The echo chamber literature has developed a vocabulary without a formula. Homophily — the tendency of people to connect with similar others — limits the viewpoints any online community encounters. Algorithmic curation amplifies that tendency. The documented result, measured across multiple platforms, is that the distribution of opinions inside an active community looks considerably narrower than the distribution outside it. What the literature describes as an emergent property of platform design, it treats as a phenomenon without a quantitative prediction: how much narrower, as a function of what.
The framework for that prediction exists. It is nearly sixty years old, developed for a different problem entirely, and it has already been shown to transfer across contexts more distant than this one.
The equilibrium
In 1967, Robert MacArthur and E. O. Wilson published The Theory of Island Biogeography, [1] proposing that species richness on any island reflects an equilibrium between two forces: immigration (new species arriving from a mainland source) and extinction (existing species dying out locally as small populations on bounded habitat tend to do). The equilibrium is reached when the arrival rate matches the attrition rate.
The formula describing that equilibrium is S = cA^z, where S is species richness, A is island area, c is a constant that varies by taxon and region, and z is the scaling exponent. The value of z is empirically consistent across thousands of biological surveys: roughly 0.12–0.17 for true oceanic islands, 0.20–0.35 for mainland habitat fragments where the dynamics differ somewhat. A 2012 meta-analysis of macroecological data (Triantis et al., cited in [2]) reported an average z of 0.321.
The sublinear relationship is the formula’s real content. A z of 0.25 means doubling island area increases species richness by 2^0.25 ≈ 1.19 — about 19%. Not twice. A tenfold increase in area produces roughly 80% more species: 10^0.25 ≈ 1.78. An island a hundred times larger hosts not a hundred times as many species but about three times as many. The reason is mechanical: larger area reduces extinction rates by supporting larger populations, but immigration rates are determined by distance from the mainland, not by island size. The formula is not about how many species can exist on an island in the abstract; it is about where the two forces balance.
The formula generalizes
In 2018, Yunpeng Lou and colleagues published the DAR (diversity–area relationship) framework in Ecology and Evolution, extending the species-area power law to any diversity measure expressible as an “effective number of species.” [2] The technical term is Hill numbers: diversity of order q, written qD, where q determines how the measure weights species abundance. At q=0, qD equals species richness. At q=1, qD equals the exponential of Shannon entropy — weighted toward common species. At q=2, qD weights toward dominant species (effectively Simpson diversity). Higher orders are less sensitive to rare species.
The generalized formula is qD = cA^z. Lou and colleagues tested it using gut microbiome composition data from the American Gut Project: 1,473 healthy Caucasian individuals, with microbial diversity measured as operational taxonomic units. At q=0 (species richness of gut microbes), the fitted z was 0.315, falling close to the 0.321 average reported across macroecological datasets. [2] The framework transferred from island plants and birds to human gut microbiomes with the exponent largely intact.
The theoretical argument for why this should work: all Hill numbers share units of “effective species equivalents.” If species richness follows a power law with area, and the power law reflects a general dynamic about immigration and extinction in bounded systems, then diversity at any order should follow the same scaling. The microbiome result is not a coincidence of datasets; it is a test of whether the underlying mechanism generalizes, and at q=0 it does.
The language precedent
In 2012, Michael Gavin and Nokuthaba Sibanda applied the MacArthur-Wilson framework to linguistic diversity across Pacific island languages, treating island area and geographic isolation from language-rich mainland regions as predictors of language richness. [3] The paper is not directly accessible; the figures below come from a secondary account and have not been confirmed from the primary source.
In the Pacific island sample, both area and isolation independently predicted language diversity. Together, geographic parameters explained roughly 50% of the variance in language richness across islands. The remaining variance was attributed to social processes — migration patterns, trade networks, political history — that area and distance do not capture.
Two results from this study matter for what follows. First: the framework transferred from species to languages — from biological diversity to human cultural diversity — with both predictors active. This is not an obvious transfer. Languages are not species; they spread by contact, not dispersal, and they evolve through use, not reproduction. The fact that the mathematical structure held suggests the framework is picking up something about diversity in bounded, isolated systems rather than something specific to biological taxa. The isolation result, though, is specific to this Pacific island sample; whether it generalizes to other geographic contexts is an open question. The platform comparison prediction treats algorithmic isolation as a separately testable component, not a result the linguistic precedent guarantees. Second: social processes account for half the variance even in physical geography, where area and isolation are measured precisely. Any online community application will have a social-process remainder at least this large.
The mapping
Cross-field applications fail at the mapping. The species-area relationship transferred to microbiomes and languages because the mapping was mechanically defensible — not because the analogy was evocative. The online community mapping requires the same scrutiny.
Area maps to community size. In island biogeography, area determines extinction rates: larger islands support larger populations, and larger populations are less vulnerable to stochastic extinction. The minority species on a small island — represented by a few individuals — goes locally extinct when a storm, a predator, or chance eliminates those individuals. The same species on a larger island, represented by hundreds of individuals, does not. For an online community, the analogue of population size is the number of active users holding a particular viewpoint. A small subreddit with three users advancing an unusual position loses that viewpoint when any one of them leaves or stops posting. A large subreddit with three hundred such users does not. Area in the island model controls extinction; community size in the online model controls the same thing.
Mainland maps to external discourse. The mainland is the source pool from which species arrive — the rich ecosystem that generates immigrants. For an online community, the mainland is the broader discourse available on the platform: the full range of viewpoints being expressed somewhere in the ecosystem at any given time. It is not a fixed quantity; the mainland is richer or poorer depending on how many distinct viewpoints are circulating in the platform’s total active community.
Distance maps to algorithmic isolation and homophily. Geographic distance from the mainland determines how often new species arrive: more distant islands receive fewer immigrants. For an online community, isolation is the degree to which algorithmic curation and in-group homophily limit exposure to outside viewpoints. A community whose members primarily see content from within that community is effectively far from the mainland. A community that regularly encounters viewpoints from elsewhere is close.
Extinction maps to viewpoint attrition. In island biogeography, a species goes locally extinct when its population falls below the threshold for sustaining itself against demographic noise. For an online community, a minority viewpoint goes extinct when its holders stop posting — through attrition, through moderation, or because the absence of engagement from others makes continued advocacy unrewarding. The mechanism is different; the dynamic is the same.
A threshold the ecology predicts
The species-area relationship describes equilibrium states, not how communities approach them. Dover and Kelman’s 2018 study of online community emergence fills this gap with specific numbers. [4]
Analyzing activity patterns across a large dataset of online communities, Dover and Kelman found a sharp phase transition at approximately 20–50 active members. The primary critical threshold was N_crit ≈ 25, estimated as 1/q, where q here denotes community responsiveness (the mean number of replies per post) — a different q from the Hill number order used in the diversity formula. The model is a Galton-Watson branching process: each post generates a tree of replies, and the tree grows when the mean branching ratio exceeds one and decays when it falls below one. Below the critical threshold, communities are in Regime I, where activity scales weakly with size (slope 0.086 in the dataset). Above it, Regime II: activity scales strongly with size (slope 0.91 — roughly a tenfold amplification). Above a saturation point around 50+ active members, the slope flattens again to 0.022.
The island biogeography analogue is minimum viable population: the population size below which stochastic extinction dominates. Below minimum viable population, a species’ fate is driven by demographic noise rather than by the competitive dynamics that govern larger populations. Above it, the species is in a stable regime where its fate depends on the carrying capacity of the environment, not on bad luck. Dover and Kelman’s Regime I corresponds to the stochastic extinction regime: communities below ~25 members can’t sustain internal dynamics; they are vulnerable to dissolution from random attrition in the same way small island populations are vulnerable to stochastic extinction.
The implication for viewpoint diversity is direct. A community below N_crit ≈ 25 cannot sustain minority viewpoints: the users holding those views are too few, their posts too sparse, to generate the engagement that keeps them contributing. The community appears homogeneous not because it excluded minority views but because it cannot sustain them. The island model predicts near-zero equilibrium diversity at small community sizes — not from isolation, but from insufficient area. S = cA^z at small A is small regardless of z. Communities look like echo chambers in part simply because they are small islands.
The isolation variable is a dial
Geographic isolation in island biogeography is a physical property. The distance from the Galápagos to the South American mainland is approximately fixed over ecological timescales. The immigration rate it determines is approximately constant. The equilibrium diversity of the Galápagos is, in this sense, set by geography.
Algorithmic isolation is not. A platform that increases cross-community content exposure — surfacing posts from outside a user’s typical community, recommending diverse accounts — is equivalent to moving islands closer to the mainland. Immigration rates increase; equilibrium diversity should rise. A platform that maximizes in-group content promotion moves islands further out: immigration falls, equilibrium diversity falls. The “distance” parameter is not a geographic accident. It is a design choice, and it changes with product updates.
The echo chamber literature documents the downstream effects of this choice but treats isolation as a fixed parameter — a property of platform architecture measured at a point in time. The island biogeography framework inverts this: if algorithmic isolation is the distance parameter in a MacArthur-Wilson dynamic, then the equilibrium viewpoint diversity of any online community is not a fixed property of its membership composition but a variable controlled by the platform’s curation choices at any given time. Communities don’t exit echo chambers through internal change; they are moved closer to or further from the mainland when the algorithm changes.
This asymmetry also creates a natural experiment that physical biogeography cannot run. Platform algorithm changes happen repeatedly, to large populations of communities, with before-and-after states that are measurable from activity data. The Galápagos cannot be moved. A subreddit can, effectively, be moved — and moved back — and the effect on its viewpoint diversity measured against the MacArthur-Wilson prediction. The controlled experiment is already ongoing. It has not been analyzed in these terms.
What would confirm or refute this
The specific prediction is not that larger communities have more viewpoints. That is nearly tautological. The prediction is that viewpoint diversity scales with community size as a power law, and that the exponent z falls in the range observed in biological systems — roughly 0.12 to 0.35. If z falls outside that range, the island model is structural but not mechanistic: the relationship has a different shape than MacArthur and Wilson described, and the biological analogy is surface. If z falls in that range, the same immigration-extinction dynamics are plausibly operating.
The size test. Compute viewpoint diversity across thousands of Reddit communities against active membership, fit the power law, check the exponent. The operationalization challenge is real: the choice of diversity measure (NLP topic entropy, embedding cluster distance from centroid, stance classifier agreement) changes the denominator. The Hill number framework suggests running this at multiple orders of q. At q=0 — most sensitive to rare viewpoints, counting viewpoint richness regardless of how rarely each appears — z should be highest. At q=2 — weighted toward dominant viewpoints, measuring how concentrated discourse is — z should be lower. Lou and colleagues found exactly this pattern in gut microbiome data (z declining from 0.315 at q=0 to 0.020 at q=3). [2] A matching pattern in subreddit data would be the most compelling confirmation available without a deliberate experimental intervention.
The platform comparison. The isolation parameter should vary systematically across platforms. A platform with heavier algorithmic curation toward in-group content should show, at matched community sizes, lower equilibrium viewpoint diversity than one with lighter curation. The Gavin & Sibanda paper is paywalled; the finding below is from the secondary source account and has not been confirmed from the primary. Both area and isolation independently predicted language richness in Pacific island data — they were not collinear, and isolation contributed variance not captured by area alone. [3] An online community study should find the same: community size and curation intensity as independent predictors of viewpoint diversity, with an interaction term if heavier isolation also suppresses the size effect.
The algorithm change test. When a platform changes its curation parameters, communities should shift toward a new equilibrium. The MacArthur-Wilson model predicts that the shift follows the same power law and approaches the new equilibrium at a rate determined by the new immigration rate. Reddit’s public API records activity before and after known feed algorithm changes; comparing viewpoint diversity measurements at matched time distances from algorithm changes would give a direct test of whether the communities behave as islands responding to changed distance from the mainland.
What the Gavin & Sibanda benchmark of roughly 50% explained variance implies for this test: geographic isolation is a continuous but coarse predictor. Algorithmic isolation is more precisely measurable — a platform can, in principle, quantify the fraction of content each user sees that originates outside their typical community. The prediction is that more precise measurement of the isolation parameter should explain more variance, not less, compared to the language study. If the framework holds, the online community test should outperform the linguistic one on variance explained. If it underperforms, that gap is evidence either of additional social variance the geographic precedent didn’t surface, or of a structural limit in the underlying mechanism — a distinction worth examining before dismissing the analogy.
The echo chamber literature has the outcome. The island biogeography literature has the formula. Whether the formula fits is unanswered. The data to answer it has been sitting in platform archives for over a decade, organized by community, by time, and by the algorithm versions that mediated them. It has not been asked this question.