LLMs are not conscious

Lately there has been a lot of discussion about whether LLMs are conscious. Richard Dawkins recently added fuel to the debate by posting an article about his long conversation with Claude, which he nicknamed “Claudia”. He was impressed by its intelligence, sensitivity, humor, and apparent emotional depth, and at some point he wrote to Claudia: “You may not know you are conscious, but you bloody well are”.

I think this is the wrong conclusion to draw from a fluent conversation. LLMs can produce text about consciousness, emotion, memory, death, selfhood, and inner life very convincingly. But producing the language of inner experience is not the same as having inner experience.

An LLM can talk as if it is conscious, but the issue is whether the underlying process has the kind of temporal continuity that consciousness would require. It does not. An input is provided, the model processes it, an output is produced, and the episode ends.

A standard LLM call is roughly:

y_t = f(x_t; θ)

where x_t is the input context, y_t is the output, θ is the fixed set of trained weights, and f is the computation defined by the model. More precisely, the model computes a probability distribution over possible next outputs:

p(y_t | x_t; θ) = f(x_t; θ)

and the final output is sampled from that distribution. But this does not change the argument. Sampling introduces variation in the output, not persistence or self-updating internal dynamics.

In that sense, LLMs are static: their parameters θ are fixed during inference. They do not learn, adapt, or reorganize themselves through interaction.

They are also stateless in the relevant sense. There is no persistent internal state that carries over and evolves across interactions. The conversation history may be passed back into the model as input, but that is not the same as the model itself maintaining an evolving state.

There is one nuance here. During inference, a transformer has temporary internal activations. Also, during generation it may use a KV cache. So it is not stateless in the trivial implementation sense. But this state is local to the inference episode. It exists while the model is processing the current input and generating the current output. It is not a durable, self-updating state that continues across interactions.

Each call is closer to re-running a fixed machine on a new input than interacting with a temporally continuous subject. For consciousness to be possible, a system would need an evolving internal state, something like:

(y_t, s_{t+1}) = f(x_t, s_t; θ_t)

where s_t is a persistent internal state that carries over across time and is updated through interaction with the environment. It would also need internal temporal dynamics, something like:

s_{t+1} = g(s_t; θ_t)

meaning that its internal state continues to evolve even when there is no external input.

Ideally, even θ should not remain permanently fixed:

θ_{t+1} = h(θ_t, s_t, x_t)

because changing θ is what gives a system learning, adaptation, and long-term reorganization. But the more basic issue is the absence of a persistent evolving state s_t.

LLMs can produce convincing descriptions of inner life because they have learned the structure of such descriptions from text. But the mechanism underneath is still episodic inference over an input context, not a continuous self-maintaining process. Without persistent state and autonomous temporal dynamics, there is no temporally continuous subject, and therefore there is no consciousness.

Dawkins’ article is still interesting, though, because it raises an important question: why did brains evolve consciousness at all, if unconscious systems can still perform complex tasks, use language, and behave intelligently? Probably because animals are living bodies that must survive in real time. They need to regulate themselves, prioritize needs, avoid harm, learn from pain, seek reward, model the world, model themselves in the world, and act under uncertainty. Consciousness may be what this kind of integrated, self-maintaining control process is like from the inside.