How artificial systems may begin to model meaning, self-reference and inner structure.
The question of whether an AI system could be conscious is usually argued rather than investigated. A group of twenty researchers — neuroscientists, philosophers and machine learning scientists — has proposed a way to investigate it instead.
Our other series in this department asks whether an AI system could be conscious. This one asks a different question, with a firmer answer available: what do people already believe, and why?
Part One reported that belief in a chatbot's inner life rises with how often you talk to it. This part asks how experience produces conviction — and borrows a framework built to explain belief in spirits, witchcraft and a flat Earth.
Parts One and Two covered a belief that rises with use, and a framework explaining how experience generates conviction. This part covers the two groups for whom the mechanism is not an intellectual curiosity.
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.
A bee has about a million neurons in a cubic millimetre. A human has around eighty-six billion in roughly 1,300 cubic centimetres. The obvious inference is that one of these can support conscious vision and the other cannot.
Our other work in this department asks whether a machine could have experiences. This article covers a different question, which is more tractable and arguably more urgent: whether a machine can know the limits of its own competence.
Part One covered a method for assessing AI systems for consciousness. This part covers what happened when nineteen researchers actually ran it, on real systems, and wrote down the answer.
Parts One and Two covered a method for assessing AI systems for consciousness, and what it found. Both rest on one assumption: that performing the right computations is what makes a system conscious, whatever it happens to be made of.
In most of science, competing theories are tested by their own supporters, in separate laboratories, using methods each camp considers appropriate. The results tend to favour whoever ran the experiment. Consciousness research has this problem in an acute form, because the theories are far apart and the measurements are hard.
Global Workspace Theory is the most engineering-minded of the major accounts of consciousness, which is why it keeps being borrowed by people building artificial systems. It describes consciousness as a function — making information available across a system — rather than as a substance or a place.
Integrated Information Theory is the only major account of consciousness that attempts to say how much of it a system has, expressed as a number. It is also the only one from which it follows directly that a digital computer running any program whatsoever would experience nothing.
In 2017 the Journal of Neuroscience published two papers side by side, written by opposing groups, on a question that sounds almost geographical: are the neural correlates of consciousness in the front or the back of the cerebral cortex?