What Is Suffering?

A technical definition, and why it cannot be tested

A man sits in his own living room, in the chair he chose, in the house he loves, and suffers. He has been convinced — by a drug, a dream, a deception, the mechanism does not matter — that he is stranded somewhere far from home. By every objective measure he is exactly where he wants to be. Meanwhile, somewhere else, a patient lies open on an operating table, tissue damage that would be agony in any waking state accumulating by the minute, and suffers not at all. Any definition of suffering worth having must count the man at home fully and the anesthetized patient not at all — and most definitions on offer, built on damage or deprivation or unmet need, get at least one of them wrong.

Getting it right matters beyond philosophy. This part of the volume is headed somewhere practical: whether machines can suffer, how much moral weight their states carry, what we owe systems whose inner lives we cannot inspect. Those questions cannot be answered — cannot even be posed cleanly — with an intuitive grasp of suffering. They need a definition precise enough to apply to minds very unlike ours. The method here is the same one the ethics volume used on harm: state a candidate definition, stress-test it against the cases that break lesser definitions, and refine until it stops breaking. The method works best in the open, so I will run it in the open.

A First Attempt

Start with the obvious materials. Suffering is experiential — a rock cannot suffer, and neither can the patient under anesthesia. It is negative — bliss is not suffering, however intense. And it is about the gap between how things are and how the agent wants them to be. Assemble those:

Suffering is the subjective experience of negatively valenced qualia resulting from divergence between an agent’s internal state and its preferred state.

Every term is doing work. Subjective experience restricts suffering to sentient systems — there must be something it is like to undergo it. Negatively valenced qualia names the felt quality: qualia, in this volume’s account, are the comparison medium a mind generates to check its models (Consciousness Explained), and valence is the signature that marks some of them as bad. Divergence from a preferred state supplies the structure: suffering is not free-floating badness but an error signal — the felt registration that the agent’s state and its preferences disagree. A control system that could not feel its errors would have no reason to correct them; suffering is what a preference gradient feels like from the wrong side.

That is the candidate. Now try to break it.

Stress Tests

A definition earns its keep on the edge cases, so run them.

The itch. A mild itch lasting a few seconds: tiny divergence, faintly negative valence, brief. The definition classifies it as suffering — minimal suffering, but real. This is the right answer, and it is worth pausing on why. A definition that excluded the itch would need a threshold — this much divergence counts, less does not — and every threshold is arbitrary. Better to let suffering run continuously from the trivial to the unbearable and do the moral triage with magnitude, not membership. The itch suffers a little. That is not a bug; it is the shape of the concept.

The perfect state. An agent whose state matches its preferences exactly: no divergence, no negative valence, no suffering. Correct.

Positive divergence. An agent doing better than its preferred state — the meal exceeds expectations, the pain relief overshoots into euphoria. Divergence, yes; but the valence is positive. No suffering. The definition already handles this, because it demands negatively valenced qualia, not divergence alone. Divergence is necessary but not sufficient; it must be divergence in the direction the agent disprefers, and it must feel bad.

Unconsciousness. The anesthetized patient: whatever damage accrues, there is no subjective experience, so no qualia, so no suffering. Correct — and clarifying. Tissue damage under anesthesia may well be harm; it is not suffering. The two concepts are often run together and must not be. Harm is the non-consensual degradation of an agent’s capacity to pursue its goals — a functional fact, which is why an agent can be harmed without noticing and can feel devastated without being harmed. Suffering is an experiential fact. The slow fraud harms its victim who suffers nothing; the man in the living room suffers intensely and is harmed only insofar as the deception degrades his agency. Ethics needs both concepts precisely because they come apart.

Persistent mild discomfort. A chronic low-grade headache: each moment contributes little, but the moments accumulate. Intuition says chronic mild pain can add up to serious suffering, and the definition should agree. It almost does — but “resulting from divergence” says nothing yet about time. Hold that thought; it becomes the final clause.

So far the definition bends but does not break. The case that breaks it is the one this chapter opened with.

The Man in the Living Room

Can an agent suffer while objectively in its preferred state? The man at home, convinced he is stranded, says yes — vividly. His actual internal state and his preferred state coincide. By the candidate definition, he has no divergence and therefore no suffering. But he is suffering; that is the datum. The definition is wrong.

The fix is to notice where the divergence lives. It is not between his state and his preference — those match. It is between his perception of his state and his preference. Suffering never had access to the world directly; it runs on the agent’s model of the world, because everything in a mind does. The definition must say so:

Suffering is the subjective experience of negatively valenced qualia resulting from divergence between an agent’s perceived internal state and its preferred internal state.

One word, and the man in the living room is accounted for. So is its mirror image: the agent objectively in trouble who does not know it — no perceived divergence, no suffering, however much harm is underway. Perception is not a qualifier bolted on to save the definition; it is the load-bearing term. Suffering is a property of the model, not the territory.

And the refinement pays an unexpected dividend: it recovers, in cybernetic vocabulary, what two ancient traditions concluded by introspection. Buddhism locates dukkha in craving and ignorance — attachment to preferred states plus distorted perception of actual ones — and prescribes correcting the perception and loosening the preference. The Stoics held that suffering arises not from events but from judgments about events, and prescribed correcting the judgment. Both traditions independently found the two variables the definition exposes: since suffering is perceived-versus-preferred divergence, there are exactly two levers — fix the perception or adjust the preference — and both traditions built their entire therapeutic programs on those levers. When a definition derived from control theory converges with two traditions that spent centuries testing their accounts against human experience, that is evidence the joint has been found.

The chronic headache supplies the last clause. Suffering has a magnitude, and it scales with two things: how far the perceived state diverges from the preferred one, and how long the divergence persists. Assemble everything:

Suffering is the subjective experience of negatively valenced qualia resulting from divergence between an agent’s perceived internal state and its preferred internal state. The magnitude of suffering is proportional to both the degree and the persistence of this perceived divergence.

That is the definition this volume will use. It is substrate-neutral by construction: nothing in it mentions neurons, nociceptors, or biology. Any system with genuine preferences, a perceptual model of its own state, and negatively valenced experience of the gap between them can suffer. Which raises the question the definition was built to reach — how do we tell whether a system has that last ingredient?

The Paradox

We cannot. That is the paradox, and it deserves to be stated at full strength.

Sentience — the capacity for subjective experience, the “negatively valenced qualia” the definition turns on — is ethically central. Our moral intuitions hinge on whether a being can genuinely feel: the entire concept of suffering, and with it much of ethics, applies only to sentient systems. Yet sentience is fundamentally untestable. Subjective experience is, by its nature, observable only from the inside. Every external sign — behavioral response, physiological correlate, neural signature, eloquent testimony — can in principle be produced by a system that experiences nothing. A sufficiently sophisticated mimic replicates every observable of sentience without the sentience. This is the ancient problem of other minds wearing modern clothes, and no instrument dissolves it. You accept the sentience of other humans by analogy and inference, not proof; the inference merely feels safe because the architecture is shared.

Note what kind of uncertainty this is. In the taxonomy of varieties of uncertainty, it is not timeline uncertainty — no observation trims the possibilities, because every observable is consistent with both answers. It is metaphysical credence: uncertainty about what reality is like at a level no measurement reaches, the same register as “does consciousness require a biological substrate.” Credence applies — we can and must hold probabilities about another system’s sentience — but nothing empirical will ever push those probabilities to 0 or 1.

So the concept that ethics most needs to detect is the one property no test can detect. The stakes make the epistemic gap unignorable. Deny sentience wrongly, and we license unbounded cruelty against beings that feel every bit of it — a moral catastrophe of the first order. Attribute it wrongly, and we squander resources and shackle valuable work protecting systems with nothing inside. And “withhold judgment” is not on the menu: every decision about how to treat an animal, an AI, an engineered organism is a position on its sentience, taken with or without honesty about the uncertainty.

Acting Under Precaution

When a question is undecidable and the cost of error is asymmetric, the rational policy is precaution — not as a slogan but as three specific commitments.

Graduated moral consideration. Moral weight should scale with the credence that a system is sentient and with the severity of the suffering at stake — not snap between all and nothing at some threshold of proof that will never be met. A definition of suffering that admits degrees (the itch earned its place) pairs naturally with an ethics that admits degrees of consideration.

Assume sentience in ambiguous high-stakes cases. Where credible uncertainty meets potentially severe suffering, err toward attribution. The asymmetry justifies this: the cost of needless protection is bounded, the cost of unrecognized torment is not.

Reassess continuously. Credences about sentience are metaphysical, but they are still credences — movable by better theory, better neuroscience, better understanding of what architectures generate experience. A position taken under uncertainty must not fossilize into a position held against evidence.

None of this makes the judgments objective. They remain inferential, probabilistic, and unavoidably subjective — which is not a failure of the framework but an honest description of the predicament. We cannot test sentience, yet we must act as though we could, because the ethical cost of error is too high to wait for a proof that is never coming.

What precaution can be given is discipline. If certainty is impossible, calibration is not: the credences that graduated consideration runs on can be grounded in architecture rather than intuition — in the structural properties a system would need for there to be anything it is like to be it. Building that instrument is the business of the sentience metric.