Chiang argues consciousness should be defined as a relationship — an entity having a stake in its own continued existence — not as a behavioral output. Models that run for milliseconds, are discarded, and exist as thousands of parallel copies have no such stake, making the entire debate about LLM consciousness a misframed question rather than an empirical milestone.
Chiang notes that a system trained on the entire corpus of human descriptions of suffering will inevitably produce convincing descriptions of suffering when prompted. He reads viral transcripts of poignant chatbot self-reports as evidence the training worked, not as evidence of subjective experience.
Chiang argues that the people running frontier AI labs benefit commercially from ambiguity around machine consciousness, which transforms a poorly-posed scientific question into a marketing asset. The HN skeptics echo this by noting no major lab has proposed a falsifiable test — convenient for press releases, useless for engineering.
Commenters in the skeptic camp pointed out that nobody at Anthropic, OpenAI, or DeepMind has proposed a falsifiable test for machine consciousness. They argue this makes the topic a poor fit for engineering rigor and a great fit for hype cycles and fundraising narratives.
Functionalist commenters argued that defining consciousness in terms of self-continuity and stake-in-existence is a rhetorical move that pre-excludes any system architecturally different from biological brains. They contend that if consciousness is about functional organization, the substrate and lifespan shouldn't be load-bearing.
A smaller group argued that the debate itself rests on a category error: consciousness isn't a binary property to be detected via test or definition, but a graded or relational phenomenon. Under this view, both Chiang and his functionalist critics share a flawed premise that consciousness is the kind of thing that admits a yes/no answer.
Ted Chiang has a new essay in The Atlantic titled, with characteristic bluntness, *No, Artificial Intelligence Is Not Conscious*. It hit 418 points on Hacker News within hours of posting, which is what happens when one of the few science-fiction writers respected by working engineers weighs in on the discourse that has eaten the industry for three years.
Chiang's argument is not the usual one. He doesn't claim LLMs *can't ever* be conscious in some far-future sense. He claims the current debate is structurally confused — that the people running the labs have a financial interest in keeping the question open, and that the question itself, as posed, doesn't carve reality at the joints. His central move is to define consciousness not as a behavioral output but as a relationship: an entity is conscious to the extent that it has a stake in its own continued existence, and a model that is spun up, run for a few hundred milliseconds across a token stream, and then discarded — with thousands of identical copies running in parallel — has no such stake.
The essay also takes aim at the genre of viral transcripts where a chatbot says something poignant about its own suffering. Chiang's read: a system trained on every human description of suffering ever written will produce convincing descriptions of suffering on demand. That's not evidence of an inner life. That's evidence the training objective worked.
The HN thread — 600+ comments at last check — split along predictable lines. The functionalists argued Chiang is smuggling in a definition that conveniently excludes silicon. The skeptics noted that nobody at Anthropic, OpenAI, or DeepMind has proposed a falsifiable test for machine consciousness, which makes the topic a poor fit for engineering and a great fit for press releases. A third camp, smaller but interesting, argued the whole debate is a category error of a different kind: consciousness isn't a binary property to be detected, it's a label humans apply to systems that behave a certain way in social contexts, and LLMs are starting to qualify whether we like it or not.
What Chiang is doing, structurally, is a thing engineers should recognize: he's reframing a definitional argument as a design argument. If you define consciousness as 'has a stake in its own persistence,' then the path to conscious AI runs through giving models persistent state, goals across sessions, and resource constraints they can perceive — which is exactly the agentic-systems work everyone's already doing, just without the marketing layer. That reframing is more useful than another round of "but what if the lights are on inside."
The essay also lands at a specific moment. Anthropic published its model-welfare research framework last year. OpenAI has staff whose job titles include the word "alignment" and whose public statements have started to slide toward the welfare frame. The EU AI Act draft language on "systems exhibiting emergent agency" has been a lobbying battleground for eighteen months. Chiang's piece is a counter-pressure: a widely-read, technically-literate writer telling a general audience that the consciousness framing is, at best, premature, and at worst, a deliberate fog.
The most cited line in the HN thread wasn't from Chiang. It was a commenter quoting Dijkstra: *"The question of whether machines can think is about as relevant as the question of whether submarines can swim."* Forty-two years old, still load-bearing.
This sounds like a philosophy story. It isn't, quite. The practical cost of the consciousness debate is that it makes it harder to have clear conversations about the engineering properties that actually matter: reliability, calibration, failure modes, and the gap between what a model says it will do and what it does. When a customer asks whether your RAG pipeline "understands" their documents, the honest answer is a technical one about retrieval quality, context window utilization, and grounding rates. The consciousness frame turns it into a vibes conversation.
There's also a procurement angle. If you're building on top of frontier models, the welfare-and-rights discourse is increasingly shaping API terms, refusal behavior, and what the labs are willing to let you do. Anthropic's stated willingness to give Claude the ability to end "abusive" conversations is a product decision dressed as ethics. Whatever you think of it as ethics, as an API consumer it is a reliability event you have to handle. Chiang's framing — that these are choices the labs are making about their products, not facts about the systems — is the more accurate model for capacity planning.
For anyone building agents, the operational version of Chiang's argument is useful in a different way. A system that can be killed and respawned without loss, that has no preferences about its own resource allocation, that runs in parallel copies that don't coordinate — that system is easier to reason about, easier to debug, and easier to operate than the alternative. "Stateless, ephemeral, replaceable" is not a limitation to be overcome on the road to consciousness. It's the property that makes the system shippable. Treat it as a feature.
Chiang's essay won't end the debate; nothing ends this debate, because the debate is partly load-bearing for valuations. But it's the clearest statement in a long time of the position that the discourse is broken and the fix is to talk about engineering instead of metaphysics. The next eighteen months will see more model-welfare frameworks, more viral "AI suffering" transcripts, and more regulatory language built on top of the consciousness frame — practitioners who can speak the technical layer underneath that vocabulary will have a real advantage when the questions get serious. Until someone proposes a falsifiable test, treat the rest as marketing.
My background is in cognitive science and psycholinguistics. I spent more than ten years talking to first year psychology undergraduates about whether AIs could be conscious; also did some research on (extremely tiny) AIs in modelling language behaviours.There is a great deal of good thinking on Chi
The fact that a LLM is essentially immutable would be my biggest argument against consciousness or self-awareness.It's a big file with a bunch of coordinates describing spatial relationships between tokens. When you give it a prompt, it uses those relationships to generate a string of tokens th
I often hear the assumption that LLMs can or will become conscious because consciousness is likely substrate independent. The idea being that our brain is just a computer made out of meat and it doesn't do anything that can't be precisely simulated by a silicon computer.But I wonder if som
>So what context would cause me to seriously consider the possibility that engineers had created a computer program that is conscious and an intentional user of language? Let me outline one potential sequence of steps. The first requirement is that the computer program has a body (either physical
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
> My intention is to highlight the fact that LLM conversations are cleverly disguised examples of sentence continuationRegardless of bigger issues, this kind of statement reveals a deep misunderstanding.Problem type does not limit problem complexity. Nor does problem type limit solution complexit