Other-Regarding Preferences¶
Core Idea¶
An agent's preference function takes as input the joint outcome vector — what I get and what others get — moving from \(u_i(x_i)\) to \(u_i(x_i, x_{-i})\) with nonzero cross-arguments. The structural fact is only that the function depends on others' payoffs at all; the sign (altruism, spite, inequity aversion, reciprocity) is left open.
How would you explain it like I'm…
Caring About Their Cookies
Imagine you're handing out cookies, and you don't only think about how many YOU get — you also care how many your friend gets. Maybe you're happier when they get some too, or maybe you'd feel it's unfair if they got way more. Either way, what they get is part of what makes you happy or not.
Their Outcome Counts Too
Other-Regarding Preferences means that when someone decides what they like, they care not just about their own outcome but about what others get too. The way they care can point in different directions: they might want others to do well (kindness), want them to do badly (spite or envy), prefer things to be fair, or want to reward people who were nice and punish people who were mean. The key point isn't that everyone is kind — it's just that what others get is part of the picture at all. So the old idea of a person who only cares about themselves becomes one special case, not the only case.
Payoffs Beyond My Own
Other-Regarding Preferences is the structural commitment that an agent evaluates outcomes based not only on its own payoff but on the payoffs received by one or more other agents. The preference function takes in the joint outcome — what I get and what they get — and returns an ordering that responds to both, not just the self-referential argument. That dependence can run in several signed directions: positive (I prefer outcomes that benefit them — altruism), negative (I prefer outcomes that harm them — spite or envy), reference-comparative (I prefer outcomes close to a fair split — inequity aversion), or reciprocity-conditioned (reward cooperators, punish defectors). The structural fact isn't that any one of these holds for everyone, but that the preference function is generically not a function of own-outcome alone. Once you name this, the classic self-interested agent becomes a special case — the one where the cross-arguments have zero weight — and real populations aren't all clustered at that point. Importantly, it is not a claim that people are nice; it's the meta-claim that the preference function has other-people's-outcomes as inputs at all, leaving the sign and shape open.
Other-Regarding Preferences is the structural commitment that an agent's evaluation of outcomes is indexed not only on the agent's own payoff but on the payoffs received by one or more other agents. The agent's preference function takes as input the joint outcome vector — what I get and what they get — and returns an ordering that systematically responds to both arguments, not only the self-referential one. The dependence can run in any of several signed directions: positive (I prefer outcomes that benefit them — altruism, kin-directed sacrifice), negative (I prefer outcomes that harm them — spite, envy), reference-comparative (I prefer outcomes close to a fair allocation — inequity aversion), or reciprocity-conditioned (I prefer outcomes that reward those who cooperated and punish those who defected). The structural fact is not that any particular signed dependence holds across all agents, but that the preference function is generically not a function of own-outcome alone. Once this commitment is named, the canonical own-payoff-maximising agent becomes the special case — where the cross-arguments have zero weight — and the empirical distribution of agents across substrates is not concentrated at that point. What other-regarding preferences is not is a specific preference content: it is not the claim that people are nice, that they care about fairness, or that they reciprocate. It is the structural meta-commitment that the preference function has cross-arguments at all, leaving open what sign and shape those arguments take in any given substrate or population. Formally it is the move from a utility function of own-outcome alone to a function of both own-outcome and others' outcomes, with the structure of the dependence — which agents enter, with what signs, under what conditions — as the empirical object of interest.
Broad Use¶
- Behavioural game theory: dictator donations, ultimatum rejections, trust-game investments, and third-party costly punishment.
- Evolutionary biology: kin selection values relatives' reproductive success at a coefficient given by relatedness.
- Welfare economics: any social welfare function must specify whether agents' orderings depend on others' outcomes.
- Mechanism design: allowing other-regarding components changes truth-revelation and optimal bid functions.
- Political philosophy: the veil of ignorance as a procedure for eliciting other-regarding components.
- Multi-agent AI: reward functions with non-self-only arguments to support cooperation.
Clarity¶
Separates self-interest from preference content: the own-payoff-only agent is one measure-zero point in a parameter space, not a privileged default, so the empirical question becomes which point a population occupies.
Manages Complexity¶
Compresses a sprawl of named "anomalies" into a small parameter space — which agents enter the function, with what signs, under what conditions — that the same paradigm suite can estimate.
Abstract Reasoning¶
Asks, of any multi-agent setting, "whose outcomes enter the focal agent's evaluation, with what weight, and under what conditions?", a question that references only abstract roles and so ports across substrates.
Knowledge Transfer¶
- Economics → biology: an altruism parameter and a kin-selection relatedness coefficient are the same weight on another agent's outcome.
- Behavioural theory → AI: "target the cross-argument weights" becomes tuning a reward-sharing coefficient in multi-agent RL.
- Game theory → organisations: peer-comparison (inequity-aversion) weights explain pay-satisfaction effects.
Example¶
In the ultimatum game, the Fehr-Schmidt form \(u_i(x_i, x_j) = x_i - \alpha\max(x_j - x_i, 0) - \beta\max(x_i - x_j, 0)\) makes the cross-argument parametric; the own-payoff-only model (\(\alpha=\beta=0\)) wrongly predicts acceptance of any positive offer.
Relationships to Other Primes¶
Parents (1) — more general patterns this builds on
- Other-Regarding Preferences is a kind of Preference — The file + dedup flag: a CONTENT SPECIALISATION of preference — the move from u_i(x_i) to u_i(x_i, x_{-i}) with nonzero cross-arguments. preference is the umbrella; the own-payoff-only agent is recovered at zero cross-weights (a measure-zero point). NOT a reparent — preference is the genus, candidate is the more-SPECIFIC child.
Path to root: Other-Regarding Preferences → Preference
Not to Be Confused With¶
- Other-Regarding Preferences is not Preference because it is the specific content commitment that the ordering depends on others' outcomes, whereas preference is the umbrella ordering relation silent on what enters it.
- Other-Regarding Preferences is not Cooperation because it is a preference structure, whereas cooperation is a behavioural pattern that can arise from pure self-interest (the folk theorem) and that spiteful other-regarding agents can fail to produce.
- Other-Regarding Preferences is not Fairness because fairness is one positive-weight content (toward equal allocations), whereas other-regarding preferences also covers spite, envy, and norm-enforcement.