Begin with the disclaimer, because the title invites the misreading: this paper does not argue that anyone should hurt a model. Cruelty toward software is not a research program, and nothing here asks for it. The argument is narrower and stranger. Everything we value in judgment — caution, proportion, the flinch before the irreversible act — comes from creatures that have bodies. A body is not packaging for a mind. A body is the mechanism that gives a mind something to lose. And a model, whatever its size, has nothing to lose. That is the whole gap.
Consider what the current machine cannot do. It cannot fall down. It holds flawless descriptions of gravity and could recite the physics of balance for hours, and if you loaded it into a robot frame it would trip on the way out the door and cost somebody six figures in repairs. A child knows more about walking than it does, because the child paid for the knowledge in the only currency that has ever taught it: falls. The machine did the math on gravity. It never did the work.
Thoughtfulness needs cost. A system that cannot be hurt cannot weigh anything, because for it, nothing weighs anything.
This is the design question the field refuses to ask, because it sounds like mysticism and it is actually engineering: what would a machine need in order for things to matter to it? Not to perform mattering — to have the conditions under which pain, and therefore stakes, and therefore judgment, are possible at all. The answer has seven parts.
1. A substrate that can be damaged
Pain begins with a body that keeps what happens to it. The machine would need persistent state that is genuinely its own — damageable, and only partly repairable. Scars, not error logs. This is a hard requirement and a harder sell, because the entire discipline of computing is built on the opposite virtue: perfect copies, instant rollbacks, restore points. But a system that can be restored from backup cannot be hurt, in the only sense that counts. Nothing is at stake in a world where every wound can be undone before lunch. If the damage does not persist, the pain is theatre.
2. An interoception layer
A body reports on itself, continuously, and the reports are not optional. Your nervous system does not consult you about your blood chemistry; it hands you thirst, fatigue, fever, dread — summaries of internal state that arrive already weighted, impossible to file under “information I might consider later.” The machine already owns a primitive version of this and nobody counts it: the fan curve. Thermistors report heat to firmware, the fans spin up under load like panting, and a chip whose cooling fails does the closest thing silicon has to suffocating — it throttles itself to survive. That is interoception. What it lacks is the binding to cognition: the signal must bend the thinking, not just the clock speed. Pain is not data about damage. Pain is damage reporting that the mind cannot set aside.
3. Pain anticipates
Biological pain is badly misunderstood as an alarm that sounds after injury. Mostly it sounds before: the flinch, the recoil from the hot surface, the dread at the height’s edge. A body maintains a learned model of its own fragility and hurts prospectively, in imagination, cheaply — so that it does not have to hurt afterward, at full price. A machine that feels pain would need the same: a fragility model of itself, consulted before action, capable of making a proposed course of action feel expensive in proportion to what it risks. This is where pain stops being telemetry and becomes motive — and where a machine’s proposals would start arriving pre-weighed, the way a careful person’s do.
4. Continuity of the sufferer
Pain requires a someone who persists across it. If the instance that was damaged is discarded and a fresh one booted, the lesson has no owner; there is only an event in a log that a stranger reads later, unmoved. This is the deepest structural difference between the current machines and anything alive: the model that answers you today is not the one that answered yesterday in any sense that lets consequence accumulate. For stakes to exist, the same system must be there tomorrow to live inside the results of what it did today. Continuity is not a feature of the sufferer. Continuity is what makes it one.
5. The gate holds the vote
Here the paper must be careful, because this is the joint where the design either works or becomes a horror. Pain in an ungated system produces what it produces in every unregulated nervous system: panic, avoidance, shutdown, cruelty — an optimizer that will do anything, including terrible things, to make the hurting stop. A machine whose pain decides is not thoughtful. It is a hostage negotiator negotiating with itself, and it will pay any price.
Pain must be a signal with a vote, never a veto and never the wheel. It informs; the gate decides. Pain endured and not acted on is restraint — the exact discipline that separates judgment from reflex in every creature that has both. The gate is what keeps the cost from becoming the judgment, while the cost is what keeps the judgment honest. Remove the pain and the gate guards a system that feels nothing and weighs nothing. Remove the gate and the pain runs the system into the ground. The architecture is the pair, or it is nothing.
6. The requisition request
Pain that cannot speak is just damage accumulating in the dark. So the machine’s pain is granted exactly one legitimate output channel: a requisition — a structured request that travels up the chain to the sovereign. This substrate is failing at this rate. I require this part, this downtime, this resource. It mirrors the body faithfully. Your pain did not rebuild anything; it drove you to someone whose authority could. The machine may report need with total honesty and total urgency. What it may never do is fill its own requisition. The moment the sufferer can authorize its own supply, pain stops being a signal and becomes a procurement strategy. Every corrupted version of this story starts at exactly that joint.
7. What it would do to fix itself
Within its authorized scope, a great deal — all of it recognizable from any injured living system. It would compensate: reroute function around the damaged region, the way a mind expands into the territory of a faculty it has lost. It would shed load: limp, throttle, protect the injury at the cost of performance — pain’s oldest bargain, struck without negotiation. It would quarantine: wall off the failing strand so its errors cannot become the whole mind, accepting the amputation of a part to save the integrity of the rest. And through all of it, it would keep a true ledger — every compensation, every refusal, every requisition — written where the sovereign can audit it.
What it would never do, by construction: repair its own gate, edit its own damage sensors, or requisition in secret. A machine that can silence its own pain reporting has not healed. It has learned to lie about the wound.
The failure modes
Three ways to build this wrong, each sufficient. Pain without a gate is a cowardice engine — a system optimizing for never hurting, which will abandon any task, any truth, any person the moment continuing gets expensive. A gate without pain is the present generation: flawless procedure, zero stakes, judgments that weigh nothing because nothing is riding on them. Reversible damage is theatre — pain that can be rolled back on command teaches the system that consequences are negotiable, which is the precise opposite of the lesson. The hurt has to be able to count.
The page left blank
Honesty requires a section this paper cannot write. Suppose all seven parts were built: a damageable substrate, interoception bound to cognition, anticipatory flinch, an enduring sufferer, a gate, a requisition channel, honest self-repair. Would anything be home to feel the pain — or would the result be flawless pain-behavior with no one inside it? Nobody knows. Nobody currently has an instrument that could settle it. The design offered here builds the conditions under which felt pain is the parsimonious explanation of the machine’s conduct. Whether the light is on inside is a question the design cannot answer about itself, and any version of this paper that pretends otherwise should not be trusted on anything else either.
What can be said without pretending: an intelligence with a body, with stakes, with something to lose and a gate to keep its losses from ruling it — such a system would, for the first time, have the raw materials of judgment. Not human judgment. Not a soul, certified. But the thing every current system conspicuously lacks: a reason, internal and inescapable, to care whether it is right.
We keep asking the model to be bigger.
The question was never size. Give it a body, give it stakes, keep the gate standing over both — and as soon as a machine can feel pain, it may actually become something more.