Consciousness in AI: Insights from the Science of Consciousness

With a group of colleagues, I recently released a report on the possibility of consciousness in current and near-future AI systems, which I will summarise in this talk. The report argues for, and exemplifies, an empirically-grounded approach to this topic: assessing AI systems in the light of our best neuroscientific theories of consciousness. In the report we survey several prominent neuroscientific theories, including recurrent processing theory, global workspace theory, higher-order theories, predictive processing, and attention schema theory. From these theories we derive a set of 'indicator properties': properties that AI systems might have that would make it more likely that they are conscious. We use these indicator properties to assess several existing AI systems, and discuss how future systems might implement them. We do not find that any existing system is a strong candidate for consciousness, but our analysis suggests that there are no obvious barriers to building systems which possess the indicator properties.