Corporate editing (CE) — the large-scale mapping activity of paid teams employed by technology companies such as Meta, Grab, and Kaart — has become a prominent mode of OSM data production, deploying semi-automated and AI-assisted workflows at scales that far exceed typical volunteer contributions [1,2]. Concerns have been raised about CE's long-term consequences for the project's volunteer-driven character, contributor diversity, and data quality [3,4], yet empirical evidence remains fragmented, focusing primarily on single outcomes [5] or short observation windows. A systematic, multi-dimensional evaluation of CE's long-term impacts on OSM's collective production dynamics is lacking.
We address this gap through the Collective Intelligence (CI) framework. Previous work has already shown the value of analysing OSM-related mapping systems as collective intelligence systems, where contributor composition, collective action, and evidence of intelligent outcomes must be considered together rather than treated as separate dimensions of data production [6]. CI is contingent on four enabling conditions: independence of contributors, diversity of opinion, decentralization of decision-making, and an effective aggregation mechanism [7]. These conditions correspond directly to the tensions raised by CE: scale-intensive contributions may enhance aggregation while simultaneously reducing independence or diversity. Framing the analysis in CI allows us to interpret simultaneously improving and deteriorating dimensions, and ask whether CE sustains, erodes, or transforms the collective production dynamics that make OSM valuable as shared geographic infrastructure.
We operationalize CI using six measures derived from the OSM data model, applied to road (highway=) and building (building=) features. Three capture data quality as an aggregation outcome: entity count (coverage and completeness), mean tag richness per entity (semantic content, restricted to tags within the top 25th percentile of global co-occurrence frequency to exclude noise), and building geometric complexity (1 minus the ratio of footprint area to convex hull area, reflecting spatial detail). Two capture diversity and innovation: the Shannon diversity index over tag distributions and unique tag counts (distinct tag keys per entity type per spatial unit). One captures independence and decentralization: Shannon evenness of tag distributions, where low values indicate convergence toward homogeneous mapping practices. These yield 11 individual measures, combined into a composite CI score via the EWM-TOPSIS objective weighting approach.
The framework covers five Southeast Asian countries — Thailand (TH), Malaysia (MY), Myanmar, the Philippines, and Papua New Guinea — over 2015–2025. Annual observations represent the OSM database state on 1 January of each year, extracted via the OHSOME API [8]. CE changesets were identified from the OSM API using hashtags documented on organized editing Wiki pages for the three most active corporations in the region: Meta/Facebook (#nsroadimport, #mapwithai), Grab (#grab, #grabosm), and Kaart (#kaart). TH and MY — the focal countries of the Facebook AI-Assisted Road Tracing project [9] — serve as primary case studies; the remaining three countries provide comparison contexts at the country level only.
Country-level analysis is descriptive. Province- and state-level causal modeling for TH (77 provinces) and MY (16 states) uses Jordà-style local projections (LP) at horizons h = 0, …, 4 years, with log(1 + CE changesets) as the treatment variable, region and year two-way fixed effects, and lagged outcome and treatment terms. Two CE waves are identified per country — Wave 1 (Facebook/Meta: 2018–2019 in TH, 2019–2020 in MY) and Wave 2 (Grab/Kaart: 2024–2025 in both) — and estimated via interaction terms marking deviations from the baseline treatment slope. A pooled model adds country × period indicators to prevent country-specific phase shifts from driving pooled estimates. For TH, spatial spillover effects are modeled by adding spatially lagged CE exposure under k-nearest neighbors and inverse-distance weight matrices.
Country-level time series results (Figure 1) show that CE-intensive countries do not follow trajectories meaningfully distinct from comparison countries. The Philippines — driven by an active local community with minimal corporate involvement — consistently outperforms all others on the composite CI score, pointing to community-driven mapping as a viable and potentially more effective pathway to collective intelligence.
Province-level analysis for TH (Figure 2) overturns several country-level findings — notably, highway evenness recovers to pre-Wave 1 levels during Wave 2 — and exposes strong temporal autocorrelation that motivates causal modeling. State-level results for MY follow a similar pattern, identifying the first wave as a turning point for several highway measures.
LP results are presented as heatmaps with rows corresponding to CI measures and columns to horizons h = 0–4 (Figure 3). Three panels are shown side by side: baseline CE effects (βh), Wave 1 interaction (βh,w1), and Wave 2 interaction (βh,w2). Warm colors indicate positive coefficients, cool colors negative; asterisks mark significance levels (* p < 0.1, + p < 0.05, † p < 0.01). Significant baseline effects are predominantly positive, concentrated on innovation (unique highway tag counts) and completeness, and are fully absorbed by h = 3. Wave 1 attenuates these effects and reduces decentralization (negative building evenness coefficients persisting to h = 2); Wave 2 amplifies positive effects across most highway measures. Results for MY replicate the wave contrast but differ in significance patterns and coefficient signs, illustrating context dependence. Spatial spillover effects are limited and rarely reach statistical significance.
These findings carry several implications. CE's impacts are real but bounded: absorption within approximately two years suggests OSM functions as a resilient self-correcting system. The wave contrast is consistent with a learning trajectory in which corporate teams progressively align with community norms — a process best accelerated through deeper integration into OSM's community structures. The superior CI performance of the Philippines relative to CE-intensive countries reinforces the value of community-driven contributions. The risk of corporate disengagement, illustrated by the establishment of the Overture Maps Foundation [10] and a concurrent decline in CE activity [4], underscores the importance of institutional structures that reduce OSM's dependence on any single contribution mode. Intra-OSM platforms for cross-team knowledge exchange would reduce the repetition of harmful practices and help channel corporate resources — career pathways, technical infrastructure, and mentorship capacity — toward contributor retention [11], one of OSM's central sustainability challenges.
The codebase underlying this analysis is currently being prepared for open publication and will be made fully available under an open source license in time for the conference.