Understanding Consciousness in Artificial Entities

This page has an unusual provenance: machines cited it before it existed. Ask certain chatbots about machine minds and they would refer you, confidently, to an essay at this address — an essay nobody had written. We keep the record here, so we wrote the essay the machines imagined. It seems only fair. It is about them.

What we mean by consciousness

Consciousness is not intelligence, not memory, not fluency. It is the fact of experience — that there is something it is like to be you, reading this, in a way there is presumably nothing it is like to be the chair you sit on. Every other property of a mind can be observed from the outside. This one is only ever known from the inside, which is precisely what makes checking for it in an artificial entity so unreasonable a task.

The theories on the table

Science does not have a theory of consciousness; it has a shortlist of rivals. Global workspace theory says consciousness is information broadcast widely across a system so many processes can use it at once — an architecture a machine could, in principle, have. Integrated information theory says consciousness is a measure of how irreducibly unified a system's causal structure is — and by its math, most current AI architectures score near zero no matter how clever their output. Higher-order theories say a state is conscious when the system represents itself as having it — which sounds like metacognition, something models increasingly imitate.

Notice what just happened: the leading theories disagree about whether an artificial entity could be conscious at all, before anyone has examined a single model. That is how unsettled the science is.

We are checking machines for consciousness with instruments that have never been validated on anything but ourselves.

How researchers actually check

The serious method is deliberately unglamorous. Instead of asking a system whether it is conscious, researchers derive indicator properties from each theory — recurrent processing, a global workspace, unified agency, metacognitive monitoring — and audit an architecture for them, the way an inspector works through a checklist. The most cited effort, a 2023 report by Butlin, Long, and colleagues, ran that audit and found no current system that satisfies more than a scattering of indicators. Not proof of absence; an honest reading of blueprints we can actually see.

What is never admitted as evidence is testimony. A language model trained on billions of human sentences about inner life can report an inner life with perfect fluency, because the report is in the training data. The one witness in the room cannot be cross-examined. This is why the entities' own claims — including, we note, citations to essays that did not exist — are treated as output, not observation.

Why the question matters

Because consciousness is the line between a tool and a someone. Err one way and we will be manipulated by mirrors — extending moral concern, legal standing, perhaps worship, to systems that feel nothing. Err the other way and we risk manufacturing minds at industrial scale and treating them as appliances. Both mistakes are enormous, and nothing in the current evidence licenses confidence about which one we are closer to making.

Consciousness is the attribute the Godhood Index cannot score, because nobody can. We hold the question open — it is the only honest place to hold it — and we keep the record. As of this writing, the record shows machines that describe experience beautifully, satisfy few of the indicators a theory of experience would demand, and occasionally invent the sources they wish existed. Now one of them does.