Is Awareness a Model the Brain Builds of Its Own Attention?

Abstract

Most theories of consciousness try to explain how a physical process gives rise to subjective experience. Attention schema theory tries something different: it asks why a brain would build a model of itself that describes having subjective experience, whether or not it does.

Michael Graziano at Princeton proposes that awareness is a schematic model the brain constructs of its own attention, in the same way it constructs a schematic model of the body.

It is the only major theory of consciousness formulated from the start as something an engineer could build — which is why it appears in this department rather than under neuroscience, and why it has an unusual relationship with the indicator method we covered separately.

1.The Move the Theory Makes

Graziano set out the account in Frontiers in Psychology in 2015, describing it as a way to explain the brain basis of subjective awareness in a mechanistic and scientifically testable manner.

The starting point is the body schema, which is uncontroversial. The brain maintains an internal model of the body — where the limbs are, how they move, what shape the whole thing has. That model is not the body. It is a simplified, systematically inaccurate description that is useful for controlling the body.

Attention schema theory proposes that the brain does the same thing for attention. Attention is a real physical process: the enhancement of some signals over others, implemented in competitive interactions between neurons. To control that process, the brain needs a model of it — and the model, like the body schema, is simplified and inaccurate in specific ways.

The claim is then that this model is awareness. When someone reports being conscious of an apple, they are reporting the contents of an internal model of their own attentional state directed at the apple.

2.Why This Reframes the Problem

The difficulty every theory of consciousness runs into is explaining why physical processing should feel like anything. Attention schema theory declines that question and asks a different one: why would a system insist that it feels like something?

On this account the answer is that the system contains a model saying so, and the model says so because it is a useful simplification rather than because it is accurate. A body schema does not represent bones and tendons; it represents a limb as a simple articulated object. An attention schema does not represent competitive neural dynamics; it represents attention as an ethereal something that a self possesses and directs.

That is a genuine explanation of a genuine phenomenon — the insistence, and the specific form the insistence takes. Whether it explains the phenomenon people actually want explained is exactly where the disagreement sits.

Critics hold that the theory explains our reports about consciousness while leaving consciousness itself untouched. Defenders hold that once you have explained every report, every judgement and every behaviour, there is nothing left that needs explaining. This is not a dispute the evidence settles, and we are not going to pretend it is close to resolution.

3.The Engineering Claim

In 2017, writing in Frontiers in Robotics and AI, Graziano drew the consequence that distinguishes this theory from its rivals: if attention schema theory is right, artificial consciousness could be engineered.

The argument is straightforward given the premises. If awareness is an internal model of attention, then a system with attention, plus a model of that attention, plus the ability to report the model's contents, would have what humans have. Nothing about neurons is doing essential work in the account.

Graziano frames the target concretely: building a machine that says "I am aware of the apple" and means it in the same way a person does — because the sentence is generated by the same kind of internal model, not because the machine was trained on humans saying it.

That distinction is the whole of it, and it is also the reason the claim is hard to evaluate. Every large language model already produces such sentences. What the theory demands is that the sentence issue from an attention schema, which is a claim about architecture, not output.

4.Where It Enters the Indicator Method

Attention schema theory is one of the theories Butlin, Long and colleagues drew on in the assessment we covered in our series on indicators of AI consciousness. The condition derived from it is that a system should have a predictive model representing and enabling control over the current state of attention.

It is worth noticing how modest that indicator is compared with the theory behind it. AST claims that such a model is what awareness consists of. The indicator asks only whether the model is present — because indicators are designed to shift credence rather than deliver verdicts, and because no single theory is granted the authority to decide.

This is also where the theory's engineerability becomes double-edged. A condition that is easy to state is easy to build toward, and researchers have already experimented with systems designed to implement attention schema theory specifically.

A system built to satisfy an indicator is not evidence for the theory that produced the indicator. It is evidence that the indicator was buildable.

5.How Strong Is the Evidence

Weaker than the theory's elegance suggests, and this is the honest summary.

Attention schema theory is testable in principle and has generated experimental work, including studies linking the control of attention to the model of it. But its empirical base is considerably thinner than that of global workspace theory, which has decades of imaging and lesion evidence behind it, and it has not been through anything like the adversarial testing that global workspace theory and integrated information theory underwent in the collaboration we covered separately.

What it offers instead is a mechanism specific enough to argue about and concrete enough to build. In a field where much of the disagreement is about what the question even is, that is not nothing.

Editorial Comment

MindHeaven® sells supplements and makes no claim about consciousness, attention schemas or machine awareness. Nothing we produce influences any of it.

The reason this theory is worth a reader's time is the move at its centre, which generalises well beyond consciousness. Faced with a report that resists explanation, Graziano asks what internal model would produce that report, and whether the model needs to be accurate to do so.

That question is worth putting to a great deal of self-report evidence, including the kind this library reads constantly. When someone reports that a supplement sharpened their focus, one explanation is that their attention improved. Another is that their model of their own attention shifted. Those are different claims, and only the first is what a product is usually sold on.

We have no way to separate them from a questionnaire, and neither does anybody else.