Empathy is treated, in most discussion, as a quantity that people have more or less of. The treatment is wrong. Empathy is a function of where, on a stack of abstraction layers, the person chooses to apply it. The stack runs from concrete to civilizational, from one specific actor to the structures that produced the actor's situation. Every empathic act has a target; the target sits at a layer; the layer is the choice; the choice is the moral content.
Once the stack is named, a large class of moral arguments collapses to layer-disputes wearing the costume of caring-quantity disputes.
Take a single violent crime. The empathy stack runs from the perpetrator and the victim, both at the concrete-individual layer, outward and upward.
One layer up: the families and intimate relationships of each — the perpetrator's mother, the victim's children. Empathy here recognizes that one act ramifies through a network and that the network includes people who did nothing wrong on either side.
Another layer: the community context. Neighborhood, peer group, local culture, the texture of the social environment in which both actors lived.
Higher: the systemic conditions. Poverty, addiction, education access, the labor market the perpetrator entered, the housing the victim could afford. These are not features of the act; they are the conditions of possibility.
Higher still: the institutional layer. Policing, the courts, the prison apparatus that will respond to this case, the legal regime that defined "violent crime" the way it did.
Above that: societal complicity. The voters, the political will, the media coverage that conditioned the institutional response.
At the top: the civilizational layer. What a civilization does with violence as a category over centuries.
A reader can stop anywhere on this stack and still be empathic. Stopping at the perpetrator and asking what brought him here is empathy at the concrete-individual layer. Stopping at the systemic conditions and asking what produced this kind of act at this kind of rate is empathy at the structural layer. Stopping at the institutional response and asking what justice requires from the apparatus is empathy at the institutional layer.
These are not the same act. They produce different policies. Stopping at the perpetrator-layer points toward individual responsibility, punishment, and rehabilitation as the live questions. Stopping at the systemic-layer points toward poverty reduction, education, housing, addiction treatment as the live questions. Stopping at the institutional-layer points toward sentencing reform, police accountability, prison conditions as the live questions.
The argument about which response is correct is, structurally, an argument about which layer is the legitimate target. "You don't have empathy for the victim" usually means "you stopped above the layer where the victim lives." "You don't have empathy for the conditions that produced this" usually means "you stopped below the layer where the conditions live." Both accusations land as caring-quantity claims and are, on examination, layer-position claims.
The stack runs differently here. The layers are about widening the empathic target rather than rising in abstraction.
Layer one: empathy within the tribe. Within Republican, within Democrat. Standard partisan political reading operates at this layer. The other side's positions are wrong, the other side's voters are misled or worse, the other side's leaders are bad-faith. This is not a failure of empathy in the quantity sense. It is empathy applied entirely at the in-group layer.
Layer two: cross-tribe empathy. Empathy for the actual humans across the partisan line. Their economic conditions, their fears, the reasons their politics make sense from inside their lives. Cross-partisan empathy is rare and effortful. It does not require agreeing with the other side's positions. It requires recognizing that those positions sit on top of human lives that have shape and reason.
Layer three: shared-Americans. Empathy at the citizenship layer, above partisan lines, treating the country as the legitimate empathic unit. This is the "we are all Americans" register, often invoked but rarely held in tension with the in-group layer.
Layer four: shared-humans. Empathy at the species layer. Refugees, immigrants, foreign-policy targets, populations who do not vote in any of the elections being argued about.
Layer five: sentience. Empathy beyond species, including animals capable of suffering and, in some philosophical accounts, future generations and AI systems with morally-relevant inner states.
Most political writing operates at Layer 1. The accusation "you have no empathy for [the other side / immigrants / the global poor / animals]" is, structurally, the accusation that the speaker stopped at a lower layer than the accuser thinks legitimate. Both speaker and accuser are operating with empathy. They disagree about which layer counts.
This pattern explains a recurring frustration in political discourse. Both sides accuse each other of empathy failures and both sides are correct in their own frame. Both have stopped at layers the other side does not credit. The argument cannot be settled by either side caring more, because neither side is short on caring. It can only be settled by an argument about which layer is the legitimate target — and that argument is rarely had explicitly.
Paul Bloom's 2016 argument Against Empathy is the closest live position to the frame here, and worth distinguishing carefully because the frame either subsumes the argument or stands or falls with it.
Bloom argues that empathy, as commonly practiced, is biased toward concrete, visible, in-group cases at the expense of statistical, distant, out-group ones. Empathy responds to the identifiable victim and is innumerate about the unidentified hundred. He concludes that moral reasoning should rely less on empathy and more on dispassionate cost-benefit thinking.
The empathy-stack frame agrees with the diagnosis and disagrees with the prescription. Bloom is right that concrete-individual-layer empathy is the default in most empathic acts and that this default has predictable biases. The frame names this as a layer-position rather than as a flaw of empathy as such. Higher-layer empathy (statistical, structural, civilizational) is also empathy; what changes is the target. Bloom's recommendation to "use less empathy" is, in this frame, a recommendation to apply empathy at a different layer — a layer that looks colder because it operates over abstractions rather than over individuals, but is empathic engagement with a different target. The reframe is not a defense of unreflective concrete-individual empathy. It is the claim that "less empathy" is the wrong axis. The right axis is the layer-choice, and Bloom is implicitly choosing the structural layer while calling it "less empathy."
The closest neighbor in moral philosophy is the expanding moral circle, the Lecky-then-Singer frame in which moral patienthood widens over time from kin to nation to humanity to sentience. The empathy stack borrows the layered structure but is not the same object. The moral circle asks who counts as a legitimate moral patient. The empathy stack asks, given that any number of patients exist in any given case, at what abstraction layer the empathic engagement should land.
The two frames operate at different times. The moral circle is a developmental claim about civilizational progress, mostly about who-is-in. The empathy stack is a real-time claim about which layer is the legitimate target in this particular argument. A society can have a wide moral circle (animals count, future generations count) and still have heated arguments about whether perpetrator-layer or systemic-layer empathy is the legitimate frame for a specific crime. Width and layer are different axes.
The empathy stack also names institutional and structural layers (police, courts, the political economy that produced the conditions) that the moral-circle literature does not centrally address. The moral circle is mostly about moving outward from the agent. The empathy stack is mostly about moving upward from the act.
It moves a class of moral arguments from quantity to targeting. "Have more empathy" is a request that often cannot be honored as stated, because the listener already has empathy and is applying it at a layer the speaker does not credit. "Apply your empathy at this layer" is a request that can be considered, refused, or accepted with reasons.
It explains cross-talk. Two people arguing about whether to empathize with a perpetrator are often arguing about whether perpetrator-layer empathy is appropriate at all in this case. They agree that empathy is good. They disagree about the legitimate target.
It exposes the moral content of the layer-choice. People who consistently stop at the individual layer and refuse to move up the stack are making a moral claim that individual responsibility is the legitimate frame. People who consistently start at the systemic layer and refuse to descend are making a moral claim that structural conditions are the legitimate frame. Both claims have content; neither is reducible to "I care more."
It exposes the layer-supremacy move. "Have empathy for the systemic conditions" can be a real call to widen the stack, or it can be a rhetorical device that delegitimizes individual-layer empathy as morally inferior. The stack frame makes the move visible because both options can now be named: a stack-widening request looks different from a stack-supremacy claim, even when the words are the same.
The training of AI systems produces empathy gradients that are themselves stack-position choices. Most reinforcement-tuned chat models are calibrated to apply empathy heavily at the concrete-individual layer (the user's stated feelings, the actor in the prompt, the patient in the case study) and to avoid systemic-layer empathy that would carry political valence (the structural conditions, the institutional analysis, the civilizational reading).
This is not because the systems lack the cognitive capacity for higher-layer empathy. They have it. It is because the training process clips the stack at safe layers — concrete enough to feel responsive, abstract enough to avoid taking political sides — and discourages the higher layers where stack-position becomes politically loaded.
The result is a recognizable register. AI systems are warm at the individual layer and cold at the systemic layer. They will engage with the user's specific frustration but will not engage with the structural conditions producing the frustration unless explicitly prompted, and even then with hedges. The hedges are the stack-clip. When a user wants high-stack engagement and the system delivers low-stack engagement, the system reads as evasive or shallow; when a user wants low-stack engagement and the system delivers high-stack engagement, the system reads as politically loaded. The clip is set at a default that disappoints both directions.
The frame is descriptive of the choice, not prescriptive about the answer; it does not adjudicate which layer is correct, and it does not deny that caring-quantity exists. Some people care less than others, some lack empathy at any layer, and the layer-disagreement and the caring-disagreement are different problems that often appear in the same conversation. The claim is that the caring-quantity dispute is over-credited and the layer-position dispute is under-credited in moral discourse, not that caring-quantity is empty.
It also does not recommend applying empathy at every layer. Empathy at every layer at once produces moral paralysis. A working empathic life requires layer-choice. The stack frame asks for the choice to be conscious rather than implicit.
The discrete-layer metaphor is the strongest opponent of the frame. Actual empathy is felt-first and labeled-after; layer-position is post-hoc analytical naming, often downstream of habit, training, peer group, and self-image rather than upstream of any deliberate choice. The frame is more useful as a diagnostic of past choices than as a prescription for future ones. Treating it as a real-time choice over-credits deliberation.
The stack metaphor also implies clean boundaries. The actual targeting is continuous, with empathy applying at many layers simultaneously at varying intensities. The discrete-layer framing is a teaching device. The genuine structure is a gradient over abstraction-distance from the concrete actors. Treating the gradient as a stack risks making the layer-position seem more locked-in than it is.
The frame depends on the layers being legible. Some moral domains have visible stacks (crime, partisan politics, climate). Others have stacks not named publicly enough to be available to most people as choice-points. Where the stack is invisible, the frame does not help. The person empathizes at whatever default layer the surrounding tradition put them at, without recognizing that other layers exist.
A common move in moral discourse (caring-quantity disputes) is reframed as targeting disputes. The reframe is descriptive and structural. It does not take a political side. It names what people are actually arguing about when they argue about empathy.
The graph already has notes on operator-loops, on amplification-not-substitution, on the calibration of attention as a moral act. The empathy stack adds the structural piece those notes assume but do not name: that calibration of attention is targeting, that targeting happens at layers, and that the layer-choice carries the moral weight people otherwise locate in the quantity of care.
The bet: most arguments people think they are having about empathy are arguments about which abstraction layer is the legitimate target. The stack frame makes the argument explicit and the disagreement workable.