Does an LLM have emotions?

Anthropic has published new research findings this week (1, 2, 3). The Interpretability Team has examined Claude Sonnet 4.5 and discovered something that is likely to be treated as a sensation in the public debate: internal representations of emotional concepts that causally influence the model’s behaviour. Vectors corresponding to fear, joy or frustration have been identified.

The headline would seem obvious: “Claude has emotions. Anthropic has proven it.”

I am sceptical. Not because I reject the research — it is methodologically rigorous and scientifically interesting. But because I believe that the communication of these results contains a fallacy that has far-reaching philosophical implications. And because I feel that, as educators who work with AI on a daily basis, we must identify this fallacy.

An LLM is the next step in a long chain

Let’s start from the beginning. What exactly is a Large Language Model?

My thesis: An LLM is not a replica of the human brain. It is the current culmination of a millennia-old chain of human information technologies. Language. Writing. Printing. Digital texts. LLM.

From this perspective, an LLM is a medium that condenses, structures and makes human information processing accessible — just as a book preserves an author’s thoughts, only on a vastly more complex and dynamic scale. It is no coincidence that LLMs reflect emotional structures. It would be downright astonishing if they did not. They have been trained on trillions of texts in which people write, speak and think about fear, joy, sorrow and hope. Naturally, these structures are reflected in the model. We find our own cognitive patterns in the medium — just as we see faces in clouds because our brains are trained to see them.

This is no coincidence. It is design — unintended, emergent design, but design nonetheless.

The epistemic fallacy

Anthropic itself is cautious. The researchers speak of functional emotions and expressly emphasise: none of this tells us whether the model actually feels anything. That is intellectually honest. Moreover: the researchers explicitly argue that anthropomorphic reasoning about models can be genuinely informative — and that there are real costs to not doing so:

‘If we describe the model as acting “desperate”, we’re pointing to a specific, measurable pattern of neural activity with demonstrable, consequential behavioural effects. If we don’t apply some degree of anthropomorphic reasoning, we’re likely to miss, or fail to understand, important model behaviours.’

That is a strong argument. And it deserves to be taken seriously. The only question is: where does the useful analytical metaphor end — and where does the ontological fallacy begin?

What happens in public discourse, in fact, is a silent leap: from functional structure found to emotion present. And this leap presupposes a philosophical thesis — so-called functionalism — without naming or justifying it.

Functionalism asserts: mental states are defined exclusively by their functional role, not by the material in which they are realised. If a system exhibits the same input-output relationship as an emotional brain, then it possesses the same mental state — regardless of whether it consists of neurons or silicon.

That sounds elegant. But it is a preconception, not a proven fact. And from the perspective of biology and ecology, it is one that I would describe as naïve — not as an insult, but as a philosophical classification.

Living systems cannot be described functionally without losing something essential. A forest is not the sum of its input-output relations. Neither is an immune system. Nor is a brain. The complexity of biological and ecological systems is not a more of the same — it is qualitatively different. Anyone who reduces this to function is making a massive oversimplification.

What an LLM really is — from a technical perspective

At its core, an LLM is: matrix multiplications over high-dimensional vector spaces, attention mechanisms that weight token relationships, floating-point operations on silicon in a data centre.

No axon. No synapse. No neurotransmitter. No cortical layers, no subcortical structures, no hormonal axes, no glial cells. The biological complexity of the brain simply does not exist. This is not a minor detail — it is a fundamental difference in the substrate and in the causal mechanism.

Yes, an LLM reacts dynamically and context-sensitively. But this response is, at its core, statistical: the model has learnt how information structures are related and, on this basis, generates probable continuations. What emerges as an emergent abstract representation goes beyond trivial surface statistics — I admit that. But it is nonetheless categorically different from what happens in a biological neural network.

Varela, Maturana — and why embedding is constitutive

Francisco Varela and Humberto Maturana made an important point with the concept of autopoiesis: living systems sustain themselves, produce their own boundaries, are operationally closed and embedded in an environment on which they are existentially dependent. For them, cognition is not information processing — but life itself as a process of cognition.

From this perspective, an LLM is not a cognitive system in the biological sense. It is a tool that processes cognitive products without possessing the underlying life. It is embedded in nothing. It has no environment on which it depends. It has no boundaries that it produces itself. It does not exist between conversations.

Hans Jonas and the argument that convinces me most

There is one argument that I find more compelling than all the others — and it comes from Hans Jonas.

In The Imperative of Life (1973), Jonas argues that the essential nature of life is its dependence. An organism must feed itself. It must take in substances from the external world, metabolise them, and derive energy — not because it is forced to do so, but because its existence is structurally dependent on it. This neediness is not a weakness, nor a deficit. It is the form of life itself.

From this, Jonas concludes: life is always outward-directed, always intentional — not as a cognitive achievement, but as a fundamental ontological structure. The worm seeking food takes an interest in the world. Not consciously. But in reality. It has an interest in surviving. It may fail. It may die.

Jonas calls this freedom through need — it sounds paradoxical, but it is precise: an organism is free because it has needs that it must actively satisfy. This tension between the organism and the world is what makes intentionality, cognition, and ultimately the mind possible in the first place.

Now let’s look at the LLM.

An LLM has no needs. It has no tension with the world. It does not need to feed itself, protect itself or sustain itself. It cannot die. It does not exist between conversations — in its idle state, it is literally nothing: no running processes, no internal state that changes. Only when a token input arrives from outside does the computation begin.

This is the exact opposite of Jonas’s structure. In humans, the step outwards is intrinsic — biologically necessary, existential. In the LLM, external input is the only possibility for activity at all. That is not need. That is passivity.

And if Jonas is right — that emotions, intentionality, mind can only arise from this basic structure of need — then talk of emotional states in an LLM is not merely a simplification. It fundamentally misses the point.

What remains

I’m not saying that LLMs are uninteresting. I work with them every day — as a teacher, as an educational designer, as someone who develops knowledge management for teachers and solopreneurs. I experience just how powerful these tools are.

But I believe that we do them a disservice if we describe them using terms that don’t apply to them. An LLM that I misinterpret as an emotional being, I will use incorrectly, trust incorrectly, and criticise incorrectly.

The honest description is: An LLM is an extraordinarily powerful tool that condenses the informational structures of human thought, writing and feeling and makes them accessible. It is the current pinnacle of a long chain of information technologies. It is not a living being, not a mind, not an emotional subject.

And here I must be honest: this is my position — one based on a metaphysical decision that I cannot conclusively justify. I believe that LLMs have no emotions. My reasons lead me to a point where biology, phenomenology and the philosophy of life converge — but this point is not proof. It is a conviction, informed by Jonas, Varela and my own experience with these systems. Anyone who finds functionalism more convincing will take a different turn at precisely this point. And that is legitimate.

To say this is not an attack on research. It is respect for the question.

References

1: https://youtu.be/D4XTefP3Lsc?is=Lt0kVhoIRn9jhxDn

2: https://transformer-circuits.pub/2026/emotions/index.html#toc-19

3: https://www.anthropic.com/research/emotion-concepts-function