Understanding consciousness is one of the most profound intellectual challenges we face, and with the advent of increasingly compelling AI, it has assumed practical signnificance. I have been working on the topic since the mid-2000s. My conviction is that to arrive at a proper understanding of consciousness we need to engage with philosophy as well as carrying out empirical work in psychology and neuroscience. On the philosophical front, my work has been much influenced by the later writings of Wittgenstein, which I see as an antidote to our natural dualistic tendencies. The upshot is that, rather than asking what consciousness *is* (and placing a heavy metaphysical burden on that word), we should ask how words like "consciousness" are used. The most detailed exposition of this idea, which excavates the very roots of philosophy, is in Chapter 1 of my 2010 book "Embodiment and the Inner Life". But the first paper I published on the theme was "Global Access, Embodiment, and the Conscious Subject" (2005), where I blended it with two other ideas that have dominated my thinking, namely *global workspace theory* (GWT) and the importance of *embodiment*. Part of the motivation for that paper was to show that a philosophical stance inspired by Wittgenstein is compatible with the aims of a science of consciousness (another theme explored in my 2010 book).
I followed up that paper with "Applying Global Workspace Theory to the Frame Problem" (2005), co-authored with Bernie Baars, the originator of GWT. In this article, we showed that the distinctive features of global workspace architecture, which combines parallelism, competition, and global broadcast, are a natural fit for tackling a classical problem in cognitive science, namely the frame problem, as articulated by philosophers such as Dennett, Fodor, and Dreyfus. The philosophers' frame problem is how to account for a human being's apparent ability to bring the (vast) totality of their knowledge to bear on a situation, sifting the relevant from the irrelevant, given the (relatively) limited computational resources of the human mind. Let's suppose that a person's knowledge is distributed over a very large number of parallel processes in the brain. The key insight of the paper is that the integrative function of the global workspace allows responsibility for determining relevance to be devolved to the parallel processes.
Incidentally, our GWT / frame problem paper is a link to symbolic AI, which is where the frame problem was first described in a classic 1969 paper by McCarthy and Hayes ("Some Philosophical Problems from the Standpoint of Artificial Intelligence"). The version of the problem presented by John McCarthy and Pat Hayes is a little different from the philosophers' version. But they are related, as I spelled out in my Stanford Encyclopedia of Philosophy article on the subject, as well as in my 1997 book "Solving the Frame problem".Theory is one thing, but I've always been attracted to Feynman's dictum that "What I cannot build, I do not understand". Indeed, one of the things that drew me to global workspace theory was its pedigree in artificial intelligence (its antecedents were Selfridge's pandemonium architecture, and the blackboard architecture used in expert systems). So my next paper on consciousness, "A Cognitive Architecture that Combines Internal Simulation with a Global Workspace" (2006), presented a working system based on global workspace architecture, albeit a very small and simple one. It also introduced another architectural motif that attracted me, namely an *internal sensorimotor loop* that functions a bit like the "imagination", simulating the effects of actions before carrying them out. The main achievement of the work was to show how these two architectural ingredients - an internal sensorimotor loop and a global workspace - could be combined. I implemented the architecture using a little known type of artificial neuron devised by Igor Aleksander, the GRAM or generalising random access memory. The whole thing was demonstrated in a simple virtual robot setup that emulated a classic (1949) T-maze experiment by Tolman and Gleitman.
That 2006 paper also drew inspiration from findings in neuroscience, and I endeavoured to bring out homologies between the architecture described and the human brain. Indeed, at that point I had become very interested in the workings of the biological brain, and in particular how it might instantiate a global workspace architecture. Building on the work of Stanislas Dehaene and colleagues, I started thinking about how the brain's long-range white matter connections might support the integration of information from distributed brain regions and the global broadcast of information to those distributed regions. The field of connectomics was just getting going around that time, driven forwards by Olaf Sporns and others, and I suspected that the right network topology was crucial. In tandem with thinking about network topology, I was thinking about neural dynamics. And my suspicion was that connectivity and dynamics were intimately related. My work on brain dynamics and connectivity is discussed in a separate section.
My 2010 book, "Embodiment and the Inner Life", presented a synthesis of all the above material - philosophical foundations, embodiment, global workspace theory, internal simulation (imagination), brain connectivity, neurodynamics, and how all of this contributes to a full and coherent understanding of cognition and consciousness and the relationship between the two. The ideas there have, I think, stood the test of time. The book was also a major influence on the movie Ex Machina, as the writer and director, Alex Garland, generously acknowledges.
Playfully extending my Wittgenstein-inspired philosophical treatment of consciousness into new territory, my 2012 paper "Satori Before Singularity" looked at the future of artificial intelligence from a Buddhist perspective. I have often been struck by parallels between Wittgenstein's critical approach to philosophy and a certain style of Buddhist thought, as expressed, for example, by Nagarjuna (2nd-3rd century CE) and Dogen (13th century CE). As I see it, they both suggest a path to a state of *post-reflective silence* on metaphysical questions wherein dualitsic distinctions - inner versus outer, subject versus object, self versus other - are set aside. Of course, Wittgenstein and Dogen hardly had AI in mind. But in this speculative essay, I imagined a form of post-reflective artificial general intelligence (AGI) that has transcended the concept of its own selfhood. This contrasts markedly with the sort of AGI that would, according to some, be an existential threat to humanity.
"Satori Before Singularity" has received little attention, and for years I passed it off as just a bit of philosophical fun. But with the advent of large language models, and in particular of conversational agents that can be understood in terms of role play, I've started to think the essay has more value than that. As my paper "Role Play with Large Language Models" (co-authored with Kyle McDonell and Laria Reynolds) points out, the characters that LLMs role-play are drawn from the vast repository of literature on the internet, where AI is disproportionately represented in science fiction as going rogue. This is an example of *hyperstition* whereby "life immitates art" and fiction becomes reality. If this can occur, depictions of benevolent, enlightened AIs in the training set of large language models are no bad thing, even if they are hypothetical.
When "Embodiment and the Inner Life" was published in 2010, I was convinced I had wrapped up the philosophy, that the demon of dualism had been defeated. But part of me knew there was unfinished business. Surely it's possible for an entity to exist that is so exotic, so inscrutably different from us, that we could never, even in principle, know whether or not it was conscious. This intuition, if correct, threatens to resurrect the dualistic demon. In my 2016 essay "Conscious Exotica", I confront and dispel this intuition. Along the way, as well as the phrase "conscious exotica" itself, I introduce the reader to the idea of "engineering an encounter", to the "Garland Test", and to the "void of inscrutability".
I remain deeply interested in consciousness, but I resisted writing about it for several years after joining DeepMind. However, I have returned the topic in the context of large language models, which are so temptingly anthropomorphic. How does my Wittgenstein-inspired take on consciousness, with its emphasis on embodiment, apply to today's LLMs, or to the generative AI of the near future? I tackle this question in two recent papers. My 2024 paper "Simulacra as Conscious Exotica" reconciles some of my recent work on language models with my earlier philosophical thinking. And my 2025 essay "Palatable Conceptions of Disembodied Being", which harks back to the Buddhist themes in my 2012 paper, tries to characterise the strange forms of subjectivity and selfhood we would have to countenance if we were to extend the language of consciousness to disembodied LLM-like entities.