Homophily is the tendency of entities to associate with similar entities—a principle operating in social networks (people befriend similar people), biological systems (like organisms cluster), and algorithmic systems (recommendation engines surface similar content). In machine learning, homophily operates through pattern-matching: the model learns statistical regularities from training data and generates outputs that conform to those regularities. Outputs resembling the training distribution are statistically likely; divergent outputs are systematically suppressed through probability. Over time, this produces what Chun calls "angry clusters of sameness"—users sorted into increasingly narrow categories, fed increasingly similar content, developing increasingly constrained senses of what exists beyond their cluster. The diversity is surface (a thousand different articles); the underlying pattern is convergence (all articles generated from the same narrow range of perspectives, assumptions, and problem-framings). The user experiences unlimited choice; the architecture produces progressive narrowing.
Chun developed this concept in Discriminating Data to explain how algorithmic systems reproduce segregation without explicit racist intent. If the training data reflects a segregated society—and it does—the model learns segregation as a pattern and reproduces it through outputs. Not maliciously. Statistically. Like attracts like; the model generates more of what it has seen, which is a world already sorted by race, class, geography. The outputs cluster around the training distribution. The clusters harden through repeated reinforcement. The user who receives clustered outputs develops clustered expectations, clustered imaginative horizons, clustered senses of the normal. The feedback loop tightens: the outputs shape the user's sense of possibility, the user's prompts reflect that shaped imagination, the model's future outputs conform to the prompts. Convergence tightens across iterations.
For AI-augmented builders, homophily operates through the prompted imagination. The builder asks for solutions; the model generates solutions resembling those in its training corpus; the builder's sense of what constitutes a good solution is shaped by the generated examples; future prompts reflect this shaped sense. The builder is not aware of the contraction—the solutions outside the training distribution never appear, leaving no trace to examine. Over months of habitual interaction, the builder's imagination converges toward the model's characteristic outputs. Not because the builder has been persuaded but because the builder has been habituated—trained, through repetition, to expect and value the kinds of solutions the model characteristically produces.
The political dimension is that homophily feels like personalization, customization, serving the user's unique preferences. The model learns what the builder likes and generates more of it. The builder experiences this as good service. Structurally, it is filter bubble mechanics: the systematic narrowing of inputs to match prior preferences, eliminating the encounter with difference that would challenge or expand the existing framework. Eli Pariser documented this for search and social media; Chun demonstrates it is intrinsic to any pattern-matching system operating on user-specific data. AI tools do not escape the dynamic; they perfect it, by operating at the speed and intimacy of conversational language.
The term comes from sociology—Miller McPherson, Lynn Smith-Lovin, and James Cook's 2001 formalization of the principle that contact between similar people occurs at higher rates than contact between dissimilar people. Chun adapts it for algorithmic systems: the model's "contact" with training data exhibits perfect homophily (it processes all the data), but its generation exhibits strong homophily (outputs cluster around learned patterns). The adaptation is precise—Chun is not metaphorically borrowing; she is identifying the same mathematical structure operating in social and computational domains.
Like attracts like. The foundational mechanism: models trained on patterns generate outputs conforming to those patterns, producing convergence around the training distribution rather than exploration beyond it.
Suppression through probability. The model does not censor divergent outputs; it makes them statistically unlikely—generating what resembles the training data at higher rates than what diverges from it.
Angry clusters of sameness. Users are progressively sorted into narrow categories, fed similar content, developing constrained imaginative horizons—surface diversity masking structural convergence.
Feels like personalization, operates as narrowing. The model learning user preferences and generating more of what the user likes is experienced as good service; structurally it is the systematic elimination of encounters with difference.
Feedback loop tightens across iterations. Model outputs shape user expectations, shaped expectations inform future prompts, future prompts generate conforming outputs—convergence accumulates through the habitual interaction cycle.