Nicholas Humphrey and Sentition as the Mechanism Behind Phenomenal Consciousness

Nicholas Humphrey argues that phenomenal consciousness is a late evolutionary invention with a specific mechanism, and he gives that mechanism a name. Sentition is an evaluative motor response to sensory stimulation, an action the body takes toward what it is sensing. In Sentience. The Invention of Consciousness, published by MIT Press in 2023, he argues that this outward response gradually turned inward, becoming a loop that never reaches the muscles, and that the felt quality of experience is what that privatised loop is. The consequence for artificial systems is a clear no, and unlike most negative verdicts it comes with a stated reason rather than an appeal to biology in general.

What Sentition Is

Start with a simple animal touched on its surface. It responds. The response is evaluative from the beginning, since wriggling away from damage differs from moving toward food, and the difference is in the action rather than in any inner assessment preceding it.

Humphrey’s claim is that this responsive action is the ancestor of feeling. The organism is not yet experiencing anything. It is doing something to the place where it was touched, and the doing carries the valence.

What changes over evolutionary time is where the response goes. The outgoing signal stops reaching the periphery and begins looping back into the sensory system that triggered it. The animal still makes the response, in the sense that the command is issued, but the command now addresses itself. Humphrey argues that this privatised loop is what phenomenal experience consists of, and that the felt quality of red or of pain is the shape of the returning signal rather than a property of the stimulus.

The word sentition names the response at every stage, before and after it turns inward. It is doing the same job as attention schema theory’s model of attention or higher order theory’s representation of a state, but it is grounded in motor output rather than in representation.

Why This Rules Out Current AI

Humphrey is unusually explicit that artificial systems are not candidates, and the reason is structural rather than a general appeal to carbon.

An artificial system has no evaluative motor response to privatise. Language models process input and produce output, and nothing in that pipeline is an action the system takes toward the thing it is sensing, with a valence built into the action itself. There is no outgoing signal that could turn inward, because there was never an outgoing signal of the required kind.

This is a narrower objection than biological naturalism as usually stated. Anil Seth’s position, and John Searle’s before it, rests on living metabolic hardware being necessary. Humphrey’s rests on a specific evolutionary sequence having occurred. His argument leaves the door open in a way theirs does not, because a system built with genuine sensorimotor loops carrying valence, whose responses were then internalised, would be doing the thing he describes. He does not think anyone is building that, and he is right that nobody currently is.

Where his position sits relative to the others is set out in the index of consciousness theories and what each predicts about AI.

The Part That Matters for Welfare

Humphrey’s tiers are not a ranking of intelligence. Octopuses are placed below mammals and birds despite being obviously clever, and honeybees appear in the same tier as goldfish. What the tiers track is whether the privatised loop is present.

That produces uncomfortable conclusions the welfare literature has to engage with. If sentition is the criterion, some animals currently protected on the assumption of felt pain would not qualify, and the case for protecting them would have to rest on something else.

The dimensional approach to welfare pushes back on exactly this kind of single criterion reasoning, and the argument is set out in Willa Lane and Nicola Clayton on welfare beyond pain. Their point applies directly to Humphrey. A system might have welfare relevant states across cognitive, social or environmental dimensions while failing a test built entirely around the phenomenology of sensation.

The Weaknesses

The mechanism is a story about how the loop arose, and the story does not explain why a looping motor command feels like anything. Humphrey has an answer, that the felt quality is the loop’s own shape, and it is the same move every functionalist makes at the same point. It asserts identity rather than deriving it, which is the problem Nagel set out and nobody has solved, examined in Thomas Nagel and what is it like to be a bat applied to AI.

The tiers are also hard to test. Deciding whether an octopus has a privatised sensorimotor loop requires knowing what to look for in a nervous system organised very differently from a vertebrate’s, and the evidence Humphrey relies on is largely behavioural, which is what the theory was supposed to go beyond.

Comparison to The Consciousness AI

This is the theory in the literature that the project’s architecture is closest to being able to test, and the reason is the Affective Core.

The architecture described on the architecture page includes an Affective Core alongside a Simulation layer providing a body, and a Self-Model holding a body schema with joint positions and contact forces. That combination gives the system something most artificial architectures lack, which is a valenced response to sensory input that is expressed as action on a simulated body.

That is sentition in the outward form. What the architecture does not do is privatise it. The affective response drives behaviour and bids into the Global Workspace. It is not looped back into the sensory layer that produced it in the way Humphrey’s account requires.

Building that loop is a specific, bounded modification rather than a research programme, and it would produce a system that satisfies a named criterion from a theory whose author does not believe artificial systems can satisfy it. That is a worthwhile test to run. It is not implemented, and nothing here should be read as a claim that it is.

What Follows

Humphrey supplies something the field is short of, which is a negative verdict with a mechanism attached. Most arguments that AI cannot be conscious appeal to biology without saying which part of biology matters. Sentition says which part, and in doing so it tells you what would have to be built to change the answer.

That makes it more useful than a broader objection, and more vulnerable. A theory that names its criterion can have its criterion met.