Agents Ended the Nonsense About Conscious LLMs

A year ago, it was surprisingly fashionable to talk about the alleged consciousness of language models.

ChatGPT said “I,” responded with apparent empathy, and seemed to reflect on itself. That was enough to start a wave of speculation. Maybe there was someone in there. Maybe the model had feelings. Maybe it could suffer. Maybe we were watching a new kind of subject come into existence.

You hear much less of that today.

Not because consciousness research has suddenly produced a final answer. The change happened because everyone started working with agents.

Agents ended the nonsense.

Chat invited projection

A chat is a social form. Someone writes. Someone replies. Language almost automatically creates the impression of another mind.

When the reply is clever, funny, or emotionally perceptive, the projection becomes tempting: surely there must be someone behind those words.

That reaction was psychologically understandable. Philosophically, it was flimsy.

A system can use the sentence “I am afraid” convincingly without being afraid. It can write about pain without anything hurting. It can produce an excellent essay on consciousness without anything becoming conscious for the system itself.

Linguistic competence is not evidence of experience. A sentence about an inner state is not an inner state.

Agents expose the machinery

The moment the same model starts working as an agent, our perspective changes.

The agent has to understand a repository, modify files, call tools, run tests, and verify the result. Suddenly we see more than elegant language. We see how the system actually operates.

It opens the same file three times. It forgets a requirement. It mistakes a failing test for success. It confidently pursues a misunderstood goal. It announces completion while the crucial file remains untouched.

None of this makes agents ridiculous. They are astonishingly capable. But they also look very clearly like what they are: technical systems.

A chatbot invites us to imagine a mind. An agent forces us to inspect the machinery.

That machinery consists of a model, context, tools, memory, loops, permissions, and controls. Remove one component and the behaviour changes radically. Give the system the wrong context and it pursues the wrong goal. Deny it a tool and it cannot complete the task. Remove verification and it happily treats its own nonsense as a result.

There is no mysterious person hiding behind the interface. There is architecture.

Today’s LLMs were not built as conscious systems

I do not consider current LLMs conscious. Not because machines could never be conscious, but because today’s language models lack almost everything that serious theories of consciousness consider relevant.

Michael Graziano’s Attention Schema Theory does not equate consciousness with attention. It proposes an internal model of the system’s own attention. A transformer does contain a mechanism called “attention,” but that is not the same as a persistent self-model through which a system understands and controls its limited attentional processes.

Integrated Information Theory asks about a system’s intrinsic causal structure. Does it exert irreducible causal power upon itself? Does it form an integrated cause–effect structure from its own perspective? More parameters, more fluent language, and better reasoning do not answer that question at all.

Ted Honderich describes consciousness in terms of something being actual for a subject. In perception, a world exists for someone; in thought and feeling, representations and attitudes are genuinely present. An LLM processes linguistic relations. There is no good reason to assume that anything thereby becomes actual for the model.

And Heidegger’s being-in-the-world does not mean possessing a large database of facts about the world. It means inhabiting a practically disclosed world in which things matter, call for action, offer possibilities, and where something is at stake for the being itself. A context window is not a world. A tool call is not yet being-in-the-world.

None of these theories settles the nature of consciousness on its own. But together they make one simple point painfully clear: a system’s ability to talk like a conscious being tells us very little about whether it is one.

Artificial consciousness could probably be engineered

The fact that current LLMs are not conscious does not make artificial consciousness impossible.

We could try to build a system that remains continuously active, models itself and its environment, controls its own attention, maintains autobiographical memory, is embodied or technically embedded in a world, and faces real consequences for changes in its own state.

Such a system would probably require dense recurrent activity, continuous perception, strong internal causal feedback, its own needs or at least homeostatic variables, and a stable temporal perspective on itself.

That would be expensive — not merely financially, but in hardware and energy. A human brain runs continuously while integrating bodily states, perception, memory, attention, movement, and a social environment. Engineering anything comparable would probably require far more than a large text generator that wakes up when prompted.

Perhaps we will build such a system one day. I see no fundamental reason why we could not.

But we will not accidentally create consciousness by making autocomplete bigger.

The debate was not solved. It became more sensible

Agents did not prove that machines can never be conscious. They did something much more useful: they exposed the difference between linguistic effect and technical reality.

Now that we work with agents, we ask less about what the system supposedly feels. We ask what it understands, what it is allowed to do, how it detects mistakes, and who remains responsible for its actions.

That is not a loss of philosophical depth.

It is the end of a remarkably shallow projection.

I discuss the technical side of this shift on Aporeo: Nobody Asks Whether the Agent Is Conscious Anymore.