Are LLMs conscious? Today’s LLMs are missing some capabilities that we typically associate with consciousness, such as persistent memory¹, continuous updating, or an ability to interact with their environment. Nonetheless, the question of whether an LLM is conscious while performing inference seems valid.

Some people argue that LLMs can’t possibly be conscious, because they only exist in a computer and are not biological systems. This is the substrate dependence assumption. However, those arguments have never resonated with me. As of today, I’m not aware of any fundamental reason for why a neural-network based AI couldn’t in principle be conscious.

A big challenge stems from the fact that consciousness is a poorly defined phenomenon. How could we even measure it in an LLM?

Functional assessment of consciousness

One approach for measuring consciousness is the functional one, which works roughly as follows:

1. Define certain properties that a conscious mind would exhibit, while a non-conscious one would not.
2. Place the mind into situations in which its observable behavior will differ depending on whether it possesses the properties in question or not.
3. If the mind exhibits enough of the behaviors that are associated with the properties in question, you can conclude that it is therefore likely to be conscious.

It should be noted that these types of arguments are inherently probabilistic, and depend on your priors about the space of plausible mechanisms that could produce a given observable behavior. In essence, you’re asking: Is being conscious the most likely explanation for explaining the behaviors we observe from the mind?

Many people have performed these kinds of experiments with LLMs. Implicitly, all of us are, to some degree, assessing the degree of consciousness in an LLM every time we talk to a chatbot.

Different people come to vastly different conclusions from these experiments. The differences are largely due to our individual prior beliefs about what kind of mechanisms plausibly could produce consciousness. In 2022, an engineer at Google famously concluded that their LaMDA LLM was conscious and sentient. Though the most prevalent opinion to this day seems to be that LLMs are not conscious, which is also the position I personally share.

If we assume for a moment that LLMs are not conscious, what makes them so good at appearing conscious, and allows them to pass many of the behavioral tests associated with consciousness?

Crystallized consciousness

I’ll start with a simpler question: Do LLMs feel anxiety?

As with consciousness, you might try to test this by exposing an LLM to a traumatic scenario, and then checking whether its behavior changes in line with what you’d expect from a being that experiences anxiety. Ben Zion et al. (2025) did just that, and found that LLMs exhibit changes in behavior consistent with anxiety.

I argue that the underlying feeling of anxiety, and the experience of consciousness observed here are, in fact, entirely real! LLMs behave consistent with anxiety and consciousness, because anxiety is actually being felt, and conscious experiences are actually being had!

However, this does not mean that these experiences are being had by the LLM, nor necessarily at the time of inference.

I call this crystallized consciousness: The consciousness whose effects we observe is not experienced by the LLM. It is not even experienced at the time where the LLM’s response gets generated. Rather, it was experienced at some point in the past, and by a human being. What the LLM produces is shaped by the collective of millions of past conscious experiences.

The anxiety example makes this more obvious: Clearly, the reason for why the LLM can provide responses consistent with the experience of anxiety is that it was trained on traces (aka writing) produced by humans who were experiencing (or recalling a prior experience of) anxiety. The effects of anxiety on observable behaviors were first documented by humans in text, and subsequently condensed into a behavioral model of anxiety during LLM training – thus, crystallized anxiety.

The same applies to consciousness: The reason LLMs behave consistent with having a conscious experience is because those conscious experiences really did happen! Humans had those very experiences in the past, which influenced their writing, which was subsequently ingested by the models.

To be clear, no single human necessarily had the precise conscious experience that matches a given prompt and its corresponding response produced by an LLM. However, what the LLM uses to generate this response is a (potentially non-trivial) combination of multiple real conscious experiences. In that sense, does this mean that LLMs are in fact capable of generating new conscious experiences after all?

1. Some of these missing properties can be added by a suitable agentic harness. Though it would – in my opinion – be weird to believe that the presence of an outer harness materially changes how the inner LLM experiences its thinking process. Intuitively, it either is conscious while doing reasoning both with and without a harness, or it it’s not conscious in either case. ↩︎