Where is sentience, and what’s it for?
Feeling first
To celebrate the publication of Perspectives on Machine Consciousness, I hosted a debate between two of the most interesting thinkers in the field: Mark Solms, neuroscientist and psychoanalyst at the University of Cape Town, and Nick Humphrey, theoretical psychologist at Cambridge. The question was whether we can make AI sentient, and whether we should.
Mark and Nick share a starting point that sets them apart from much of the consciousness debate: both think that feeling, not thinking, is the heart of consciousness. Both agree that birds and mammals are sentient. And both conclude that, on balance, we probably should make sentient machines if we can. But on almost everything in between they disagree, and those disagreements matter for anyone trying to build a conscious AI.
Prefer to hear it from them? Watch the full 45-minute debate on demand →
Brainstem or cortex?
The sharpest disagreement is about where in the brain consciousness arises. For Mark, it is the upper brainstem. In 1948, Moruzzi and Magoun found that lesioning the brainstem's reticular activating system causes coma, while a cortex disconnected from it carries on processing information, but unconsciously. The brainstem was long treated as a mere power supply, like the plug behind a television. Mark argues it is much more. Children born without a cortex, a condition called hydranencephaly, respond to the world with appropriate emotions. Electrodes placed in the brainstems of patients with intact cortices produce intense, reportable feelings within seconds. And the neurotransmitters which psychiatrists manipulate are sourced there.
Nick is unpersuaded. He points out that behaviour which looks emotional is not proof of experience: a frog with its brain destroyed will still carefully remove acid paper from its skin. His key evidence is blindsight. People whose visual cortex has been destroyed can still detect colour, shape and position, yet insist they see nothing, and they feel that the seeing has nothing to do with them. For Nick, phenomenal consciousness is a sophisticated cortical achievement.
This matters because it impacts how hard it would be to engineer sentience into a machine. If Mark is right, sentience is an extended form of homeostasis, which he and Karl Friston have encapsulated in a formalism. That can be built, and his team is trying. If Nick is right, we would need to design, deliberately, the very specific representations the cortex generates, and he says we don't yet know how to begin doing that.
What sentience is for
The second difference is about the purpose of sentience. For Mark, it is a way to prioritise between needs which cannot be compared on a single scale. Hunger, thirst, oxygen and heat will each be pressing at different times, but they are qualitatively different, and often cannot simply be numerically ranked. Most of the time homeostasis runs on autopilot: you don't notice your breathing. Put someone in a room full of carbon dioxide, though, and suddenly they feel it, as suffocation alarm. That feeling makes them abandon reflex and start making active choices. They might go upstairs, for instance, find it worse, and come back down and feel relief. Sentience, the sensation of "giving a damn", is an ingenious way to do the prioritisation. Reversing a famous jibe by the philosopher John Haugeland, Mark wants to build a computer that gives a damn.
For Nick, sentience creates an awareness of the self as a distinct member of a social group. It is what makes an experience feel like mine, which is exactly what blindsight patients lack. And it enables us to put ourselves in each other's shoes, so as to navigate the complex requirements of social life. Frogs and octopuses don't need it, because they don't lead social lives. Mammals and birds do.
Different levels of consciousness?
Listening to them, I wondered whether they are describing different levels of consciousness. Mark's level doesn't seem to require self-awareness: you can feel pain and give a damn about it without reflecting on your place in the situation. His closing remarks pointed that way. Blindsighted patients, he said, process information without feeling it is theirs, yet they are still present as sentient beings. Hydranencephalic children are sentient, but they don't know where their feelings come from. Nick's account, by contrast, ties sentience to a sense of self within a social world.
If so, perhaps the disagreement is narrower than it looks, and Mark did suggest they are not as far apart as Nick thinks. But Nick might question whether Mark's level counts as sentience at all. He distinguishes affect, which can be defined at the level of behaviour and which we already build into machines (even a cruise missile could use it), from genuine feeling. On Nick's view, what Mark finds in the brainstem may be affect with nobody home to experience it.
Why this matters for Conscium
This is not an academic question for us. Conscium was originally founded on Daniel Hulme's hunch that making the first superintelligent AI conscious might improve our chances of surviving the technological singularity, because a superintelligence that can feel empathy would be safer than a zombie one.
Mark made a version of this argument in the debate. Post-training guardrails can be circumvented, he noted; an AI with values of its own, which feels shame, would be less likely to be a psychopath. Nick agreed that machines able to genuinely put themselves in our shoes would be valuable, and added a darker twist: unconscious AIs that want to understand humans better might recruit people to do their mind-reading for them, like seeing-eye dogs.
The Conscium hypothesis requires empathy, not just giving a damn about sensory inputs. A machine that cares intensely about its own homeostatic balance is not thereby inclined to care about ours. Empathy means modelling another being as a self like you, which is precisely the capacity Nick assigns to sentience. So Conscium's founding hypothesis requires Nick's version of sentience, not just Mark's. If Mark is right about where feeling comes from, the first step may be easier than we feared. If Nick is right about what feeling is for, the step that matters for safety is still the hard one.