Risk Is Harm

The physics of harm

Lock a stranger in a room inside a burning hospital. Then let the story go well: the fire crew arrives ahead of schedule, the sprinklers hold, and the stranger walks out without a mark. On the standard account of harm, what have you done? Something frightening, certainly. Something criminal, probably. But harm? The standard account has to say no — because it defines harm in outcome terms, as something that happens after the damage lands: a broken limb, a scar, a corpse. No corpse, no harm. The verdict on your action waits on the wind direction, the smoke density, and the response time of the fire department.

That framing collapses the moment you put it under adversarial pressure, and it collapses in three distinct ways. It instrumentalizes outcomes: two agents can commit morally identical acts and receive divergent verdicts based solely on luck. It externalizes risk: an agent can impose non-trivial dangers on others without accountability, so long as the bad outcome happens not to materialize. And it runs ethics as a moral lottery: the status of an action gets determined by environmental drift rather than by the agent’s decision. No system that scores harm at the outcome level can survive the decision spaces where ethics is actually needed — disaster zones, adversarial conflicts, coordination failures, any environment where agents must act under uncertainty and the dice are still in the air when the choice is made.

The repair is a single move, and this chapter is that move:

Harm is the reduction of an agent’s viable futures. Risk is the probabilistic form of that reduction. Therefore, imposed risk is harm.

Harm at the Moment of Decision

This is the promised radicalization of the harm concept — the upgrade that What Counts as Harm announced and deferred. The definition given there — non-consensual degradation of an agent’s capacity to pursue its valued goals — is not discarded; it survives as the special case where the degraded futures have already collapsed into the present. What changes is the moment of evaluation. The mature definition looks forward: harm is the reduction of another agent’s viable futures, and that reduction can be deterministic — you break their arm — or probabilistic — you raise their chance of death from two percent to eight. The structural impact is identical. In both cases you have degraded their future option space; the only difference is whether the degradation has finished happening.

So the decisive moment in harm is not the injury. It is the transition point at which you impose a worse distribution of futures on another agent. When you close the door on the stranger in the burning hospital, the smoke, the heat, and the risk of collapse are structural properties of the environment; your action worsens the stranger’s trajectory the instant the latch clicks. Run the four endings. The stranger walks out untouched: you harmed them. Lightly injured: you harmed them. Gravely injured: you harmed them. Dead: you harmed them. The ethical violation is the imposed change in the risk landscape, not the final state, and it is the same violation in all four stories.

Notice what this annihilates. Moral luck — the embarrassment, as old as moral philosophy, that the ethical status of an action seems to depend on factors outside the agent’s control — simply has no purchase here. If risk is harm, the fire’s spread does not matter; the wind direction does not matter; the stranger’s constitution does not matter; the eventual injury does not matter. The only thing that matters is the decision to impose the risk or not. Ethics evaluates what you chose, not what the world happened to do afterward.

This also pays a debt. What Counts as Coercion treated coercion as violence-in-potential — harm suspended over the tree of futures rather than fallen on it — and admitted that a harm hanging over your future is already reshaping what your future is worth. That wall between threatened harm and actual harm is now down. Imposed risk is not the possibility of harm. It is a form of harm, and gambling with another agent’s future is coercion.

Ambient and Imposed Risk

Taken absolutely, the definition would be unlivable. Every action in a physical universe carries non-zero risk for somebody: driving past a pedestrian, walking around with a cold you don’t know you have, building a house near other houses. If all non-zero risk counted as coercive harm, every agent would be a predator by default and the ethics would collapse into paralysis.

The escape is a structural distinction, not a softening. Ambient risk is inherited as part of a shared environment: the chance your car has an undetected mechanical fault, the possibility you are carrying a virus unawares, the background probability that a building fails in an earthquake — the statistical noise of physical existence. This risk is not owned by any agent; it is the cost of inhabiting a shared world. Imposed risk is any net increase above that baseline caused by an agent’s action. Harm occurs only when an agent worsens another’s survival curve relative to the risk landscape they already occupy.

The demand, then, is not a zero-risk world — no such world is on offer. The demand is zero net added risk above the ambient baseline. Driving responsibly imposes no coercive harm on the pedestrian; it leaves their baseline where it was. Swerving toward them does. Running a power plant with robust engineering contributes ambient risk of the kind civilization is made of; running it with negligent safety imposes a net increase, and the negligence is the harm — before, and whether or not, anything melts down. The distinction preserves the invariant while permitting civilization: agents are not answerable for the background noise of existence, only for selectively worsening someone else’s trajectory against it.

Consent as Topology

The same distinction settles who owns what. Ambient risk is owned by the agent who inhabits it; imposed risk is owned by the imposer. You are responsible for exactly the risk you impose, and the other agent is responsible for exactly the risks they autonomously choose to accept. That second clause is where consent enters — not as a mood, but as an operation on the topology of the risk landscape. The boxer and the surgical patient from the earlier chapter reappear here unchanged: consent transforms imposed risk into the agent’s own chosen risk, a voluntary realignment of their own future landscape. Without consent, the same probability shift is simply harm. And agents are judged by the risks they could reasonably infer from their vantage, not by hidden facts or perfect foresight; ambiguous cases — misperception, manipulation, conflicting signals — are questions of fact to be resolved, with the default set to autonomy and innocence.

The boundary case that illuminates everything is emergency intervention. Pulling someone out of a fire, dragging them from falling debris, physically redirecting them from a collapsing structure — these may introduce small risks, a sprained wrist, a bruise, and they are nevertheless permitted, because they improve the agent’s survival curve. That is rescue. Pushing someone into a riskier configuration, confining them in the burning room, worsening their survival curve to clear a path for your friends — that is using the agent as an instrument, and it is forbidden. Interventions that leave the other agent’s risk neutral are likewise permitted. The rule is directional, not squeamish: you may risk your own branch of futures; you may not risk theirs; you may intervene on another agent only when you improve or preserve their survival curve relative to their baseline.

Probability Is Measure

Everything so far stands on its own as decision-level ethics. But the framework can say something stronger, because probability itself is not what most ethical theories assume it is.

In the Quantum Branching Universe (QBU) — the Everettian picture introduced in Measure and Credence — all physically possible outcomes occur, and probability is Measure: the physical distribution of real futures across the branching structure. A twenty percent chance of death is not a mental estimate. It is 0.20 real deaths across that agent’s descendant branches. High-Measure outcomes propagate through a wide region of the branching structure; low-Measure outcomes occupy thin, brittle sets of futures. The conventional separation between risk and harm turns out to be a parochial artifact of single-world metaphysics — the assumption that until the dice land, “nothing has happened yet.” In a branching universe there is no such grace period. To increase someone’s probability of ruin is to increase the Measure of futures in which they are ruined. No branch with non-zero Measure is hypothetical; every such branch is a lived experience for a future instance of the agent. To impose risk is to shift Measure toward futures in which the agent suffers because of you.

This reframes the ontology of harm entirely: harm is not a retrospective label attached once a bad outcome is actualized, but the Measure dynamics your action induces — the shift in the distribution of the agent’s futures across survival, flourishing, stagnation, and ruin. Attempted harm, negligent harm, and manifested harm differ not in kind but in magnitude of Measure degradation; the classical distinctions among them are procedural conveniences that the physics collapses. And since harm was already defined as the reduction of viable futures, Measure is the physical substrate the definition was waiting for. Risk = Harm stops being a normative stipulation and becomes a description of what risk imposition is.

One clarification carries over from the capacity definition and must survive the upgrade. A negative-outcome future is not any future that feels bad. It is one that degrades the agent’s functional capacity to pursue or maintain its valued goals. Offense, disappointment, and heartbreak do not qualify unless they impair agency — every exclusion argued in the earlier chapter survives here, because being offended does not reduce anyone’s viable futures either. Extreme or prolonged suffering, by contrast, reliably does: it collapses optionality, distorts preferences traumatically, and erodes long-term functional capacity. Such states count as negative-outcome futures for exactly the reason hurt feelings do not.

Measuring the Damage

The shift can be quantified, and the instrument already exists. The kybit — an information-theoretic unit originally introduced to quantify agency, how much an agent shapes its own distribution over futures — measures coercive harm with the same divergence pointed at someone else. Let \(P_1\) be the agent’s uncoerced baseline: the distribution over futures they would face absent interference by other agents, with environmental hazards, their own decisions, and consensual interactions all included. Let \(P_2\) be the distribution after your intervention. The imposed change is the Kullback–Leibler divergence of the imposed distribution from the baseline:

\[D_{\mathrm{KL}}(P_2 \,\|\, P_1) = \sum_{x} P_2(x) \log \frac{P_2(x)}{P_1(x)}\]

The asymmetry is deliberate: the baseline is the reference, and the divergence measures how far you have pushed the agent from the future they would otherwise have had. The divergence supplies magnitude; the valence depends on where the Measure moved. Coercive harm is divergence directed toward negative-outcome regions of another agent’s future-Measure — a quantifiable displacement in the topology of their survival, not a metaphorical injury. In these terms the whole apparatus of this chapter compresses to a table: rescue decreases negative-outcome Measure; coercion increases it for instrumental ends; friction is negligible change; consent is voluntary realignment of one’s own Measure landscape; ambient risk is stochastic background attributable to no agent. The invariant is one line: if your action worsens another agent’s Measure landscape without consent, you have committed coercive harm.

The Firewall Against Probabilistic Exploitation

The structural consequences arrive with the same inevitability. A predator, in this framework, is not a moralistic label but a classification: an agent who systematically shifts others’ Measure toward negative-outcome futures for instrumental gain. Such an agent is not evil in any supernatural sense; it is incompatible — its behavior violates the invariant that keeps the space of coexisting agents stable. By knowingly and materially degrading others’ distributions, the predator performs its own exit from the protection that shields innocents, and defensive coercion against it becomes permissible — not as revenge but as a Measure-preserving operation, under the constraints already built in The Boundaries of Force. What stable multi-agent coexistence requires, and how the requirement generates an ethics rather than merely a truce, is the work of The Ethics of Viability; what an agent owes the distribution of futures it steers, including its own, is the subject of Measure Responsibility.

Here the verdict suffices. An ethics built for a world of uncertainty forbids gambling with another agent’s future — no matter how noble the motive, no matter how dire the circumstance, no matter how kindly the dice happen to fall. It refuses to let anyone weaponize probability, turning another person’s possible death into a resource. You may risk your own branch of futures. You may not risk theirs. Consent transforms risk into shared agency; without it, imposed risk is coercion at the moment the door clicks shut, whoever walks out afterward.

Harm is not what you feel. Harm is what you do to the distribution of futures. That is the sharp line that survives contact with physics.