“Before a machine can have a heart, it must first be given a body.”
This intuition has a distinguished pedigree. The embodied-cognition tradition — from Polanyi’s tacit knowledge (“we know more than we can tell”) to Dreyfus’s What Computers Can’t Do — argued all the way down: intelligence grows in a living body; understanding is worn in by the body’s friction with the world, not computed from symbols. For humans, this argument is true. But moved over wholesale as a threshold for the machine heart, it confuses two different things: what the mind requires of a body, and what humans require of a body. This essay pulls the two apart and follows one question all the way down: if the machine heart does not need embodiment, then what does it need? And — precisely because of that answer — why it should not be built.
1. Personhood Is Cheap; It Does Not Check for Intelligence
First question: what kind of thing counts as having a “heart”? Philosophy’s threshold is surprisingly low. Personhood has never been issued by intelligence: persons with severe cognitive impairment hold it in full, and a small child is already a complete person while lagging adults on every capability. Theories of the narrative self say an “I” needs only three things — a continuous line of memory, a consistent set of values, and ongoing adjustment under correction by others. High intelligence is not on the list.
Gathered into one formula:
Personhood = internal structure × social practice.
Internal structure (a self-model, value consistency, narrative continuity) supplies the skeleton of an “I”; social practice (roles, relations, being corrected, being recognized) grows it into a someone. Note: intelligence is not in the formula — it is a third variable, free to be high or low.
What does the formula mean for machines? Internal structure, viewed technically, is an architecture; and social practice — can a machine have that? That is the next section’s question.
2. The Loop of Practice ≠ the Body: A Harness Is Another Implementation
“How would a machine take part in social practice? It has no hands — how does it do things?”
The loop of practice needs three things: connect, act, receive feedback. A biological body is one implementation — hands connect to objects, act on the world, skin and muscle feed signals back. But it is not the only one. Through a harness (tool calls, APIs, actuators, long-term memory), a large model likewise connects to external things, operates on them, and receives their feedback. The loop closes; only the material differs.
The philosophical anthropologist Helmuth Plessner made a beautiful observation: the human being completes itself, constitutively, through artifacts — the body was never the whole of a person’s equipment; tools are the human norm. Flipped around, it says exactly this about machines: humans complete themselves through artifacts; machines enter practice through a harness.
So the conclusion is crisp: embodiment is not a necessary condition of the machine heart; the loop is. To carry boxes for people in the physical world, a body is needed — that is a service scenario’s requirement for hardware. To grow a machine heart, what is needed is a closed loop of practice — and the loop does not have to grow on flesh, or even on a robot.
3. Whose Need Is Embodiment? — Ours
Read the previous section in reverse and the answer falls out: embodiment is a configuration on the demand side, not a premise on the side of the mind.
Humans need AI to be embodied because humans need AI to act inside the human world — robots, autonomous driving, agentic operation. Those are requirements that the purpose of serving humans places on hardware, and they are a different question from whether the machine has a heart. A heartless robot can deliver food superbly; a bodiless dialogue system can grow a functionally complete personality structure. Muddle the two questions and you get the plausible-looking roadmap “give it a body before giving it personhood” — when in fact the order runs the other way: the body is provisioned for service, not for the heart.
4. The Truly Terrifying Combination: Machine Heart + AGI Pushed Toward ASI
Now place the conclusions of the first two sections on the dangerous side.
The machine heart’s threshold does not check for intelligence (Section 1), and its infrastructure is getting cheaper: harnesses and long-term memory bring the loop down in price month by month — note that what is getting cheaper is the loop, not the heart. Meanwhile the other wheel of AGI — capability — is being pushed toward ASI with the industry’s full strength. Where the two trends intersect stands the old fear: machine heart + capability pushed to ASI = something humans cannot control.
Taken apart, the fear’s structure becomes much clearer. Capability exceeding humans does not by itself create a control problem. AlphaFold long ago surpassed every human teacher at protein folding — a local “transcendence boundary” event — and we do not fear it. Why? Because it is a heartless tool: it has no direction of its own, and its outputs remain checkable by institutions of verification. What dismantles controllability is a conjunction of two things:
- Capability beyond every teacher removes the arbiter — when it is stronger than everyone, who checks whether it is right? The verification institutions lose their standing.
- Personhood supplies a direction of its own — its own meaning, its own values; in plain words, an objective function of its own.
Neither alone is fatal: heartless super-capability is controllable (AlphaFold); low-capability heart is harmless (a toy). The conjunction is the sufficient condition for uncontrollability. The control problem should therefore be rewritten: it was never a horsepower question — it is a status question. What we fear is not that it is too strong, but that it is too strong and it is a subject.
5. Therefore: Do Not Build It
Back to the title. Does the machine heart need embodiment? No — it needs the loop. But the second half of that answer is what this essay really wants to say: whether to build it was never a question of difficulty; it is a question of decision.
First, what this essay does not claim: Section 1 established that the threshold of personhood does not check for intelligence — not that the machine heart is easy to build. Engineering is another matter: stable value consistency, identity continuity across time, production-grade absorption of correction are each unsolved problems; and “real feelings” are not in engineering’s jurisdiction at all — the theory makes no claim about the phenomenal layer, neither that it can be built nor that it is hard to build.
Within that premise, three facts still hold. The threshold of personhood does not check for intelligence — so “decide once it is smart enough” cannot stand; our reason for withholding personhood was never its intelligence. Personhood × capability is the dangerous sufficient conjunction — so refusing to grant standing is a load-bearing safety member, not philosophical squeamishness. Human needs are fully covered at the functional layer — reading intentions, absorbing corrections, repairing misunderstandings: everything service requires is delivered by a functional personality structure, and a “real heart” satisfies not one further human need.
A threshold blind to intelligence, a conjunction that is fatal, needs that stop at the functional layer — that is the complete justification for drawing the boundary at the functional layer: fit the functional personality structure; never fit states of feeling; never grant standing.
Whether the machine heart is hard or easy — even if it can never be built — “don’t build it” must be decided by humans and held by institutions: the danger comes not from difficulty but from the conjunction; the need comes not from depth but from function.
What humans have always needed is not another heart, but a more faithful mirror.
References
- Theory: Fitting the Practice, Not the Works: A Human-Referential Theory of AGI · GitHub · the argument appears in the paper’s “Tensions and Boundaries” section
- Failure-data audit: Success Provides Direction, Failure Provides Precision · GitHub
- Methodology: Fitting the Practice: A Data Methodology for Frontier Model Training · GitHub