as a vague, folk concept or a “family resemblance” cluster, not a single well-defined property.
Definitions and Criteria for Consciousness
- Proposed ingredients include: subjective experience (qualia), self-modeling, emotions, desires, embodiment, persistent memory, continuous operation, and agency.
- Pushback: some of these are contested or come from narrow anthropocentric intuitions.
- Several note that many criteria would exclude some humans or animals, which makes them suspect.
LLMs as Predictors vs Minds
- One camp: LLMs are sophisticated autocomplete / “stochastic parrots,” with no inner life; understanding is just statistical pattern-matching over text.
- Others reply that “just prediction” is also a good description of human brains; prediction machinery can in principle implement rich world models, reasoning, maybe even consciousness.
- Comparisons to thought experiments (Chinese Room, pen‑and‑paper execution) are used both to deny and to support computational consciousness.
Embodiment, Time, and Memory
- Many commenters see embodiment, sensory input, and affect (pleasure/pain, fear, desire) as central; without these, talk of “anxious” or “happy” models is seen as misleading.
- Others argue that virtual bodies and synthetic “drives” could suffice, and that humans can be conscious even in diminished states (coma, amnesia, dreams).
- Lack of persistent self-updating and intrinsic passage of time in current LLM deployments is frequently cited as a major gap.
Ethics and Marketing
- Broad concern that anthropomorphic chat UX and “constitutions” encourage people to treat tools as persons (“pseudanthropy”).
- Some argue: if companies really thought models might suffer, present practices would look like slavery or torture.
- Others say, regardless of AI consciousness, how we treat AI systems matters because it shapes our character (virtue ethics) and our treatment of animals and humans.
Intelligence vs Consciousness
- General agreement that LLMs showcase impressive reasoning and language skills, weakening the idea that such abilities require consciousness.
- Several see this as the real philosophical shock: intelligence and consciousness may be more decoupled than previously assumed.