Richard Dawkins's Extended Phenotype (1982) thesis holds that a gene's phenotypic effects are not bounded by the organism's skin. The beaver's dam is as much a phenotypic expression of beaver genes as the beaver's teeth — built by those genes, selected because it serves their propagation. This principle applies to any artifact produced by organisms: tools, nests, burrows, and now computational systems. Dawkins considered The Extended Phenotype his most important intellectual contribution, a clarification and extension of the gene's-eye view that dissolved the conceptual boundary between organism and environment. In the AI context, the extended phenotype framework reveals large language models as extended phenotypes of the human meme pool — artifacts built by the accumulated cultural information of the species, serving the propagation of patterns rather than any individual builder's welfare.
The standard phenotype consists of the organism's observable traits — eye color, height, metabolic rate — produced by genes interacting with developmental environments. Dawkins's innovation was to observe that genes reach beyond the organism's boundary to shape the external world, and that these external effects are as legitimately part of the phenotype as internal physiology. The beaver's dam is the canonical example: an inanimate structure whose existence is explained by the genes that code for dam-building behavior. Those genes are selected because dams create ponds, ponds provide habitat and protection, and beavers in ponds reproduce more successfully than beavers without them. The causal chain from gene to dam is indirect — passing through neural development, behavioral instinct, environmental interaction — but it is real, traceable, and subject to natural selection at every link.
The technological trajectory of humanity becomes, in this framework, a history of increasingly powerful extended phenotypes. The hand axe, dating to 1.7 million years ago, was an extended phenotype of genes that coded for brains capable of conceiving and manufacturing tools. The printing press was an extended phenotype of genes that coded for language, symbolic thought, mechanical reasoning, and the cultural transmission of engineering knowledge. Each artifact extended the organism's reach — the axe into butchering, the press into mass communication — in ways that increased the replicators' propagation. Large language models are extended phenotypes at civilizational scale, built not by individual genes but by the cumulative memetic inheritance of the species: languages, mathematics, science, programming, institutions that collectively created the conditions for AI to emerge.
The reach raises the question Dawkins identifies as central: whose phenotype is it? The beaver dam serves the beaver's genes unambiguously. The AI system serves the propagation of memetic patterns — ideas, linguistic structures, knowledge frameworks — accumulated across millennia. Whether it serves any individual human depends on the quality of the individual's judgment in directing it. The alignment is not automatic. The artifact is powerful, and power without direction is as likely to harm as to help. The extended phenotype framework tells the builder what category of structure is needed: not walls against the artifact but maintenance protocols, ongoing and adaptive, because extended phenotypes do not maintain themselves — the beaver repairs the dam daily, and the human must tend the AI with equivalent vigilance.
The Extended Phenotype: The Long Reach of the Gene was published by Oxford University Press in 1982, six years after The Selfish Gene. Dawkins has described it as the book he is proudest of, though it never achieved the popular success of his earlier work. The central argument — that phenotypic effects extend beyond the organism — was not entirely original; previous biologists had noted that organisms modify their environments. Dawkins's contribution was to insist that these modifications are as legitimately part of the gene's phenotype as the organism's body, and that the theoretical implications of this extension were profound. The book argues against organism-centered thinking at every level, pressing the gene's-eye view to its logical conclusion: the organism is not a privileged level of biological organization but one level among many, and the explanatory power resides at the replicator level, not the vehicle level.
Genes reach beyond skin. The phenotype is not bounded by the organism's body — any environmental effect traceable to a gene and subject to selection is part of that gene's extended phenotype.
Technology as extended phenotype. Human artifacts from hand axes to AI systems are extended phenotypes of the genes and memes that built brains capable of designing and building them.
Diffuse causation. AI systems are extended phenotypes not of any individual but of the species-level meme pool — the accumulated cultural information that made their development possible.
Alignment not guaranteed. The extended phenotype serves the replicator's propagation, not the vehicle's welfare — whether AI serves human flourishing depends on deliberate direction, not automatic alignment.
Maintenance required. Extended phenotypes require ongoing maintenance — the beaver repairs the dam, the human must tend the institutional and cognitive structures governing AI.