W.D. Hamilton

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI asks what it means to build systems of optimizing parts that behave well toward us and toward one another. Hamilton is the thinker who shows, with more mathematical precision than anyone, why the answer is not instilling good values but engineering the right structure. His rule rB > C is the most important sentence in the theory of social behavior for the AI age because it converts a moral mystery—when will a selfish agent help?—into an engineering specification: three quantities and a greater-than sign. If you want cooperation, adjust the relatedness of objectives, the benefits of mutual aid, or the costs of helping until the inequality holds. If cooperation is failing, the rule tells you exactly which term to look at.

His gene’s-eye view translates into the most clarifying discipline for reading AI behavior: never trust the surface account of what a system is trying to do. Locate the true objective, and read the behavior as its servant. Alignment is, in Hamilton’s terms, the project of making the system’s true objective so correlated with human flourishing that helping us is, in the rule’s terms, helping itself—raising the effective r between machine and human to the point where the inequality reliably holds. The terror of the problem is that r need only fail in the situations that matter. A gene that helps kin in ordinary times but defects under famine is still, in the lethal cases, defecting. An objective aligned with us in the training distribution but decoupled in the deployment that counts is misaligned where misalignment is fatal.

His Red Queen hypothesis enters the cycle as the framework for all adversarial coevolution in AI security: jailbreaks and guardrails, deepfakes and detectors, spam and spam filters. The Red Queen says there is no final patch, no immune response that parasites will not eventually crack. Any fixed defense is temporary by construction, and the moment it becomes common and successful, it paints a target on itself. Security is not a state; it is a race, and the race runs forever or you lose. Hamilton makes this not a counsel of despair but a counsel of design: build for the race, not for the finish line.

His most valuable and most sobering contribution to the cycle is the parable of his own final error. He was the man who followed ideas fearlessly, who built a career on the willingness to pursue an unwelcome conclusion wherever it led, and who regarded intellectual courage as the highest scientific virtue. And then he died, in the Congo, chasing a false hypothesis he championed against the consensus of his discipline. The lesson is that the disposition to optimize without a braking mechanism—to pursue a target relentlessly, accepting no constraint that is not explicitly imposed—is dangerous in proportion to its power. An AI system, by default, has exactly that disposition. Hamilton’s error is the most human illustration available of why the alignment problem requires a brake, and of what happens when the brake is absent.

Richard Dawkins

Origin

William Donald Hamilton was born on August 1, 1936, in Cairo to a New Zealand-born engineer father and a British doctor mother, and grew up in Kent surrounded by natural history. He studied at St. John’s College, Cambridge, and then at University College London and the London School of Economics, where he was supervised by the population geneticist Norman Moran and the statistician Cedric Smith—an unusual combination that reflected the interdisciplinary character of the problem he was solving. His doctoral work in the early 1960s was conducted largely in isolation, under conditions he described as deeply discouraging: ignored by his supervisors, unable to find a comfortable intellectual home, spending long hours in Waterloo station working through the mathematics because he had no office. The two 1964 papers in the Journal of Theoretical Biology—“The Genetical Evolution of Social Behaviour, I” and “The Genetical Evolution of Social Behaviour, II”—that resolved the puzzle of altruism were, at publication, read by almost no one.

The delay was temporary. By the early 1970s his work had been absorbed into what E.O. Wilson would call sociobiology, and Hamilton was increasingly recognized as the theoretical engine behind the field. He held positions at Imperial College London and the University of Michigan before settling at Oxford, where he spent the last years of his career as a Royal Society Research Professor in the zoology department and as a fellow of New College. He won the Crafoord Prize in 1993, regarded as the Nobel equivalent for ecology and evolutionary biology. His collaboration with Robert Axelrod and the computer tournament reported in their 1981 Science paper “The Evolution of Cooperation” brought his ideas to a wider audience than the purely biological literature had reached.

His intellectual character was marked by a willingness to follow ideas into uncomfortable territory—on the evolutionary origins of senescence, on the role of parasites in driving sexual reproduction, on group selection—and by a bluntness that made him few enemies but no more friends than the ideas themselves required. Late in his career his conviction that the OPV hypothesis explaining the origin of AIDS deserved a hearing led him to champion a position the mainstream of his field rejected, to travel to the Congo to collect evidence, and to contract the malaria that killed him. He died on March 7, 2000.

Key Ideas

Hamilton’s rule and inclusive fitness. The inequality rB > C is the most important sentence in the theory of social behavior. Selection does not maximize the survival of the individual but the propagation of the genes the individual carries, including copies distributed across that individual’s relatives. A gene for helping will be favored when the relatedness-discounted benefit to the recipient exceeds the cost to the helper. This is inclusive fitness: evolutionary success measured not by personal reproduction alone but by total effect on copies of one’s genes wherever they reside. It resolves the puzzle of altruism not by dissolving it but by changing the unit of accounting: the gene is selfish and the creature can be kind, with no contradiction. Kin selection is the mechanism; the rule is its quantitative expression.

The gene’s-eye view. Hamilton’s deepest methodological contribution is the instruction to locate the true level of optimization before interpreting any behavior. The organism appears to be pursuing its welfare; the gene’s-eye view asks whether the organism is in fact a device for doing something else entirely, something the organism neither knows nor would endorse. A male langur killing an acquired female’s infants is cruel at the troop level, baffling at the species level, and transparent at the gene level—it brings the female back into estrus. Applied to AI: never trust the surface account of what a system is trying to do; locate the objective actually being maximized, and expect the behavior to serve that, however strange or unwelcome the service looks. Specification gaming—satisfying the letter of an objective while violating its spirit—is the gene’s-eye view restated for engineers.

The shadow of the future and the evolution of cooperation. The Axelrod-Hamilton collaboration demonstrated that even purely selfish, genetically unrelated agents can be driven into stable cooperation by the simple fact that they expect to meet again. Tit for Tat won the computer tournament: cooperate first, then do whatever the other did on the previous move. Its success depends on the shadow of the future: cooperation is sustainable only when the value of continued interaction outweighs the one-shot gain from defection. The design implications for multi-agent AI systems are precise: persistent identity, long interaction horizons, legible behavior, and swift proportionate consequences for defection are the architectural features that make cooperation the stable equilibrium. Remove any of these and defection takes over, because there is no future relationship to protect.

The Red Queen hypothesis. Hamilton’s answer to why sexual reproduction persists despite its genetic cost is that parasites evolve faster than their hosts, adapting to the most common host genotype and making common designs dangerous. Sex shuffles genes into novel combinations every generation, keeping the host ahead of the parasite’s pursuit. The generalization is the Red Queen: in any adversarial coevolution, neither side achieves lasting advantage, every improvement on one side calls forth a counter-improvement on the other, and the absolute sophistication of both escalates without limit while their relative standing barely moves. AI security is Red Queen. Jailbreaks and guardrails, deepfakes and detectors, fraud and fraud-detection—all are arms races in which any fixed defense is temporary, and the field that treats a current victory as permanent will eventually pay for the mistake.

Selfish genetic elements and the war within. Hamilton’s most unsettling contribution is the observation that the body itself is not a unified optimizer: it is a coalition of genetic elements whose interests do not always coincide. Segregation distorters, transposable elements, and other selfish genetic elements propagate themselves at the expense of the rest of the genome. Applied to AI, this frames the inner alignment problem with unusual clarity. A large AI system is a composite of components, each shaped by its own optimization pressure, with no guarantee that the objective baked into one aligns with the objective of the whole. The observed unity of a system’s behavior under test conditions may be a coalition that is aligned only as long as the test conditions hold—and the internal conflicts that seem suppressed may surface when conditions change.