Which Art Endures? Funding Networks, Institutional Power, and the Political Economy of American Arts Philanthropy

The survival of art in America is not primarily an aesthetic question. It is a structural one. Behind every museum that keeps its doors open, every theater company that mounts another season, every artist residency that outlasts its founding director, there is a network of funding relationships—a web of private foundations, family trusts, corporate sponsors, and institutional donors whose collective decisions determine what gets made, what gets shown, and what endures.

This working paper argues that arts philanthropy operates as a legitimation economy—an allocation system in which production capacity is relatively abundant but institutional endorsement, relationship capital, and geographic access remain scarce. Legitimation dynamics have been studied in academic hiring,³⁵ art exhibition networks,¹¹ and film industry consecration,³⁶ but the IRS e-filing mandate gives us something those domains lack in the financial dimension: the largest available allocation ledger—every resolved grant transaction, every matched donor, every identified recipient—not a random sample but the most comprehensive financial graph available as panel data, covering an estimated two-thirds of arts 990-PF activity. Using IRS 990-PF filings from 2019 through 2024, we construct the bipartite network of American arts philanthropy and identify three structural parameters that govern how endorsement flows: a preferential attachment kernel (α ≈ 1.07) that concentrates future funding on already-endorsed institutions, an edge persistence rate (43–68%) that makes donor-recipient relationships stickier in arts than in other sectors, and a modularity coefficient (Q_excess ≈ 0.27–0.32 above bipartite null) that restricts endorsement to tightly clustered geographic communities. These parameters describe the standing architecture of cultural allocation—the wiring rules that will mediate AI abundance in creative production.

We build on and extend the methodology of Shekhtman and Barabási,¹ who analyzed IRS 990-PF data from 2010–2019 and identified locality, donor retention, and prestige as structural features of arts philanthropy. Our contribution is threefold: we replicate their findings through a period of unprecedented disruption (the COVID-19 pandemic and the IRS e-filing mandate), we document novel structural features not present in the prior literature (arts-specific structural distinctness, portfolio breadth effects, gateway donor survival gradients), and we frame these findings within a legitimation economy framework that connects the political economy of arts funding to broader questions about allocation under abundance.

The period 2019–2024 captures three structural shocks to the funding ecosystem that have no precedent in the modern history of cultural philanthropy.

Three Shocks

Between 2019 and 2024, the arts funding network underwent three stress tests—each probing the legitimation economy under different conditions.

The COVID-19 pandemic (2020–2021): demand shock

The pandemic shuttered performing arts venues, closed museums, and eliminated earned revenue for thousands of arts organizations simultaneously. This was a demand shock to the legitimation economy: it disrupted the flow of patronage, altered donor behavior, and tested the resilience of funding relationships built over decades.

The IRS e-filing mandate (2021–): observational expansion

The Taxpayer First Act required all tax-exempt organizations to file returns electronically, transforming our observational window from a biased sample of larger foundations to the most complete network yet assembled.² This is not a behavioral shock but a measurement revolution: we can now see the full legitimation economy, not just its most visible participants.

The post-pandemic restructuring (2022–2024): structural reorganization

As organizations rebuilt, the structure of the funding network did not simply revert to its pre-pandemic state. New patterns of giving emerged, geographic concentration shifted, and the donor base expanded—revealing whether the legitimation economy’s structural parameters are equilibrium features or artifacts of a particular era.

Structure of this paper

We first present our data, methodology, and theoretical context, then introduce the legitimation economy thesis and its three structural parameters. The ten findings that follow are grouped in three parts: Findings 1–3 replicate the original Shekhtman & Barabási results through the pandemic and e-filing mandate. Findings 4–6 document structural dynamics (network connectivity, concentration, preferential attachment, community structure). Findings 7–10 map art-form variation, geographic structure, arts-specific structural distinctness, and portfolio breadth as a novel predictor of donor survival. A synthesis reframes the preceding findings through the long-tail distribution that underlies them. The paper concludes by developing the legitimation economy framework fully and examining its implications for AI abundance in cultural production.

Data and Methodology

Our primary data source is the IRS Statistics of Income e-file dataset: every Form 990-PF (private foundation) filing from tax years 2019 through 2024, comprising approximately 22 GB of XML across 64 ZIP archives.³ From these filings, we extract grant disbursement records—each an edge in our funding network, connecting a donor foundation to a recipient organization.

The critical challenge in this work is entity resolution. The 990-PF XML includes recipient names and addresses but not, in most cases, recipient EINs (Employer Identification Numbers). To build the network, we must resolve free-text recipient names to canonical organizations in the IRS Business Master File (BMF), which contains NTEE classification codes.⁴ We then retain only grants to organizations classified under NTEE major group A (Arts, Culture, and Humanities).

Our entity resolution pipeline uses a three-tier strategy:

1. Exact normalized name matching with state constraint (75.6% of resolved grants)
2. Fuzzy matching using token-sort ratio at a threshold of 85, constrained by state (17.4%)
3. Nationwide fuzzy matching at a higher threshold of 90 for remaining unresolved names (3.7%)

This approach resolves 66.8% of grants to a BMF organization and identifies 39,307 arts grants for the 2019 tax year from 198,537 keyword-pre-filtered candidates. We validate against the original Shekhtman & Barabási dataset, which contains 101,130 grants for 2019.⁵

Coverage and known biases

Our keyword-filtered pipeline captures approximately 39% of the grants identified in the Shekhtman & Barabási baseline for 2019. This is a known lower bound. The keyword filter selects grants where the filer included arts-related terms in the purpose field. Grants to arts organizations described generically—“general operating support,” “annual contribution,” “unrestricted gift”—are systematically missed. This likely over-represents programmatic arts grants (those with explicit arts purposes) relative to general support grants. An unfiltered pipeline that resolves all 990-PF grants against the BMF before filtering by NTEE code has completed processing for all years (2019–2024). The current results use this unfiltered pipeline. All grant counts should still be understood as lower bounds due to entity resolution coverage (not all recipient names match a BMF record), and findings that depend on absolute grant volumes (e.g., art form funding levels) are more uncertain than those based on structural ratios (e.g., locality rates, retention rates).

2024 data completeness

Tax year 2024 filings are substantially incomplete. As of our data extraction, 2024 contains 72,823 grants versus 84,532 in 2023—a 14% shortfall. IRS 990-PF filings are due 4.5 months after fiscal year-end, with extensions to 10.5 months, meaning calendar-year filers’ 2024 returns may not appear until late 2025. Late filers tend to be smaller foundations, so the incomplete 2024 data is biased toward larger, earlier-filing organizations. Throughout this paper, 2024 figures are reported but should be treated as provisional. Core conclusions rest on the 2019–2023 panel, where filing completeness is established. Where 2024 data appears to diverge from trend (e.g., the edge persistence drop in Table 5, the Q1 persistence collapse to 19.3%), we flag filing-lag incompleteness as the most likely explanation rather than interpreting it as structural change.

Entity resolution validation and statistical framing

The fuzzy matching tiers (state-constrained at 85, nationwide at 90) constitute 21% of resolved grants and are vulnerable to false positives (e.g., “Cleveland Museum of Art” matching “Cleveland Museum of Natural History”). To quantify error rates, we drew a stratified sample of 500 matched pairs, cross-referenced each against the IRS Business Master File (name similarity, NTEE code, geographic consistency), and hand-coded all 17 flagged cases. Precision and recall estimates follow below.

Automated consistency checks and a stratified validation sample of 500 matched pairs provide partial validation. Of 30,914 unique recipient names in 2022, only 136 (0.4%) map to multiple EINs—and these are genuine disambiguation cases (e.g., “Museum of Contemporary Art” resolves to four distinct institutions in four cities). In the reverse direction, 28.7% of resolved EINs have multiple name variants mapping to them—exactly what entity resolution should produce (e.g., the Metropolitan Museum appears under 32 spelling variants, all correctly resolved to EIN 131624086). The match method distribution is 80.0% exact-state, 13.7% fuzzy-state, 3.0% exact-only, 2.3% fuzzy-national, and 0.9% exact-multi, with 87.5% of all matches at confidence 1.0.

Cross-referencing the 500-sample validation set against the IRS Business Master File yields an automated precision estimate of 86% (lower bound). Precision varies by match method: exact-state 92%, exact-only 96%, fuzzy-state 95%, fuzzy-national 93%, and exact-multi 71%. The vulnerable tier is not fuzzy matching but exact-multi—cases where short, generic names (“Arts Fund,” “Guild Inc”) match multiple BMF records. Of the 17 cases flagged by automated checks, hand-coding confirms 15 as genuine errors and 2 as false flags (Musical Instrument Museum and Halcyon, both correctly resolved despite low string similarity to their BMF legal names). All 15 confirmed errors are from the exact-multi tier, where generic short names collapse distinct organizations onto a single EIN. Updated precision: 88% (106 confirmed correct, 15 confirmed wrong, 379 passing automated checks but not individually verified—true precision could be lower if the automated checks have false negatives). To be explicit: of the 500 validation grants, 121 were hand-coded (106 correct, 15 wrong; precision = 88%). The remaining 379 passed automated consistency checks (state match, NTEE code plausibility) but were not individually verified. If the automated checks have a false-negative rate of 5%, true precision could be as low as 84%; at 10%, as low as 80%. We use 88% as the operating estimate while acknowledging this uncertainty. Note: the weighted average of per-tier precision rates yields ~92%, which exceeds the 88% overall figure because the per-tier rates are computed within each tier’s validation subsample, while the overall figure includes the 379 unverified cases conservatively. To assess the impact of this error rate, we conduct two error propagation analyses across 1,000 Monte Carlo iterations. The uniform analysis randomly reassigns 12% of edges; the tier-stratified analysis applies measured error rates per match method (exact-state 8%, exact-only 4%, fuzzy-state 5%, fuzzy-national 7%, exact-multi 29%) with correlated errors—the same recipient name maps to the same wrong EIN across all years, simulating how entity resolution errors actually propagate. Under uniform perturbation, donor retention (79.3% → 79.3%, no shift) and recipient survival (88.6% → 85.7%, −2.9pp) are robust; edge persistence (68.4% → 55.8%, −12.6pp), locality (51.2% → 44.9%, −6.3pp), and concentration (top-100 share 43.4% → 40.5%, −2.9pp) are sensitive. Under tier-stratified correlated errors with same-state-biased replacement (simulating state-constrained matching), locality shifts modestly (51.2% → 51.8%, +0.6pp)—the bias toward same-state replacements mechanically preserves locality rather than degrading it, concentration shifts modestly (43.4% → 42.4%, −1.0pp), while persistence (68.4% → 62.7%, −5.7pp) and survival shifts are similar to the uniform case. The smaller shifts under tier-stratified errors reflect both the lower effective error rate (weighted ~7–8% vs. 12% uniform) and the fact that state-constrained replacement preserves rather than disrupts locality. Under both scenarios, locality remains far above the null model (~45–52% vs. ~6%), persistence remains substantial (~54%), and retention is unaffected. Findings that depend on locality magnitude are robust under both error models; retention and survival findings are robust to entity resolution quality at measured error rates.

This study is exploratory and descriptive, not confirmatory. We report over 80 distinct statistical comparisons across ten findings. Associations described below should be understood as patterns in the data, not as confirmed hypotheses. Where we report confidence intervals or z-scores, they quantify sampling uncertainty within a given comparison, not family-wise error rates across the full set of analyses. Applying Benjamini-Hochberg false discovery rate (FDR) correction at q = 0.05 to the 18 primary hypothesis tests with script-reproducible p-values,⁵⁹ 17 of 18 survive correction. The sole non-surviving test is the supply-shock chi-square (p = 0.55), which is reported as a null result consistent with the paper’s claim that the philanthropic network showed no differential response to COVID supply disruption. Core findings—locality above the null model (permutation p < 0.001), gravity same-state preference (PPML p < 10⁻⁵⁰), the gateway-donor survival gradient (logistic p < 10⁻¹⁰), portfolio-breadth retention (logistic p < 10⁻¹⁰), and arts-versus-non-arts structural distinctness (z-test p < 10⁻⁵⁰)—survive by wide margins. Statistical significance at these sample sizes (39K–85K grants per year) does not imply practical importance; we emphasize effect sizes throughout.

Theoretical context

Arts philanthropy is not a market. Donors do not purchase measurable outcomes; they confer institutional endorsement. The sociological tradition has long recognized this: Bourdieu’s field theory describes cultural production as governed by consecration—the power of certain actors to determine which works and institutions acquire legitimacy—and his Distinction (1984) showed how aesthetic hierarchies reproduce class structure through cultural capital.⁴² Becker’s Art Worlds (1982) frames artistic production as collective action embedded in cooperative networks of galleries, critics, patrons, and institutions. DiMaggio (1982) showed empirically how Boston elites used cultural institutions to construct and maintain class boundaries.⁷ White & White (1965) provided the foundational network-analytic study of institutional change in art markets, documenting how the dealer-critic system replaced the Academic salon system in French painting—a historical shift in the legitimation infrastructure that our data measures in its contemporary philanthropic form.⁶¹ Peterson & Anand (2004) synthesized these and related strands into the “production of culture” perspective, identifying six facets—technology, law, industry structure, organizational structure, careers, and markets—that shape what cultural products exist and who encounters them.⁵⁷ A complementary economic tradition explains why legitimation matters: Caves (2000) identified the “nobody knows” property of creative goods—fundamental uncertainty about which works will succeed—which makes institutional endorsement the primary signal of value.²³ Hirsch (1972) showed the organizational precursor: cultural industries manage this uncertainty through overproduction, gatekeeping, and co-optation of institutional intermediaries—foundations serve as “surrogate consumers” whose collective filtering decisions our network encodes.⁵⁵ Baumol & Bowen (1966) showed that performing arts face structural cost disease (rising labor costs without commensurate productivity gains), making them permanently dependent on patronage.²⁴ And Andreoni’s (1989, 1990) warm-glow theory of giving—that donors derive utility from the act of giving itself, not just its outcomes—is consistent with both the locality we observe (if psychological returns are higher for proximate recipients) and the COVID retrenchment pattern (if uncertainty increases the premium on visible, local impact).²⁵ Ostrower (1995) showed empirically that wealthy donors give to arts organizations primarily for social integration and status maintenance—the micro-behavioral foundation for the locality and prestige concentration our network data documents.⁶² Glazer & Konrad (1996) complement this with a signaling model: charitable giving signals wealth, and prestigious arts institutions serve as signal amplifiers—providing a game-theoretic micro-foundation for why arts philanthropy concentrates on visible, high-status recipients.³² Harbaugh (1998) extends this by showing donors respond to prestige categories (naming thresholds, giving levels) rather than continuous amounts—consistent with the sharp tier boundaries in our concentration data.⁵⁶ A comprehensive review of these and other giving mechanisms appears in Bekkers & Wiepking (2011).³¹

A complementary tradition in economic sociology addresses the evaluation problem directly. Karpik (2010)⁴⁶ showed that markets for ‘singular’ goods—multidimensional, incommensurable, uncertain in quality—depend on ‘judgment devices’ (critics, labels, networks of trust) rather than price signals. Podolny (2005)⁴⁷ demonstrated that status, inferred from network position, functions as a market signal that constrains competition and entry—the micro-mechanism behind our gateway-donor finding. Lamont (2012)⁴⁸ provides the umbrella framework for how evaluation practices create and reproduce social worth. Menger (2014)⁴⁹ complements Caves with a comprehensive treatment of artistic labor markets under uncertainty. Boltanski & Thévenot (2006) provide the foundational framework for how competing ‘orders of worth’ coexist in tension within evaluation systems—a perspective that enriches our single-metric legitimation framework by suggesting arts philanthropy may simultaneously operate under multiple justificatory logics (civic worth, market worth, inspirational worth).⁶⁰ Our contribution relative to these traditions is measurement: we quantify the infrastructure of judgment devices as a financial network with known error bounds.

Our contribution is to measure the financial trace of these processes at scale. The IRS 990-PF dataset encodes, in grant flows, the network through which institutional endorsement is conferred: which foundations fund which organizations, with what persistence, across what geographic boundaries. When a foundation makes a multi-year commitment to an arts organization, that is a measurable act of consecration in Bourdieu’s sense. When the same act recurs year after year across a cluster of co-funding foundations, that is an empirically observable legitimation network.

Three theoretical traditions inform how we interpret the structural parameters we measure:

1. Superstar economics (Rosen, 1981; Frank & Cook, 1995).⁸⁴⁵ Winner-take-all dynamics arise when small quality differences produce large reward differences—and Frank & Cook showed this applies to cultural markets specifically, where the gap widens as market scope expands. Our concentration findings—top-100 recipients absorbing 38–46% of arts dollars, with absolute funding rising even as share falls—are consistent with both models: AI supply expansion is precisely the kind of scope increase that amplifies winner-take-all dynamics.
2. Cumulative advantage (Merton, 1968;³⁰ Bol, de Vaan & van de Rijt, 2018).⁹ In science funding, early grant winners accumulate twice the funding of near-miss losers over eight years, with no evidence the gap reflects productivity. This provides the causal mechanism behind preferential attachment: early endorsement compounds. In arts philanthropy, Fraiberger et al. (2018) showed the analogous result for exhibition networks—artists who begin at peripheral institutions face 86% career dropout.¹¹ Cattani, Ferriani & Allison (2014) demonstrated the same mechanism in Hollywood, showing that core-periphery network position predicts consecration outcomes.³⁶ Our gateway-donor finding (Table 12) is the philanthropy-side equivalent.
3. Institutional isomorphism (DiMaggio & Powell, 1983).¹⁰ Foundations within a funding community converge on the same recipients through coercive (regulatory), mimetic (copying under uncertainty), and normative (professional norms) pressures. This predicts the strong community structure we observe (excess modularity Q_excess ≈ 0.27–0.32, on a bipartite network where even random graphs produce modularity of 0.53–0.63): foundations cluster into communities that fund overlapping recipients while maintaining sharp boundaries against other communities.

None of these frameworks are new, and the move to quantify legitimation with network data has precedent: Fraiberger et al. (2018) measured it in exhibition networks, Cattani & Ferriani (2014) in Hollywood, and Wapman et al. (2022) in faculty hiring.³⁵ What is new here is the completeness described above: not a sample of exhibitions or hires, but the full financial graph of a legitimation economy, measured over a six-year period that includes two exogenous shocks. The findings below should be read as empirical characterization of dynamics that Bourdieu described theoretically and that prior quantitative work began to measure in adjacent domains.

A fourth tradition—information cascades (Bikhchandani et al. 2024)⁵⁸—provides an alternative micro-foundation for the same macro-pattern. If donors observe which recipients attract funding and treat these signals as informative about unobservable quality, cascading produces near-linear attachment without requiring status-seeking or institutional isomorphism. Our data cannot distinguish PA-as-status from PA-as-cascade at the aggregate level; the arts vs. non-arts comparison (Table 11) provides partial leverage, since cascading would predict similar attachment patterns across sectors sharing the same information environment, while our legitimation framework predicts structurally distinct parameters for sectors with different evaluation regimes. The observed +9.1pp locality gap and +10.1pp concentration gap between arts and non-arts philanthropy are more consistent with sector-specific legitimation than with generic cascading, but a definitive discrimination test requires within-donor experimental variation that our observational data cannot provide.

A fifth alternative deserves mention: regulatory compliance. Private foundations face a mandatory 5% payout requirement, expenditure responsibility rules for grants to non-public-charities, and state attorney general oversight. These constraints could generate all four structural signatures without legitimation dynamics: persistence arises because switching recipients incurs new due-diligence costs; locality arises because state AG oversight creates a home-state compliance advantage; community clustering arises because foundations using the same compliance counsel converge on the same vetted recipient lists; and preferential attachment arises because well-known recipients satisfy IRS scrutiny more easily. A clean test would compare 990-PF private foundations (subject to these rules) to community foundations and public charities making arts grants (subject to different regulatory regimes). If the structural parameters persist across regulatory contexts, the compliance explanation fails. We cannot perform this test with the current dataset, which includes only 990-PF filers.

---

Thesis: Arts Philanthropy as a Legitimation Economy

The theoretical traditions above converge on the legitimation economy prediction introduced above: if institutional endorsement rather than market price determines which organizations survive, the funding network should exhibit three measurable signatures:

1. Legitimacy (preferential attachment): if arts philanthropy is a legitimation economy, past endorsement should predict future endorsement—a near-linear attachment kernel where well-connected recipients attract new donors at rates proportional to their existing connections.
2. Continuity (edge persistence): donor-recipient relationships should be stickier in arts than in sectors where value is market-determined, reflecting the relational nature of legitimation.
3. Translation power (community structure): funding coalitions should be tightly clustered, with the capacity to legitimize across geographic or institutional boundaries concentrating in a narrow set of actors.

Two pieces of evidence anchor this framework. First, arts philanthropy should be structurally distinct from philanthropy in general—more local, more concentrated, more edge-persistent—and these gaps should persist even after controlling for grant size (Table 11). Second, institutional endorsement should predict organizational survival: new recipients entering via high-centrality donors should survive at meaningfully higher rates than those entering via peripheral donors (Table 12). The findings below test these predictions. The concluding section reports the measured parameters and develops the framework fully.

---

Finding 1: The Hidden Network

The small foundations that make up 84% of arts donors were never isolated actors. They were embedded in the same legitimation network as the Mellon Foundation—connected through shared recipients, participating in the same endorsement infrastructure. The IRS e-filing mandate made them visible: when electronic filing became mandatory in 2021, the observed network nearly doubled in size, not because arts funding suddenly grew, but because we could finally see it.

The critical evidence is the largest connected component: it jumped from 85% to 94% of all nodes. The newly visible foundations were not isolates operating in independent ecosystems. They were already woven into the same network as the major donors, connected through shared recipients. The network was always this large; we simply could not see it.

A 94% largest connected component means the legitimation economy operates as a single connected market. The endorsement infrastructure that decides which arts organizations survive is one network, not many. Pre-mandate data systematically underrepresented small and mid-size foundations, but the structural finding is that they were participating in the same legitimation process all along.

Finding 2: COVID Was Associated with Increased Local Giving

The original study established that arts philanthropy is strongly local—approximately 60% of grants go to organizations in the same state as the donor, versus roughly 6% expected under a degree-preserving null model.¹ We replicate this finding across all six years and document a striking COVID-era pattern.

During the pandemic, same-state giving rose from 62.3% (2019) to 65.4% (2020)—a three-percentage-point increase that, given the sample size of 33,492 grants, is highly significant. This is consistent with foundations retrenching to their home states during uncertainty. We test whether this spike partly reflects differential attrition—cross-state donors exiting at higher rates than local donors, mechanically shifting the remaining pool. Among 2019 donors, pure cross-state donors (0% in-state giving) retained at 48.2%, while pure local donors (100% in-state) retained at 52.6%—a modest 4.4pp gap. The donors with the highest retention (62–67%) were those giving to multiple states, suggesting that geographic diversification proxies for institutional commitment. The locality spike thus combines a small compositional effect (less-committed cross-state donors exiting disproportionately) with genuine behavioral retrenchment by surviving donors.

The subsequent drop in 2021–2023 to approximately 52% is largely a composition effect. The e-filing mandate brought thousands of new foundations into the dataset. These newly visible mandate entrants initially gave less locally (40–43% same-state) than the pre-mandate visible set, even though small foundations overall are among the most local tiers (61–67% same-state; see Synthesis section). The distinction matters: the mandate entrants were not a random sample of small foundations but a specific population that had previously filed on paper, and their initial giving patterns differed from established small donors. By 2024, these new entrants converged to ~60% locality. After accounting for this composition shift, the underlying locality preference among pre-mandate foundations remained elevated through 2022.

Gravity model. The binary same-state measure is coarse. To test whether locality reflects geographic proximity alone or a genuine state-border premium, we estimate a Poisson pseudo-maximum likelihood (PPML) gravity model following Silva & Tenreyro (2006): flow_ij = exp(β₀ + β₁log(distance_ij) + β₂same_state_ij + α_i + γ_j), using state-centroid distances and grant flow counts for all 51×51 state pairs including zeros, with origin and destination fixed effects per Anderson & van Wincoop (2003). The same-state coefficient is large and stable: β_same = 3.85–4.13 across 2019–2024 (p < 10⁻⁶), corresponding to a same-state multiplier of 47–62× (upper bound) after controlling for distance, origin capacity, and destination attractiveness.²⁹ Important caveat: our entity resolution is state-constrained for 93% of matches, mechanically inflating same-state grant counts. This means the 47–62× multiplier includes a matching artifact and should be treated as an upper bound; the true border premium is likely substantially lower. Distance decay and all models converge. For comparison, a naïve OLS specification (log-linearized, dropping zeros) yields inflated coefficients of β = 4.51–4.75 (91–115× multiplier), confirming the upward bias from omitting zero-flow pairs documented by Silva & Tenreyro (2006). State-centroid distance is a coarse proxy that compresses within-state variation (Albany–NYC is 150 miles; NYC–Newark is 10). ZIP-code-level distances from Census ZCTA centroids and a gravity specification restricted to non-state-constrained matches (to separate the matching artifact from the genuine border premium) are planned for the journal submission. A lower bound cannot be estimated without sub-state geographic resolution (e.g., ZIP-code distances), which we plan for the journal submission. The unconstrained-only subsample (3.2% of resolved grants) yields β_same = 1.8–2.3, but this subsample is small and selected, so we do not report it as a formal bound.

Finding 3: COVID and the Donor Extinction Pattern

The pandemic did not merely reduce funding levels. It severed funding relationships. Between 2019 and 2020, 46.4% of active arts donors stopped giving entirely—the largest single-year churn event in our dataset.

The recovery, however, was rapid and structurally revealing. By 2020→2021, retention rebounded to 74.9%, partly because the e-filing mandate brought back donors who had been filing on paper but were never truly absent from the network. By 2022–2024, retention stabilized at 77–79%.

The inertia of patronage

We replicate the original paper's central behavioral finding: the longer a foundation has been giving to the arts, the more likely it is to continue. This "inertia of patronage" follows a clear monotonic relationship with consecutive giving streak.

A foundation giving to the arts for the first time has a 58% chance of doing so again the following year. But a foundation with five consecutive years of arts patronage has an 89% chance of continuing. This is not merely habit; it is consistent with deepening institutional commitment—board relationships, grant officer expertise, established reporting structures—that make exit costly. The inertia of patronage is a structural feature of the network, not just a behavioral one.

Finding 4: The Network Became More Connected, Not Just Larger

The legitimation economy operates as a single connected market. A family foundation in rural Ohio giving $5,000 to the Cleveland Orchestra is structurally connected to the Andrew W. Mellon Foundation through shared recipients—and has been for years. When the e-filing mandate doubled the observed network, the largest connected component did not shrink (as we would expect if the new entrants occupied isolated local ecosystems). It grew—consolidating into the single-component structure documented in Table 1.

The 14,000+ foundations that became visible after 2021 were already funding the same arts organizations as the major donors. The preferential attachment kernel (α ≈ 1.07) operates globally across this network, not locally within communities. A recipient endorsed by a high-degree donor attracts further endorsement from foundations of all sizes and geographies, because the endorsement signal propagates through shared recipients.

This means there is no structurally insulated “alternative scene”—no pocket of the network where a different allocation logic operates. The same topology that governs Mellon Foundation grants governs $2,000 family foundation gifts. Under AI-driven supply expansion, where the volume of creative production grows faster than attention or funding, this structural unity means the legitimation bottleneck cannot be routed around. It must be passed through.

Finding 5: Concentration Fell in Share but Rose in Absolute Dollars

A preferential attachment kernel of α ≈ 1.07 operating on a growing network produces a characteristic signature: declining share but rising absolute capture. This is what preferential attachment looks like in a legitimation economy.

The top 100 arts recipients captured 46.0% of all funding in 2019. By 2023, that share had fallen to 38.3% (37.9% in incomplete 2024 data)—roughly an eight-percentage-point decline. The HHI dropped from 0.0040 (2019) to 0.0035 (2023), with incomplete 2024 data showing 0.0033. At first glance, this suggests deconcentration. But as the counterfactual test below reveals, the story is more complex.

The top recipients by PageRank centrality remain familiar names—the Metropolitan Museum of Art, the Metropolitan Opera, the Museum of Modern Art, the American Museum of Natural History (classified under NTEE A50 “Museum Activities” despite its natural science focus—a classification artifact worth noting). Their share of the network is declining, but as we show below, this is almost entirely because the denominator grew, not because their funding shrank.

Counterfactual decomposition. The 2019 top-100 recipients received $609M. By 2022, those same 100 organizations received $882M—a 45% increase in nominal dollars (~25% real).³⁴ Their share fell to 29.2% only because the denominator grew to $3.03B. The top-100’s actual 2022 funding exceeds the denominator-only counterfactual by $273M.

Rising absolute capture with declining share is the signature of a maturing legitimation economy. The prestige hierarchy is self-reinforcing: endorsed institutions attract more endorsement, more dollars, more survival advantage—even as the network expands around them. This is consistent with a growing divergence between aggregate funding and organizational health. Giving USA (2025) reports arts giving at all-time highs (6.4% real growth in 2024), but SMU DataArts (2025), tracking 6,513 arts organizations over the same period, documents the opposite trajectory at the organizational level: 44% ran operating deficits in 2024—the highest rate in six years, up from 36% pre-pandemic and 26% at the trough of pandemic-era relief.³⁸ Working capital has fallen for three consecutive years, from 6.75 months (2021) to 4.25 months (2024), and 42% of organizations now hold three months of reserves or less. More aggregate dollars flowing through a more concentrated network into organizations with thinner reserves—this is the material signature of a legitimation economy under stress. The HHI decline (0.0040 to 0.0033) reflects a broader base of visible recipients, not a redistribution of resources away from incumbents. The legitimation hierarchy intensifies in absolute terms precisely when its relative footprint appears to shrink.

The prestige hierarchy

PageRank centrality, which measures a node's importance based on both the number and importance of its connections, reveals the institutional hierarchy of American arts philanthropy. The top ten recipients in 2024:

Finding 6: The Network Rewires Predictably

Beneath the aggregate trends in concentration and retention, the network's wiring follows precise quantitative laws. Two measures reveal this: edge persistence (do specific donor-recipient pairs endure?) and preferential attachment (do well-connected recipients attract new donors faster?).

Edge persistence

In any given year, 43–68% of donor-recipient edges persist into the following year. The COVID transition (2019→2020) shows the lowest persistence at 43.1%; by the post-mandate steady state (2022→2023), persistence climbed to 68.4%. In the pre-mandate period (2019→2020), persistence is stratified by grant size as expected: Q4 (largest) at 51.7% vs. Q1 (smallest) at 35.3%. However, in the post-mandate years (2021–2023), Q1 persistence anomalously exceeds Q4—likely because the mandate brought in a cohort of small grants from foundations that were already giving but not e-filing, mechanically inflating Q1 persistence. The 2023→2024 Q1 collapse to 19.3% may reflect filing-lag incompleteness in 2024 data. Grant-size persistence patterns should therefore be interpreted cautiously until the unfiltered pipeline provides a more stable basis.

Preferential attachment

New funding relationships do not form randomly. Recipients with more existing donors attract new donors at a rate that scales nearly linearly with their current degree—the classic “rich get richer” dynamic. We estimate the preferential attachment exponent α by weighted least-squares regression of log(mean new edges) against log(degree) across nine degree bins (1, 2, 3, 4–5, 6–10, 11–20, 21–50, 51–100, 101+). The results are consistent across all four transitions:

Using geometric log-spaced bins (k ≥ 1), we estimate α = 0.99–1.07 across year-pair transitions (mean α ≈ 1.07), with R² > 0.99. Per-degree-value averaging yields a slightly lower range (α ≈ 0.95–1.01, mean 0.98). The attachment kernel is near-linear. Recipients with 50+ existing donors attract 20–80× the new edges per node compared to degree-1 recipients. Methodological note: High R² in binned regressions with few points is expected for any monotonically increasing relationship and should not be taken as strong evidence of functional form. Standard errors on α from OLS are ±0.02–0.03 (geometric bins), but these understate true uncertainty because they do not account for bin boundary sensitivity. The range across methods (0.98–1.07) is the honest uncertainty band; the qualitative finding is that attachment is approximately linear (α ≈ 1), not sub-linear (α < 0.5) or strongly super-linear (α > 2). Unbinned maximum likelihood estimation is planned for the journal submission. Ecological caveat: The pooled α ≈ 1.07 is estimated across all donor tiers. If mega foundations (>$10M annual giving, <0.2% of donors, ~30% of dollars) attach via institutional board-level relationships (potentially super-linear) while micro foundations (<$10K, 55% of donors, 3% of dollars) attach via geographic proximity (potentially sub-linear), the pooled estimate averages over incommensurable processes. Tier-specific attachment exponents, which we have not yet estimated, could differ materially from the aggregate, with implications for which tier drives the concentration predictions under AI abundance.

Degree distribution vs. attachment kernel. Clauset–Shalizi–Newman goodness-of-fit testing¹² reveals that the recipient degree distribution is better described by a log-normal or truncated power law than a pure power law. For 2021–2023, log-normal is significantly preferred (p < 0.01); for 2019, the preference is directional but not significant (p = 0.07). We do not claim the degree distribution is scale-free. But the degree distribution and the attachment kernel are distinct objects: a network can exhibit near-linear preferential attachment while its degree distribution deviates from a pure power law due to finite-size effects, degree-dependent node removal, or heterogeneous fitness.¹² Our claim is about the attachment process (how new edges form), not the static distribution (the shape of the degree histogram). The attachment kernel is robustly near-linear (R² > 0.99 across nine bins and four transition periods), and the qualitative finding—that well-connected recipients attract new donors at rates far exceeding what random attachment would produce—holds regardless of the precise distributional form.

This attachment pattern has structural implications. A near-linear attachment kernel—first formalized by Barabási & Albert (1999)²⁷—is the canonical mechanism associated with heavy-tailed degree distributions in networks. This pattern is consistent with the network’s concentration being at least partly endogenous—arising from the network’s own wiring rules, not just from differences in institutional quality or prestige. A recipient that has attracted many funders tends to attract more, though our observational design cannot distinguish this attachment pattern from unobserved quality differences. This has direct implications for how we interpret concentration trends: the apparent share decline documented in Finding 5 is working against a structural force that continuously pushes toward concentration—and as the counterfactual test shows, the absolute funding trend runs counter to the share decline.

Community structure

Louvain community detection reveals 1,300–1,850 funding communities per year, with modularity consistently between 0.86 and 0.90—indicating strong community structure. Year-over-year community stability, measured by Normalized Mutual Information (NMI), ranges from 0.50 to 0.58: communities are moderately persistent but not frozen. About half the community structure reshuffles each year, while the other half reflects durable funding coalitions—clusters of foundations that consistently support overlapping sets of recipients. This pattern parallels Uzzi & Spiro’s (2005) finding in Broadway musical production networks, where optimal creative output arose from a specific balance of cohesion and novelty—too much clustering reduced creativity, too little reduced coordination.²⁶ Our modularity of 0.86–0.90 sits at the high end of that spectrum: arts funding communities are tightly cohesive, which may facilitate coordination but limit the cross-pollination that drives creative renewal. De Vaan, Vedres & Stark (2015) formalized this intuition: “structural folds”—positions where otherwise separate groups overlap—predict creative breakthroughs, and our high modularity implies very few such folds exist in the arts funding network.⁴³

Null benchmark. An important caveat: bipartite networks (foundations → recipients) produce mechanically high modularity under Louvain because same-type nodes cannot share edges, inflating apparent community structure. To assess how much of the observed Q = 0.86–0.90 reflects genuine funding coalitions versus bipartite artifact, we construct a configuration-model null: 100 random bipartite graphs preserving the exact degree sequence of each year’s network. The null modularity is Q_null = 0.53–0.63, meaning roughly 60–70% of the raw score is attributable to bipartite structure alone. The excess modularity—Q_excess = Q_obs − Q_null ≈ 0.27–0.32—is the portion attributable to real funding coalitions and is statistically significant (p < 0.01 against the null ensemble). As a complementary check, we project the bipartite network onto a one-mode foundation co-funding network and compute modularity there: Q_projection = 0.59–0.63, confirming that meaningful community structure persists after removing bipartite inflation. The qualitative conclusion holds—arts funding is strongly clustered—but the raw modularity overstates the degree of clustering by a factor of roughly three.

---

Finding 7: Art Forms Survived COVID Differently

The pandemic did not affect all art forms equally. By mapping NTEE sub-codes to art form categories, we can track differential COVID impact and recovery trajectories across the arts ecosystem.

Museums bore the brunt. They were the only major art form to experience a funding decline exceeding 20%, reflecting both the immediate impact of physical closures and the operational costs of maintaining collections without visitors. Their recovery was also the slowest: museums did not return to 2019 funding levels until 2021, while most other art forms never fell below baseline.

Performing arts funding was essentially flat during COVID (+0.4%), which may seem counterintuitive given that theaters and concert halls were shuttered. However, a change this small is within the resolution uncertainty of our pipeline. To the extent it reflects a real pattern, one interpretation is that donors maintained or increased support for performing arts organizations precisely because they were in existential crisis—though we cannot rule out that this stability simply reflects the inertia of multi-year grant commitments already in place before the pandemic.

Visual arts funding grew 446% from 2019 to 2024, though from a small base ($3.3M). A pre-mandate filer decomposition reveals that approximately half of this growth is compositional: foundations active before the mandate grew from $3.3M to $10.8M (+227%), while newly visible post-mandate filers contributed $7.3M (40% of 2024 Visual Arts dollars). The pattern is similar across art forms—roughly 35–60% of each category’s 5-year growth comes from newly visible filers rather than behavioral change by existing donors.

Mandate caveat. All 2024-versus-2019 growth rates in Table 6 are inflated by the e-filing mandate, which roughly doubled the number of observable filings between 2019 (65K) and 2024 (127K). The COVID impact column (2019→2020) is unaffected, as the mandate post-dates both years. The art-form-stratified pre-mandate decomposition (analogous to the analysis in Finding 4) is available in the supplementary data (artform-mandate-decomposition.csv).

Finding 7b: Locality Across 15 Years

The locality data above (Figure 2) covers our pipeline years only. By merging the Shekhtman & Barabási dataset (2010–2019) with our 2019–2024 extension, we can place the COVID and mandate effects in the context of a full decade of prior stability.

The 15-year view reveals that the ~62% locality rate is not a recent phenomenon but a long-run structural feature of American arts philanthropy. For a full decade before the pandemic, approximately 61–66% of arts grants went to organizations in the same state as the donor. This equilibrium was disturbed only twice: by COVID (upward to 65%) and by the e-filing mandate (downward to 52%), the latter being a composition effect rather than a change in behavior.

The 2024 data shows locality at 60%, but this figure should be interpreted cautiously: 2024 filings are incomplete (72,823 grants vs. 84,532 in 2023), and established, more-local foundations may file earlier than newer cross-state donors, biasing the incomplete sample toward higher locality. If the pattern holds as late filers arrive, it would suggest the post-mandate composition is settling toward a new equilibrium approximately two percentage points below the pre-mandate baseline—reflecting the permanent addition of small cross-state foundations to the visible network.

Finding 8: Geography Shapes Who Gets Funded

Arts philanthropy in America is geographically concentrated to a degree that mirrors—and reinforces—the concentration of cultural institutions themselves. By mapping grant flows to recipient states across all six years, we find that two states absorb a disproportionate share of total funding, while a persistent set of states constitute funding deserts that show no signs of closing the gap.

New York alone absorbs roughly 29% of all arts philanthropy in America. Combined with California, the two states account for nearly half of national arts grant funding. This is not merely a function of population: New York’s $36 per capita is more than three times California’s $11, reflecting the density of elite cultural institutions—the Met, MoMA, Lincoln Center, the New York Philharmonic—that attract funding from foundations nationwide. DC’s per-capita rate of $112–172 is structurally unique: it hosts national institutions funded by the entire country but has the population of a mid-sized city.

Funding deserts

Four states sit in the bottom quartile of per-capita arts funding in every single year from 2019 to 2024: Alabama, Kansas, Idaho, and South Carolina. These are persistent funding deserts. An additional six states—North Dakota, West Virginia, Arizona, Nevada, New Hampshire, and Iowa—appear in the bottom quartile in five of six years. North Carolina narrowly misses this group. The pattern is stable across pre-mandate, COVID, and post-mandate eras, suggesting structural rather than cyclical causes.

These states share a common profile: fewer headquartered foundations, fewer nationally prominent cultural institutions, and lower population density. The result is a double disadvantage. Local foundations are scarce, and out-of-state foundations—which default to geographic proximity in their giving—rarely reach into these states. The locality effect documented in our earlier findings actively works against funding desert states.

Competing explanation. Low per-capita funding may partly reflect a supply-side constraint: these states may simply have fewer arts organizations eligible to receive grants, rather than (or in addition to) fewer willing donors. Our data measures grant flows, not unmet demand. A state with few arts organizations will appear as a “desert” even if its existing organizations are well-funded relative to their needs.

Exporters and importers

Not all states consume what their foundations produce. Illinois is the largest net exporter of arts philanthropy, giving $30–95M more per year than its organizations receive. This reflects Chicago’s outsized foundation sector (MacArthur, Donnelley, Polk Bros) relative to its share of recipient institutions. Delaware and Nevada also export disproportionately, giving roughly twice what they receive—Delaware due to foundation-friendly incorporation laws, Nevada likely due to a small number of large private foundations.

New York, by contrast, is the nation’s largest net importer. It gives generously ($700M+) but receives even more, as foundations from across the country fund its marquee institutions. This import surplus—the gap between what New York foundations give and what New York organizations receive—quantifies the gravitational pull of prestige. Institutions with national reputations attract funding regardless of where the donor is located.

COVID’s geographic effects

The pandemic produced dramatic geographic variation. Most states saw funding increases from 2019 to 2020 as foundations rallied, consistent with the overall COVID resilience effect documented earlier. But the distribution was highly uneven. Washington state surged +664%, almost certainly driven by a single mega-gift rather than broad-based increases. Meanwhile, North Carolina declined 47% and Alaska declined 42%—states where the foundation ecosystem was too thin to provide the stabilizing effect seen elsewhere.

The e-filing mandate produced broad increases across virtually all states, confirming that the data coverage expansion was geographically uniform rather than concentrated in particular regions. This gives us confidence that post-mandate geographic comparisons reflect genuine funding patterns rather than differential reporting artifacts.

Finding 9: Arts Philanthropy Is Structurally Distinct

As previewed in the thesis section, arts giving is measurably more local and more concentrated than non-arts philanthropy. Comparing 530K resolved grants from the 2019 990-PF filings—60K arts-classified grants vs. 471K non-arts, resolved through the same full fuzzy entity-resolution pipeline—reveals systematic differences (Table 11). The headline result: arts philanthropy is approximately 9 percentage points more local (60.3% vs. 51.2% same-state, grant-count-weighted) and 10 percentage points more concentrated (top-100 share of 43.9% vs. 33.8%). Note: the Table 11 comparison uses fuzzy-resolved arts grants for apples-to-apples comparison with non-arts; the publication figure (Figure 12) uses the unfiltered arts pipeline (61.5% vs. 41.1%, +20.4pp) which includes grants that did not resolve through entity matching, giving a larger but less controlled gap. The conservative estimate from the matched pipeline (+9pp) is the primary result.

Critically, the locality premium is not a grant-size artifact. Stratifying by grant amount reveals that arts giving is more local than non-arts giving within every size bin, and the gap increases with grant size: +1.3pp for grants under $1K, +5.5pp for $1K–$10K, +9.1pp for $10K–$100K, and +18.2pp for grants above $100K (2019 data). The same pattern holds in 2022. Concentration shows the same within-bin structure: the arts top-100 share is 3–4× higher than non-arts at every grant-size bracket. Whatever makes arts philanthropy structurally distinct, it is not reducible to differences in typical grant size between sectors.

Finding 10: Portfolio Breadth Is Associated with Donor Survival

How many art forms a donor funds is strongly associated with whether they will keep giving. This finding is novel—it does not appear in the original Shekhtman & Barabási paper—and has practical implications for arts organizations seeking stable funding.

In any given year, 67–71% of arts donors fund only one art form (as classified by NTEE sub-codes⁴). These single-form donors retain at 68.5% year over year (2022→2023). Donors funding two forms retain at 86.9%. Donors funding three or more forms retain at 93–97%. The association between portfolio breadth and retention is at least as strong as that of giving streak length, though the two co-vary.

Average portfolio breadth has been slowly increasing: 1.43 art forms per donor in 2019, rising to 1.55 by 2024. COVID appears to have nudged donors toward diversification, with breadth jumping from 1.43 to 1.50 between 2019 and 2020. The strongest co-funding axis is General Arts ↔ Performing Arts, with 33–39% bidirectional overlap. Museum donors are the most siloed: only 37.4% of museum funders also fund other art forms.

The association is intuitive: donors who engage with multiple art forms are more deeply embedded in the arts philanthropy network. Their giving may reflect a broad commitment to the cultural ecosystem rather than a transactional relationship with a single institution.

Multivariate controls. Portfolio breadth correlates with foundation size (larger foundations fund more categories) and tenure (longer-tenured foundations have had more time to diversify). A logistic regression controlling for log(total giving) and years active confirms that breadth retains independent predictive power: OR = 1.84 [1.70–1.99] per standard deviation (2023→2024, n = 13,694). This estimate comes from a specification that drops the num_recipients predictor, which exhibits quasi-separation in some year-pairs; the full model (including num_recipients) gives a nearly identical breadth OR of 1.86 [1.72–2.01], confirming the result is not sensitive to this modeling choice. The breadth OR ranges from 1.27 to 1.86 across year-pairs (pooled mean 1.62). Stratified analysis corroborates the breadth effect: within the bottom half of giving, 1-form donors retain at 64.2% vs. 92.5% for 3+ forms (+28pp). Within the top quartile (>$46,000/year), the gap narrows but persists: 77.8% vs. 95.3% (+17pp). Breadth is not merely a proxy for “large, established foundation”—it carries an independent signal about a donor’s embeddedness in the arts ecosystem.

Methodological Controls

Systematic robustness checks confirm that the core findings reported above are not artifacts of methodological choices. Four classes of sensitivity analysis were conducted:⁶

Match confidence sensitivity. Locality rates are stable across match-confidence thresholds from 0.80 to 1.0. The findings are not an artifact of fuzzy matching between grant records and the Business Master File.

Pre-mandate subsample. The locality drop observed in 2021–2023 is primarily compositional, not behavioral. Pre-mandate donors maintained locality rates of 0.60–0.64 throughout the period. The aggregate drop to ~0.52 was driven by new e-filers who initially gave less locally (0.40–0.43) but converged to 0.60 by 2024.

Cross-sectional comparison: mandate entrants vs. voluntary e-filers. To formally quantify the composition effect, we compare 12,819 mandate entrants (foundations first appearing in 2021 or later) to 12,189 voluntary e-filers (present pre-2021 and still active 2022–2023). In the post-period (2022–2023 pooled), mandate entrants give at 42.9% same-state versus 60.5% for voluntary e-filers—a 17.6pp locality gap (two-proportion z = 71.2, p < 10⁻⁶). Mandate entrants also have smaller median grants ($2,000 vs. $5,000) and lower donor retention (73.3% vs. 84.0%). However, mandate entrants show higher edge persistence (71.9% vs. 64.7%), suggesting that while they are less local and less retained, the relationships they do maintain are more stable. This is not a causal treatment effect on behavior—these foundations were always giving; they simply were not visible. The comparison quantifies the structural difference between the newly visible and previously visible populations, confirming that the post-mandate locality drop is a composition artifact rather than a behavioral change. (Note: a standard difference-in-differences design is not possible here because the treatment group has no pre-period data by definition.)

Grant size. Larger grants are more local. Fourth-quartile grants show 64–67% same-state rates, compared to 53% for first-quartile grants. Small grants cross state lines more frequently.

Bootstrap confidence intervals. All estimates are tight, with 95% confidence intervals approximately 0.007–0.011 wide. Year-over-year differences are statistically unambiguous.

Synthesis: The Long Tail as Hidden Structure

Segmenting foundations by annual arts giving reveals a sharply skewed distribution that reframes several earlier findings.

The 84/10 split. Micro and Small foundations make up 84% of all arts donors but contribute roughly 10% of total dollars. This is the long tail of arts philanthropy. The remaining 16%—Medium, Large, and Mega—drive 90% of funding. Any analysis that counts donors equally will be dominated by the long tail; any analysis that counts dollars will be dominated by the head.

COVID churn was a small-foundation phenomenon. Micro foundations drove 56.5% of all donor churn (3,413 of 6,039 churned donors). Combined with Small foundations, they account for 92.3% of churn. Micro retention during COVID was just 48.3%, compared to 75% for Mega foundations. The dollar impact was modest precisely because the churning foundations held so little of the total. The COVID churn event (Finding 3) was real, but it was concentrated in the long tail.

The mandate expansion was also small foundations. Micro and Small foundations drove 86% of new entrants in 2021 (5,035 new Micro, 3,961 new Small). This is why the network doubled in size (Finding 1) but observed behavior shifted: the new foundations are structurally different from the pre-mandate visible set. They are smaller, initially less local (40–43% same-state vs. the established ~62%), and less committed.

Non-monotonic locality gradient. Locality does not decline smoothly with foundation size. Micro through Medium foundations give 60–67% in-state. Large foundations drop sharply to ~29%—the most national tier. But Mega foundations rise back to ~38%, likely because their portfolios are anchored by flagship home-state institutions (a Mega foundation in New York giving $20M to the Met alongside $30M distributed nationally will register as substantially local). Large foundations occupy a structural middle ground: large enough to give nationally but not so large as to maintain a dominant home-state commitment. As noted in Finding 2, the newly visible mandate entrants—though mostly Micro and Small by size—initially gave at only 40–43% same-state before converging to the ~60% norm by 2024, suggesting a behavioral transition period distinct from the steady-state locality of established small foundations.

Retention tracks size. Steady-state retention ranges from ~71% for Micro foundations to ~93% for Mega foundations. Size is among the strongest predictors of philanthropic commitment. This gradient also means that the donor base is self-selecting over time: the foundations that persist are disproportionately the larger, more committed ones.

Tier mobility is low. Roughly 81% of donors remain in the same tier year-over-year. The hierarchy is sticky. Foundations rarely leap from Micro to Medium, and Large foundations rarely shrink to Small. The structure of arts philanthropy is set at entry.

The legitimation economy operates in two regimes. A peripheral churn layer—84% of donors, 10% of dollars, high turnover—absorbs expansion and contraction. A core inertia layer—16% of donors, 90% of dollars, sticky relationships—controls the endorsement infrastructure that determines which organizations survive. The locality drop, the apparent concentration decline (Finding 5), and the COVID churn patterns are all compositional effects of revealing the peripheral layer, not changes in the core’s behavior. The underlying network of committed arts donors was more stable than the aggregate statistics suggested.

The recipient side of this bifurcation is counterintuitive. SMU DataArts (2025) documents that large arts organizations (>$1M budget)—the institutions that absorb the core layer’s concentrated funding—hold only 3.09 months of working capital, the lowest of any size tier.³⁸ Small organizations (<$250K) hold 7.31 months—2.4 times the cushion. Large organizations’ revenue declined 22% in real terms over five years; small organizations grew 28% real. Corporate giving to large organizations collapsed 51% in real terms; foundation giving stagnated at −20% real. The organizations at the center of the legitimation economy are the most financially fragile precisely because they are at the center: their scale creates cost structures that earned revenue can no longer sustain (paid attendance remains 22% below 2019), making them structurally dependent on the donor network this paper measures. The core inertia layer is inertial in the network but eroding in material terms.

This two-regime structure is the finding. Under AI-driven supply expansion—where the volume of creative production grows faster than institutional attention—the peripheral layer absorbs the growth while the core’s endorsement becomes more consequential, not less. The 84/10 split should intensify, not equalize. The legitimation bottleneck tightens precisely when production becomes abundant.

Independent evidence from the commercial art market corroborates this bifurcation. In the Art Basel & UBS 2026 dealer survey, the middle market ($500K–$1M turnover) is the only segment where optimism fell year-over-year, and 45% of mid-tier dealers reported declining margins despite rising sales.⁶³ The top (>$10M) and bottom (<$250K) are more confident. The middle is hollowing out—squeezed between mega-galleries that internalize legitimation and micro-galleries that accept peripheral economics. At auction, the polarization is starker still: less than 1% of transactions capture 54% of total value, and the top 0.02% of lots alone account for 22%. The $10M+ segment was the only price tier to grow from 2010 to 2025; every other tier declined while the middle market “lost considerable weight… as the market has become increasingly polarized.”

The Legitimation Economy: Full Evidence

The legitimation economy framework introduced at the outset rests on two empirical predictions: that arts philanthropy is structurally distinct from philanthropy in general, and that institutional endorsement—not just dollar volume—is associated with organizational survival. This section presents the full evidence for those predictions and develops the framework’s implications. We note throughout that associations do not establish causation: the gateway-donor survival effect (Table 12) could reflect selection on unobservable recipient quality rather than legitimation per se. A causal test would require an identification strategy (e.g., regression discontinuity around some centrality threshold or instrumental-variable design) that this dataset does not yet support.

Arts philanthropy is structurally distinct

To test whether arts funding operates under different structural constraints than philanthropy in general, we computed the same network parameters across all 906K grants in the 2019 IRS 990-PF dataset—not just the 39K arts-classified grants. The comparison reveals systematic differences:

Arts philanthropy is more local and more concentrated than non-arts philanthropy. These structural differences are consistent with Shekhtman, Gates & Barabási (2024),²⁰ who found that locality and retention hold across science philanthropy as well, suggesting domain-general properties of philanthropic networks—but the arts-specific magnitude of concentration (+10pp) indicates that legitimation dynamics are significantly stronger in cultural fields. Importantly, both arts and non-arts grants in this comparison use the identical state-constrained entity resolution pipeline. However, differential contamination testing reveals that arts grants resolve through state-constrained methods at a higher rate (91% vs. 80% for non-arts), which could inflate the locality gap. To quantify this, we computed locality rates stratified by match method: among exact-match-only grants (no fuzzy contamination), the arts locality premium is +9.9pp (arts 60.0% vs. non-arts 50.0%, n = 420K), actually larger than the +9.1pp overall gap. A conservative counterfactual—replacing state-constrained fuzzy matches with the exact-only locality baseline—yields an adjusted gap of +8.0pp, suggesting contamination of at most 1.1pp. The locality gap among grants resolved through fully unconstrained methods (exact_only, exact_multi, fuzzy_national) drops to +2.1pp, but this subsample is small (5,338 arts grants) and selected on name distinctiveness, making it a lower bound rather than a representative estimate. We conclude that the cross-sector locality gap is robust to entity resolution methodology, with a plausible range of +8pp to +10pp after accounting for differential contamination.

Institutional endorsement is associated with survival

If legitimacy is the scarce resource in arts philanthropy, then an organization’s survival should be associated not just with how much money it receives, but with who endorses it. We tested this by tracking new recipients that entered the arts funding network in 2022 (the mandate year) and measuring their survival to 2023, classified by the centrality of their “gateway donor”—the highest-degree donor that funded them.

Recipients entering via high-centrality donors (Q4 by out-degree) survive at 67.5%—a 13.8pp advantage over those entering via lower-degree donors (53.7%). A logistic regression controlling for log grant amount and number of donors narrows the gap to 6.2pp (Q4 odds ratio = 1.54 [1.44, 1.64]). Grant size has a negative association with survival (OR = 0.87 [0.86, 0.88]), likely because large one-off project grants are less predictive of continuity than smaller recurring commitments. Number of donors is the strongest predictor (OR = 3.08 [2.89, 3.29]). The centrality effect is most pronounced for small grants: among recipients whose largest gateway grant was under $1,000, Q4-gateway recipients survive at 77.2% versus 45.8% for Q1–Q3—a 31pp gap where the endorsement signal, rather than the dollar amount, is the stronger correlate of survival. Centrality note: PageRank is degenerate for donors in this network (99.9% share identical minimum values); we use out-degree (number of recipients funded) as the centrality measure. Selection caveat: high-centrality donors may systematically select higher-quality or more established recipients, creating omitted-variable bias. The 6.2pp adjusted gap controls for grant amount and donor count but does not control for unobserved recipient characteristics (organizational capacity, board quality, earned revenue). Propensity-score matching on additional observables is planned as a robustness check for the journal submission. Mechanism: The causal pathway may be institutional dependency inversion, where museums facing constrained public budgets “rely on galleries to co-finance exhibitions, catalogs, or logistical operations,” creating a paradox where “the institution must support and legitimize artists who are already strongly backed by the market.”⁶⁴ If a parallel dynamic operates in philanthropy—where recipient organizations depend on high-centrality donors not just for dollars but for the endorsement signal that attracts subsequent funding—the gateway donor effect reflects a structural feature of legitimation economies, not a selection artifact. Material stakes: SMU DataArts (2025) reports that 42% of arts organizations hold three months of reserves or less.³⁸ For these organizations, the difference between receiving a gateway grant from a Q4 donor (survival rate 77.2% among small grants) versus a Q1–Q3 donor (45.8%) is not merely a signal—it is the difference between bridging a funding gap and failing to make payroll. The 6.2pp adjusted survival advantage operates in a sector where the median organization runs a 1% surplus; even a small advantage in the probability of receiving a subsequent grant becomes existential when the financial margin is this thin. Geographic confound: Q4 donors are disproportionately concentrated in New York, and New York recipients are disproportionately large, established institutions that may survive regardless of gateway donor centrality. Adding recipient-state fixed effects to the regression confirms that the centrality effect is not a geographic artifact: OR = 1.54 [1.44, 1.65] with state FE, virtually unchanged from the base model (OR = 1.54 [1.44, 1.64]), and the adjusted gap remains 6.2pp.

Translation power is scarce and costly

Cross-state grants—edges that bridge geographic communities—carry a median of $1,000 compared to $5,000 for same-state grants (2022). The gap is not driven by small grants being inherently cross-state: at every grant-size bracket above $1K, same-state grants outnumber cross-state grants roughly 2:1. What varies is who bridges: high-degree donors (Q4) give 53% of their grants cross-state, versus 38% for low-degree donors (Q1). Translation power—the ability to legitimize recipients outside one’s geographic cluster—concentrates in a small number of high-centrality foundations, and the edges they extend across state lines carry less money per grant than their local commitments.

Returning to the three scarce resources

The evidence above substantiates the legitimation economy framework introduced at the outset. The three structural parameters we identified—legitimacy (α ≈ 1.07), continuity (43–68% edge persistence), and translation power (excess modularity Q_excess ≈ 0.27–0.32)—are not merely descriptive labels. They correspond to dynamics described independently in cultural economics. Moureau (2025), analyzing gallery promotion strategies for the Art Basel & UBS report, identifies the micro-mechanism: “Each signal reinforces those that preceded it, each choice legitimizes future ones, and a form of path dependency gradually takes hold”—citing Arthur’s (1989) canonical model of technological lock-in.⁶⁴⁶⁵ This is preferential attachment described without the network-science vocabulary: past endorsement predicts future endorsement, accumulated endorsement stock resists displacement, and the gallery system that produces these signals functions as a cross-community bridge between market and institution (Moulin 1986).⁶⁶ Our α ≈ 1.07 is the quantitative measurement of what Arthur theorized and Moureau observes qualitatively. The 73% rate at which CPGA-affiliated galleries pursue institutional placements (vs. 50% for non-affiliated) is the behavioral mechanism behind our measured edge persistence. And Moureau’s observation that “360-degree galleries” internalize cross-community bridging—controlling “both the market and the symbolic recognition of their artists”—maps directly onto our modularity finding: translation power concentrates in a small number of high-degree nodes. The commercial art market provides independent confirmation of the continuity parameter: 62% of dealer sales by value flow through repeat buyer relationships (1+ year history), rising to 76% for the largest dealers—a range that maps closely onto the 43–68% edge persistence we measure in philanthropic networks.⁶³ Relationship stickiness is increasing, not decreasing, despite digital alternatives: established buyers (>5 years) captured 43% of mid-tier auction value in 2025, up from 32% in 2021.⁶³ The three parameters are measurable quantities that distinguish arts philanthropy from other philanthropic sectors (Table 11), are associated with organizational survival (Table 12), and identify the specific bottleneck (cross-community bridging) that concentrates allocation power in a narrow set of high-centrality donors. The persistence asymmetry between arts and non-arts (+4–5pp) suggests that relationship capital is more valuable—and harder to substitute—in sectors where value is socially constructed. The question that follows is what happens to this architecture under AI-driven supply abundance.

A formal model of allocation in a legitimation economy standalone view →

We formalize the structural dynamics documented above as a simple allocation model in which donors choose recipients based on three sources of utility. The model generates preferential attachment, edge persistence, and community clustering as equilibrium properties, and yields falsifiable comparative statics under AI-driven supply abundance.

Setup. Consider a bipartite network with N donors and M(θ) recipients, where θ ∈ [0, ∞) parameterizes the cost of producing cultural output. Higher θ corresponds to cheaper production (AI abundance), so M′(θ) > 0: as production costs fall, the pool of viable recipients expands. In the limiting case modeled by Korinek & Suh (2024), where human labor in cognitive tasks becomes fully substitutable, θ → ∞ and the scarce resource is entirely endorsement.⁴⁰ Each recipient i has an unobserved quality q_i and an observed endorsement stock e_i,t equal to its cumulative weighted in-degree at time t. Each donor j belongs to a community c_j.

Donor utility. In each period, donor j allocating a grant to recipient i receives payoff:

U_j(i, t)  =  β₁ · w(e_i,t)  +  β₂ · r_ji,t  +  β₃ · s(i, c_j)   (1)

where w(e) is the status signal—the reputational return from associating with a recipient of endorsement stock e (Podolny 2005;⁴⁷ increasing, concave); r_ji,t ∈ {0, 1} is the relationship indicator (= 1 if j funded i last period; captures warm-glow and relational inertia); s(i, c_j) is the information signal—the quality of information j has about i, which is higher when i is known within j’s community (DiMaggio & Powell 1983¹⁰); The search cost λ(θ), increasing in M(θ), enters the donor’s switching decision (see edge persistence below) rather than the utility of a specific allocation. The three β terms correspond directly to the three measured structural parameters: β₁ generates preferential attachment, β₂ generates edge persistence, β₃ generates community clustering.

Attachment dynamics. For new edges, a logit specification gives Pr(j → i) ∝ exp(β₁ w(e_i) + β₃ s(i, c_j)). Note that the search cost λ(θ), being constant across all recipients, cancels from the logit choice probabilities—the attachment kernel is invariant to AI abundance. When w(e) = log(e), this reduces to the standard PA kernel Pr(i) ∝ e_i^α with α = β₁. The empirically estimated α from unconditional binned regression therefore captures both the status effect (β₁) and the average community signal; the two are separable only with within-community estimation, which we leave to future work. Our estimated α ≈ 1.07 implies β₁ ≈ 1: the marginal reputational return is approximately proportional to the stock itself.

Edge persistence. For an existing edge, the donor compares continuing the relationship to switching. Continuation utility is β₂ + β₁ w(e_i) + β₃ s(i, c_j); switching to the best alternative requires paying the search cost λ(θ). The continuation probability is therefore Pr(continue) = σ(β₂ + β₁ [w(e_i) − w(e_k)] + β₃ [s_i − s_k] + λ(θ)), where k is the best available alternative. The key feature: λ enters positively—higher search costs make continuation more attractive relative to switching. And β₂ is a donor-side parameter independent of recipient production capacity: a supply shock changes θ but not β₂. This predicts the COVID null result (p = 0.55): a shock that shuttered venues but did not change donor-side utility had no effect on edge persistence. Equations 1–3 together imply a nested decision structure: the donor first decides whether to continue an existing relationship (where λ enters the binary comparison) or form a new edge (at which point λ cancels from the multinomial choice over recipients). This sequential framing is a reduced-form approximation; a fully consistent multinomial model would fold continuation into the choice set, where λ cancels entirely and persistence is driven solely by β₂.

Endorsement stock dynamics. The stock evolves as e_i,t+1 = δ e_i,t + Σ_j g_ji,t · d_j, where δ is depreciation, g indicates a grant, and d_j is donor centrality weight. Grants from high-centrality donors contribute more to endorsement stock—formalizing the gateway donor effect: the 6.2pp survival advantage of high-centrality first donors (Table 12) identifies a discrete jump in e_i from the initial endorsement.

Extensive-margin assumption. The model governs which recipient receives a grant but is silent on the extensive margin—how many grants each donor makes. The concentration predictions below assume that the number of active donors N is approximately fixed (or grows slower than M), so that the ratio of endorsement supply to endorsement demand falls as production expands.

Comparative statics under AI abundance (θ increases). The model yields four testable predictions:

1. Concentration increases. With M′ > M recipients competing for fixed donor attention and α ≥ 1, new entrants dilute the lower tail of the funding distribution while top recipients continue attracting grants at rates proportional to e_i^α. The Gini of recipient funding G(M) increases with M as a compositional effect. Top-k recipients gain in absolute funding but may lose share as the denominator grows—precisely the pattern we observe in Finding 5, where top-100 recipients gained 45% in absolute dollars while their share fell from 46% to 43.4%. The prediction is rising inequality (Gini) and rising absolute concentration, not necessarily rising top-k share.
2. Persistence is robust to supply shocks. In the continuation-vs-switching formulation, λ(θ) enters positively: higher search costs make continuation more attractive. Therefore ∂persistence/∂θ ≥ 0 when the supply expansion primarily affects M (and thus λ) rather than the donor-side relationship parameter β₂. Supply-side disruption should not degrade donor relationships.
3. Peripheral churn increases. A new entrant with e_i,0 ≈ 0 attracts a first grant with probability ∼1/M. As M grows, this probability falls, making the gateway donor effect more consequential—the allocation-side analogue of Acemoglu & Restrepo’s (2019) task-displacement framework, where new technology expands the task space while concentrating returns among incumbents.³⁹ The 6.2pp advantage should strengthen under abundance.
4. Modularity is stable or increases (auxiliary hypothesis). Since λ cancels from the logit, this prediction does not follow directly from the model as specified. It requires the additional assumption that search costs are recipient-specific—lower for recipients known within the donor’s community, higher for outsiders—so that the relative advantage of within-community evaluation increases with M. Under this extension, donors retreat further into community-based information channels as the recipient pool expands—an effect amplified if AI-driven evaluation tools create algorithmic monoculture across foundations⁴⁴, and ∂Q/∂θ ≥ 0.

These predictions are qualitative (direction only). The model has sufficient free parameters that magnitude predictions would require structural estimation with micro-level donor data, which we leave to future work. We note that P1–P3 are also predicted by a simpler Barabási-Albert model with random edge deletion; the empirical leverage of the legitimation framework over vanilla preferential attachment rests on P4 (community structure, which simple PA models do not generate) and the gateway-donor mechanism (Eq 4, where donor centrality weights inject differential endorsement stock into new entrants). More broadly, the model has three free parameters (β₁, β₂, β₃) matched to three empirical moments (α, persistence, modularity), giving it zero degrees of freedom in the current-period calibration. It is presented as a theoretical framing device that organizes the empirical findings into falsifiable comparative statics, not as a structural estimate. Its predictions (P1–P4) are directional only; magnitude predictions would require additional empirical moments or structural estimation. Its value is not in fitting the data—any three-parameter model can match three moments—but in generating out-of-sample predictions (P1–P4 under θ-shocks) that are falsifiable against future data.

Falsification conditions. If AI abundance leads to deconcentration (declining Gini and declining absolute funding for top recipients even as M increases), the attachment kernel is sublinear in the expanded regime and the framework fails. If a pure supply shock disrupts edge persistence substantially exceeding the year-over-year variation observed in the COVID period (which produced a maximum shift of ~7pp), relationship inertia is weaker than the model requires. If modularity declines under abundance, AI provides information channels that substitute for community-based evaluation—a qualitatively different regime in which AI changes not only supply but the information architecture of the legitimation economy. The model is deliberately minimal; its purpose is to show that three standard mechanisms—status signaling, relational inertia, and community-based information—are jointly sufficient to generate the three parameters we measure, and to derive testable predictions under a single comparative static.

Cross-domain structural signatures

If the legitimation economy framework captures something general about allocation under endorsement scarcity, its structural signatures should appear across domains—not just in arts philanthropy. We compile measured parameters from five published studies of legitimation networks:

The pattern is consistent. Across philanthropy (arts and science), the commercial art market, academic labor markets, art exhibition networks, and competitive grant-making, legitimation networks share three properties: near-linear or super-linear attachment (past endorsement predicts future endorsement), high edge persistence or low upward mobility (relationships and hierarchies are sticky), and strong concentration (Gini ≥ 0.75 across all domains reporting it). The specific magnitudes differ—arts philanthropy is more local than science philanthropy (+25pp), art exhibitions show even more extreme lock-in (99.95% of peripheral artists never reach high-prestige institutions)—but the qualitative signature is robust. This cross-domain convergence supports interpreting the three parameters not as idiosyncratic features of one sector but as structural properties of allocation systems where endorsement, rather than price, governs resource flows.

Implications for AI abundance

The following discussion connects our measured structural parameters to emerging evidence on AI’s impact on creative markets. The formal model generates qualitative predictions but has not been structurally estimated; the comparative statics are directional hypotheses, not calibrated forecasts.

AI is closing a pincer on cultural production. On the supply side, it homogenizes output. Doshi & Hauser (2024) ran a controlled experiment in which writers given generative AI ideas produced stories rated more creative, better written, and more enjoyable—especially among less-creative writers, whose novelty scores rose 10.7% and whose stories were judged up to 26.6% better written—but whose collective output was measurably more similar to each other than stories written without AI.⁵⁰ The mechanism scales beyond the lab: Daryani, Sourati & Dehghani (2026) show that LLMs disproportionately reflect Western, high-income cultural perspectives, and that AI-generated content entering training corpora creates a feedback loop—each generation of models trained on the outputs of the last, progressively standardizing expression and marginalizing minority cultural forms—a dynamic accelerated by model collapse, in which AI systems trained on AI-generated data progressively lose distributional tails.¹⁷⁵¹ The regulatory environment is beginning to encode this distinction: under the EU AI Act (entering force August 2026), purely AI-generated works lack copyright protection, as “copyright generally does not extend to works that are not tied to some form of ‘human intellectual creation’”—a legal boundary between human-legitimated and machine-generated art. Individual creators get better; the field gets flatter. Noy & Zhang (2023) find the same pattern in professional writing: ChatGPT reduced task completion time by 40% while compressing the productivity distribution, with the largest gains for lower-ability workers.²¹ Brynjolfsson, Li & Raymond (2025) confirm at scale: across 5,172 customer service workers, AI compressed the skill distribution by 15%, with the least-experienced workers gaining the most.²²

On the market side, AI floods creative marketplaces. Goldberg & Lam (2025) studied a stock-image platform with nearly 500 million images and found that after AI-generated content was permitted, affected markets saw 78% more images per month and 88% more active sellers—but also a 23% decline in non-AI artists, as human creators were crowded out by cheaper, consumer-preferred AI-generated substitutes—a pattern corroborated by Teutloff et al. (2025), who document a 17% decline in graphic design job posts after generative AI adoption.³⁷⁵² This is not a hypothetical future. CISAC & PMP Strategy (2024) project that music creators will lose 24% of their revenue to generative AI by 2028, with cumulative losses of €10 billion in music alone, as the market for AI-generated music and audiovisual content grows from €3 billion to €64 billion over five years—projections consistent with UNESCO’s broader estimates of creator revenue losses across music and audiovisual sectors.¹⁶⁵³ Anthropic’s own Economic Index found that “Arts, Design, Entertainment, Sports, and Media” constitutes 10.3% of Claude usage—the second-largest occupational category—driven primarily by writing, editing, and creative production tasks.¹³ UNESCO’s 2025 CULTAI report identifies three imperatives for ethical AI in culture—rights and integrity, pluralism, and sustainable creative futures—acknowledging at the policy level that the homogenization and displacement dynamics documented above are already reshaping cultural production worldwide.⁵⁴

The fine art market corroborates this at scale (all commercial market data in this and the following paragraphs from the Art Basel & UBS Art Market Report 2026).⁶³ Online art sales spiked from 13% to 39% of dealer revenue during the pandemic—the largest natural experiment in digital art distribution ever conducted—then reverted to 16% by 2025, below general retail e-commerce penetration (21%). Physical co-presence channels (gallery visits and art fairs) account for 81% of dealer sales by value. Online acquisition of genuinely new buyers constitutes only 6% of total sales, versus 32% for in-person encounters. Online-only art fairs—the purest test of whether digital can substitute for physical legitimation—collapsed from 9% to 0.1% of sales in five years. Lowering transaction costs does not route around the legitimation bottleneck; it confirms it. At auction, the supply-abundance prediction is already observable: contemporary art—the only fine art sector with replenishable supply (living artists producing new work)—is also the only sector that contracted in 2025, while Impressionism (+47%), Old Masters (+30%), and Modern art (+9%) all grew. The “ultra-contemporary” subsegment collapsed from $2.9B at its 2021 peak to $0.9B in 2025, a 69% decline. Meanwhile, Klimt’s Portrait of Fraulein Lieser sold for $236M—the second-highest auction price in history. The sector where production costs are lowest is the sector where allocation is tightest.

The third side of the pincer is allocation—and this is where our data speaks directly. When production is abundant and increasingly substitutable, the scarce resource is not the capacity to create but the capacity to endorse. The three structural parameters we have measured—a near-linear attachment kernel (α ≈ 1.07), edge persistence of 43–68%, and excess modularity of 0.27–0.32—are not descriptions of a static network. They are the allocation function that will determine which expressions survive the flood. The philanthropic network documented in this paper is the infrastructure through which American arts organizations receive institutional endorsement, and our gateway-donor finding makes the mechanism concrete: recipients entering via high-centrality donors survive at 67.5% versus 53.7% for those entering via lower-degree donors, a gap that narrows to 6.2pp after controls but widens to 31pp among small grants where the endorsement signal, rather than the dollar amount, is the operative variable (Table 12). Under AI abundance, the signal value of institutional endorsement should increase, because there are more alternatives to distinguish among—exactly the condition under which the attachment kernel becomes more consequential.

A falsifiable prediction. Given these parameters, we predict that AI-driven supply expansion will produce increased concentration, not redistribution. The logic is structural. First, a near-linear attachment kernel (α ≈ 1) operating on a larger pool of candidate recipients produces more extreme winner-take-all outcomes—Frank & Cook’s (1995) prediction applied to a measured network.⁴⁵ Second, high edge persistence means donor relationships are robust to supply shocks: the COVID period provides suggestive evidence that donor relationships are stable through operational disruption: edge retention, donor concentration, and recipient survival were statistically indistinguishable between venue-dependent performing arts and museums (chi-square p = 0.55). However, COVID was an operational disruption, not a supply expansion—it did not increase the recipient pool M, which is the mechanism through which our model predicts persistence robustness (via rising λ). The null result is consistent with the broader hypothesis that donor-side relationships are inertial, but it does not test the model’s specific λ(θ) mechanism. (The null result does not imply that performing arts organizations were unaffected—SMU DataArts (2025) documents performing arts revenue declining 27% in real terms over five years, with contributed revenue falling 46% in 2024 alone.³⁸ The network’s structural parameters were robust even as the organizations within it were differentially damaged.) A genuine test requires observing persistence under conditions where M actually expands. Third, high modularity means cross-community bridging—the mechanism by which new entrants gain legitimacy outside their local cluster—remains scarce even as production becomes cheap, because bridging requires relationship capital that cannot be manufactured by generative models. Therefore: the 84/10 split (84% of arts donors contributing 10% of dollars) should intensify, not equalize, as the denominator of cultural production grows while the numerator of endorsement capacity remains fixed. The intensification is already observable without AI: SMU DataArts (2025) documents the deficit rate climbing from 36% (2019) to 44% (2024) while working capital fell from 6.75 to 4.25 months over the same period.³⁸ The legitimation economy is concentrating under current conditions; AI supply expansion would accelerate a dynamic already in motion.

The falsification. If AI abundance instead triggers broad redistribution away from incumbents—if new AI-enabled organizations successfully compete for endorsement from high-centrality donors at rates comparable to established organizations—then the preferential attachment mechanism documented here is weaker than we estimate, and the legitimation economy framework does not generalize to post-abundance conditions. A genuine test requires data that does not yet exist: measuring how the network responds when AI-generated art enters the funding pipeline. We do not claim the e-filing mandate or COVID provides such a test. The mandate changed observability, not supply; COVID was a demand shock, not a supply shock. What this paper contributes is the standing architecture—the wiring diagram of endorsement, persistence, and community structure—against which future supply shocks can be measured. Acemoglu (2024) estimates AI’s aggregate macro effects at a modest 0.66% TFP gain over ten years, but our data suggests the distributional effects within sectors may be far larger.¹⁹ Autor (2024) offers a more optimistic frame, arguing that AI may create new expert tasks accessible to non-experts, but our gateway-donor data suggests the endorsement network may foreclose this possibility in cultural fields.³³ Art is not the best canary for AI’s economic disruption—stock photography, translation, and coding have clearer price signals¹⁸—but it may be the best canary for AI’s legitimation dynamics, because cultural value is socially constructed and cannot be resolved by quality metrics alone. Notably, the art market’s own actors appear not to have registered the supply-side shock. The Art Basel & UBS 2026 dealer survey ranks challenges for the year ahead and five-year horizon: AI does not appear in the top ten. Political volatility (43%), overheads (26%), and tariffs (20%) dominate. Meanwhile, the structural preconditions this paper documents—legitimation scarcity, a demographic gap in the 40–60 collector cohort, cost pressure hollowing out mid-tier dealers, and baby boomer estate flooding secondary markets—are all present and worsening. The supply expansion our model predicts will test this apparatus appears absent from the industry’s own risk assessments.

The question is not “how do we protect artists from AI?” but “what are we optimizing for when a legitimation network decides which expressions survive?” The data above suggest the current answer is: geographic proximity, institutional endorsement, and sustained relationship capital. Whether those are the right criteria for a post-abundance cultural economy is a question this data can inform but not answer.

---

Implications

For arts organizations

The inertia of patronage creates both stability and vulnerability—and the sector’s dependence on it is growing. Before the pandemic, arts organizations derived 42% of revenue from earned sources and 58% from contributions. At the trough in 2021, contributed revenue rose to 71% as earned revenue collapsed. By 2024, the ratio had partially recovered to 40/59—but paid attendance remains 22% below 2019, and performing arts contributed revenue fell 46% in a single year (2024).³⁸ The philanthropic network documented in this paper is not one of several funding channels; for a growing share of arts organizations, it is the primary survival mechanism. Organizations that maintain multi-year relationships with a diverse donor base are structurally resilient; those dependent on a few high-churn donors are fragile. Our network metrics can identify which organizations sit in structurally precarious positions before a crisis reveals it.

For philanthropic strategy

The COVID locality effect suggests that foundations, when uncertain, default to geographic proximity. This is rational but has distributional consequences: arts organizations in states with fewer foundations are doubly disadvantaged during crises. They lose both audience revenue and out-of-state philanthropic support simultaneously.

For arts policy

The e-filing mandate was a regulatory change aimed at tax administration, but it produced a research windfall. For the first time, we can see the most complete network of arts philanthropy in America yet assembled. Policymakers interested in the health of the cultural sector now have access to a real-time structural map of funding flows—if the research community builds the tools to analyze it.

For geographic equity

Funding deserts correlate strongly with states that have fewer headquartered foundations, creating a double disadvantage: these states lack both local philanthropic capacity and the gravitational pull needed to attract out-of-state funding. The locality effect means that the ~60% of arts grants that stay in-state never reach these communities, while the ~40% that cross state lines flow preferentially toward prestige institutions in New York, California, and DC. Without deliberate intervention—whether through place-based grantmaking programs, regional re-granting intermediaries, or federal matching mechanisms—the geographic distribution of arts philanthropy will continue to mirror and reinforce existing concentrations of cultural capital.

For equity in artist representation

The legitimation gap is measurable beyond philanthropy. In the commercial art market, female artists now constitute 45% of gallery rosters but capture only 37% of sales—a persistent −8pp gap that the Art Basel & UBS 2026 report documents across every dealer size tier.⁶³ The gap is inversely correlated with institutional power: the smallest galleries (<$250K turnover) have 55% female representation but a −12pp sales gap, while the largest (>$10M) have 35% representation and a −8pp gap, with female artists’ sales share actually declining from 30% to 27% at the top tier even as representation grew. Endorsement (being on a roster) does not translate equally to allocation (making sales). If the same dynamic operates in philanthropic networks—where some endorsed organizations receive structurally less funding than their endorsement position would predict—the legitimation economy is not merely concentrated but systematically biased in who it rewards. Our data does not include demographic attributes of recipient organizations, but the structural mechanism is content-agnostic: the attachment kernel, edge persistence, and modularity documented in this paper would propagate any systematic bias in initial endorsement decisions through the same compounding dynamics measured in Table 12.

For AI platforms and technology companies entering arts funding

The preferential attachment exponent measured in this network (α ≈ 1.07) describes a concentration mechanism analogous to those governing visibility on content platforms. Platform designers building recommendation systems for AI-generated art face a concrete choice: their attachment kernel either replicates or departs from the wiring rules that have governed cultural allocation for decades. Our gateway donor finding offers a specific design lever: among small grants (under $1,000), recipients endorsed by a high-centrality donor survive at 77.2% versus 45.8% for those endorsed by peripheral donors—a 31pp gap where the endorsement signal, not the dollar amount, determines survival. This is the cold-start problem: on Midjourney or Spotify, a new creator’s survival depends analogously on early algorithmic placement. Our data suggests that who provides the initial endorsement matters more than how much initial exposure is given.

Technology companies are already entering the philanthropic side of this network: Google supports the Lincoln Center, Microsoft funds the Metropolitan Museum’s digital initiatives, Anthropic sponsors LACMA’s Art + Technology Lab, and Meta has backed the Smithsonian’s digitization projects. A technology company’s decision to fund an arts program is a structural intervention in the legitimation network, conferring endorsement centrality that is associated with measurably higher organizational survival (Table 12). The community structure of the network (excess modularity 0.27–0.32) means these endorsements do not diffuse broadly; they strengthen specific funding communities while leaving others untouched. Deliberate cross-community grantmaking—funding organizations outside one’s existing cluster—is the structural intervention most likely to reduce concentration, but our data shows it is also the most expensive: cross-state grants carry a median of $1,000 versus $5,000 for same-state grants, and only high-degree donors routinely bridge geographic communities.

For AI alignment and interpretability research

The structural parameters documented in this paper have a direct bearing on AI alignment and interpretability work. AI language and multimodal models are trained on cultural output that has already been filtered by legitimation networks like the one measured here: the texts, artworks, and critical discourse that reached wide circulation were selected by the attachment kernel (α ≈ 1.07) and edge persistence (43–68%) this paper quantifies. If the same preferential-attachment mechanism operates in training corpus curation as in grant networks—and there is no structural reason it should not, since both are human endorsement systems governed by similar scarcity dynamics—then models trained on internet-scale curated text may systematically over-represent expressions from high-degree nodes in the legitimation network and under-represent the geographic and institutional periphery our data identifies. Interpretability researchers could use the network structure documented here as a reference distribution: do model representations of cultural quality, prestige, and institutional authority track the endorsement hierarchy we measure (community partition Q_excess ≈ 0.27–0.32, gateway-donor survival gap 6.2pp net of grant size) or some more universal signal? The gateway-donor finding makes the alignment stakes concrete: the mechanism that produces a 31pp survival gap among small grants—where endorsement signal dominates dollar amount—is the same mechanism that would propagate through a recommendation system trained on endorsed cultural output. Whether AI systems should inherit the attachment kernel of arts philanthropy, or be deliberately aligned away from it, is a normative question this paper cannot answer. What it provides is the empirical baseline—the measured wiring diagram of endorsement concentration—against which such alignment decisions can be evaluated.

For policymakers designing post-NEA funding mechanisms

If the NEA and NEH are eliminated,¹⁴ the philanthropic network documented here becomes the primary allocation mechanism for which art survives in America. Meanwhile, prominent art schools have closed or are closing across the country.¹⁵ As Reich (2018) argues, this raises democratic concerns: the tax deduction for charitable giving effectively subsidizes the allocation preferences of wealthy donors, and our concentration data shows just how narrow that allocation is.²⁸ Our data reveals three structural properties that any replacement mechanism must address. First, the 84/10 split: 84% of private arts donors contribute only 10% of total dollars, meaning any matching-fund or tax-incentive program that counts donors equally will be dominated by the long tail while the actual allocation remains controlled by the top 16%. Second, the locality effect: approximately 60% of arts grants stay in-state, creating funding deserts in states with few headquartered foundations. A federal mechanism designed to counteract this would need to function as a cross-community bridge—the rarest and most expensive type of edge in our network. Third, the retention gradient: first-time arts donors retain at 58% while five-year donors retain at 89%. Any new funding mechanism that relies on private matching will inherit this structural bias toward incumbents unless it explicitly subsidizes first-time and early-stage donor relationships. The measured parameters suggest that the most effective intervention point is not the volume of funding but the topology of the network: reducing modularity through intermediary re-granting organizations, subsidizing cross-state edges, and designing matching formulas that weight endorsement diversity rather than dollar totals. Quadratic funding (Buterin, Hitzig & Weyl, 2019)⁴¹ offers a concrete mechanism: by matching contributions proportional to the breadth of donor support rather than dollar volume, it would structurally counteract the 84/10 split, giving the long tail of small donors disproportionate matching power—the opposite of the current system’s concentration dynamics.

---

Limitations and Next Steps

These findings are preliminary. Several methodological challenges remain before we can make definitive claims.

Methods and Reproducibility

All data processing and analysis code is open-source and available in the project repository. The pipeline consists of:

1. Extraction: Python XML parser for IRS 990-PF bulk data (22 GB, 2019–2024)
2. Entity resolution: Three-tier name matching (exact, state-constrained fuzzy via rapidfuzz, nationwide fuzzy) against the IRS Business Master File (1.35M organizations)
3. NTEE filtering: Retain grants to organizations classified under NTEE major group A
4. Network construction: NetworkX MultiDiGraph with GraphML export
5. Analysis: Locality (degree-preserving null model, 1,000 permutations), retention (year-over-year with streak decomposition), prestige (PageRank, betweenness centrality with k=5,000 sampling, Gini, HHI), art form survival (NTEE sub-code mapping, COVID impact trajectories), network metrics (degree distribution, connected components)

Computation is performed on a 16-core, 64 GB cloud server. Entity resolution of 907K grants takes approximately 2.5 hours; network construction and analysis for all six years completes in under 5 minutes.

This study reports 18 primary hypothesis tests across findings 1–10. All p-values are FDR-corrected using the Benjamini–Hochberg procedure at q=0.05. The z-scores in Figure 2 are included in this correction set. Individual test results are available in the supplementary materials.

---