Addressing GeoAI Governance: An Automated Gatekeeper for Building Outlines in OpenStreetMap

Abstract

Geospatial Artificial Intelligence (GeoAI) enables the automated generation of built environment map features, such as building outlines/footprints, on a global scale. However, the integration of these AI-generated datasets into Volunteered Geographic Information (VGI) platforms like OpenStreetMap (OSM) risks incorporating ‘AI slop’, consisting of geometrically inconsistent/unreliable data, into the online map. While the OSM “Code of Conduct for Automated Edits” provides a policy framework for data ingestion, it lacks a machine-enforceable mechanism for real-time quality gating. This paper proposes a GeoAI-Gatekeeper to perform this task—an automated process that applies empirical Acceptable Quality Thresholds (AQT) to address the GeoAI data governance problem. Because the Gatekeeper utilizes an intrinsic, no-reference evaluation of geometric fidelity, it can assess incoming AI-generated data streams in real-time without requiring ground-truth benchmarks. Importantly, it focuses exclusively on the geometric validation of building footprints, acknowledging for now that semantic enrichment, such as tagging, remains a human-centric task. The presented GeoAI-Gatekeeper is a working prototype developed for a specific urban area, systematically triaging incoming AI-generated data into three tiers; Auto-Accept, Manual Review, and Reject. It provides a Web-GIS interface for Human-in-the-Loop (HITL) functionality to ensure the OSM community remains the final arbiter of acceptable data quality. Testing the Gatekeeper in Dublin (Ireland) demonstrates that our solution can auto-ingest 93.6% of features with a 14x reduction in human review effort while still adhering to OSM’s cartographic integrity standards. By implementing qualitative community guidelines into machine-enforceable thresholds, our approach introduces a viable methodology for next-generation hybrid VGI systems. Importantly, it ensures that the transition towards automated data ingestion reinforces, rather than undermines, the reliability of global crowd-source mapping datasets.

1. Introduction

1.1. The GeoAI Revolution and Its Discontents

The convergence of computer vision, deep learning, and remote sensing technologies has established an era of automated spatial data production at scales previously unimaginable [1,2]. Such GeoAI systems can now process satellite imagery to extract millions of building footprints in hours, a previously manual task that would require decades of intensive human effort [3]. For example, Microsoft’s Building Footprints dataset alone contains over 1 billion geometries globally [4], while Ecopia AI (https://www.ecopiatech.com/ (accessed on 12 February 2026)) and Google’s Open Buildings Dataset have contributed hundreds of millions more [5,6]. This diffusion of geospatial data production promises to accelerate global mapping initiatives, particularly in data-scarce regions where manual digitization efforts remain incomplete. While the capacity for large-scale spatial data generation remains concentrated among a small number of well-resourced corporations, the advent of multi-source geospatial datasets provides a significant opportunity for maintaining UpToDate online maps.

However, this promise carries a fundamental tension: VGI platforms such as OpenStreetMap (OSM) were architected around human judgment, local knowledge, and community consensus, not algorithmic inference [7,8]. The OSM “Automated Edits Code of Conduct” explicitly warns against “mechanical edits” that lack contextual validation, yet provides no computational framework for enforcing quality standards at scale [9]. As AI-generated data threatens to overwhelm human review capacity, the VGI mapping community faces a governance crisis: how can crowd-source mapping platforms maintain Editorial Control (a core OSM principle) when VGI mappers/editors cannot feasibly inspect every imported feature [10]?

1.2. The “AI Slop” Problem

Recent discourse in the VGI mapping community has identified a specific pathology: ‘AI slop’—which refers to geometric artifacts characterized by excessive vertices, irregular shapes, and topological inconsistencies that emerge from imperfect AI model inferences [11,12]. Unlike traditional human digitization errors (which tend to be sparse and correctable through editing), AI slop exhibits systematic patterns: irregular edges from over-segmentation, phantom buildings from cloud shadows, and elongated distortions from perspective artifacts [13,14,15]. These defects are not merely aesthetic; they degrade routing algorithms, complicate building classification, and erode trust in OSM dataset quality.

Current approaches to quality control such as manual inspection, community flagging, or post-import audits fail to address the scale asymmetry between AI production (millions of features per day) and volunteer review (hundreds of features per day). What is needed is a machine-enforceable governance layer that can triage incoming AI data at source, accepting high-quality contributions while filtering out slop with minimal human intervention [16,17].

1.3. Research Contribution

This paper introduces a GeoAI-Gatekeeper to perform this task, a computational framework that addresses spatial data quality governance through:

1.

Empirical Threshold Derivation: Using the 95th percentile of Shape Quality Metric (SQM) scores from high-quality reference (ground truth) data to establish building type specific Acceptable Quality Thresholds (AQT).

2.

The Three-Tier Triage System: Automatically classifies features into Tier 1 (auto-accept, $\mathit{SQM}<\mathit{AQT}$), Tier 2 (manual review, $\mathit{AQT}\le \mathit{SQM}<1.5\times \mathit{AQT}$), and Tier 3 (reject, $\mathit{SQM}\ge 1.5\times \mathit{AQT}$).

3.

Metric-Aware Processing: Ensuring all geometric calculations use Universal Transverse Mercator (UTM) coordinates (x,y) to manage distortions when projecting spherical 3D geographic coordinates (lat/long) onto a 2D plane.

4.

Interactive Dashboard: Providing VGI mappers with a Web-GIS interface to visualise triage decisions and adjust confidence thresholds based on local context.

Experimental validation on 16,118 GeoAI generated building footprints in the Dublin area demonstrate that 93.6% of buildings can be auto-ingested with minimal false acceptance, representing a 14× reduction in required human review effort.

1.4. Introduction to the Shape Quality Metric (SQM)

The core computational engine in the Gatekeeper framework is the Shape Quality Metric (SQM) which is a multi-dimensional geometric evaluation mechanism developed specifically to capture the cartographic veracity of building outlines [18]. Unlike standard accuracy measures like Intersection-over-Union (IoU) [19] or F1-scores [20], which only measure spatial overlap, SQM assesses the intrinsic geometric properties that define a “realistic” building footprint [21].

SQM is computed (Equation (1)) as a weighted aggregate of six key geometric indices, each targeting a specific failure mode of deep learning models

$$\mathit{SQM}={\displaystyle \frac{(\mathit{perimeter}\times \mathit{elongation}\times \mathit{convexity})}{(\mathit{area}\times \mathit{compactness}\times \mathit{solidity})}}$$

(1)

where:

• Perimeter ($\mathit{P}$) and Area ($\mathit{A}$): Fundamental dimensions used to calculate dimensionless ratios.
• Convexity: The ratio of the polygon area to its convex hull area. This is highly sensitive to shape irregularity and non-orthogonal “stair-like” edges often found in over-segmented GeoAI outputs.
• Solidity: Measures the density of the shape. A low solidity often indicates fragmented predictions or “hallucinated” holes within a building roofline.
• Compactness: Calculated as $\left({\displaystyle \frac{4\pi A}{P^2}}\right)$. This identifies unnaturally thin or elongated artifacts that can occur when an AI model incorrectly merges adjacent buildings or follows shadow projections.
• Elongation: Derived from the oriented bounding box, helping to distinguish between residential dwellings and larger industrial sheds or warehouses.

Figure 1 illustrates varying “complexity signatures” of building types. Simple residential dwellings (top) exhibit high compactness and low SQM scores, whereas complex institutional structures (bottom) naturally possess higher SQM values. Such variation requires a stratified Acceptable Quality Threshold (AQT) to address the false rejection of valid, complex architectural forms.

In our previous study, it was established that SQM could distinguish between authoritative (ground truth) data from Ordnance Survey Ireland (OSi) and crowd-sourced data such as OSM by detecting subtle differences in generalization and vertex density [18]. A key contribution of this new work extends SQM beyond being a descriptive metric used for comparison, to a prescriptive metric used for real-time data governance. By establishing a statistical “baseline of quality” from existing OSM community map feature data, SQM allows the Gatekeeper to answer the central question of VGI governance: Does this AI-generated map feature look like a manually digitized building?

To implement such a systematic governance mechanism, we introduce a Three-Tier Triage System that translates statistical SQM scores into definitive administrative actions. Rather than providing a binary ‘pass/fail’ output, this system categorizes AI-generated features into three distinct workflows (i.e., Auto-Accept, Manual Review, and Reject) based on their proximity to empirically calculated Acceptable Quality Thresholds (AQT).

By automatically ingesting and flagging high-fidelity geometries as “Ready for Semantic Qualification”, while at the same time isolating ‘AI slop’ for rejection, the Gatekeeper framework effectively triages the data stream to protect the map’s integrity. Importantly, this tiered approach preserves the Human-in-the-Loop philosophy of VGI; it optimizes the limited bandwidth of the mapping community by routing only the most ambiguous cases to human editors, thereby transforming the geometric validation process from a manual burden into a scalable, supervised oversight mechanism.

2. Background

2.1. The Quality Gap in GeoAI and VGI

Current literature on VGI quality emphasizes that data quality is multi-dimensional, encompassing positional, thematic, and temporal accuracy [22,23]. However, there is a distinct lack of research focusing on cartographic veracity—the degree to which a vector outline mimics the intentional, generalized, and orthogonal nature of human-digitized map features.

Related work using indices such as Intersection-over-Union (IoU), F1-Score, and Root Mean Square Error (RMSE) have become the industry standard for benchmarking Convolutional Neural Network (CNN) and Generative Adversarial Network (GAN) quality [3,24,25,26,27]. Some GeoAI evaluations in particular focus predominantly on pixel-level or overlap-based metrics [21,24]. While these metrics are effective for assessing an AI model’s ability to locate a feature in geographic space, they are fundamentally “shape-blind” and lack awareness of contextual form [28].

For example, as noted in [21], a building footprint can achieve a high IoU (e.g., >0.90) while possessing hundreds of redundant or irregular vertices that would fail the “eye-test” of a professional cartographer (or a VGI mapper). In the context of OSM, this results in ‘AI Slop’—data that is spatially accurate in its location but topologically and geometrically unusable for many traditional GIS/LBS applications like directional querying, routing, or urban planning [29].

2.2. The OSM Code of Conduct and the Scalability Crisis

The OpenStreetMap Foundation maintains a strict “Code of Conduct for Automated Edits”, built on three pillars: documentation, community consultation, and caution [9]. The “caution” principle essentially serves as a brake on mass-importing AI data, requiring that automated edits must not degrade the existing quality of the map.

Historically, quality control in OSM has been a reactive, human-dependent process [22,30]. Tools such as Osmose (https://wiki.openstreetmap.org/wiki/Osmose (accessed on 12 February 2026)) and KeepRight (https://keepright.at/ (accessed on 12 February 2026)) flag errors after they have been ingested into the database, while platforms like MapRoulette (https://maproulette.org/ (accessed on 12 February 2026)) rely on crowdsourcing the labour of manual verification. However, online crowd-source mapping is currently facing a Scale Asymmetry: GeoAI can generate building footprints at a rate of millions per day, while the global volunteer community of mappers can only realistically review a few thousand features in the same period. This “Reviewer Bottleneck” means that without a machine-enforceable governance layer, the VGI community is forced to choose between embracing AI at the cost of data integrity or rejecting AI to preserve trust. To bridge this gap, we propose that VGI mappers require a proactive “Gatekeeper” to address the “caution” principle before data ingestion occurs.

2.3. The Role of SQM

The Shape Quality Metric (SQM) was introduced to fill a gap left by contemporary overlap-based metrics as a more comprehensive geometric measure of overall building data quality. By aggregating six variables (i.e., Perimeter, Elongation, Convexity, Area, Compactness, and Solidity), SQM creates a “Geometric Signature” for every building footprint [18].

The selection of SQM as the primary evaluative framework, over traditional metrics such as Intersection over Union (IoU), Root Mean Square Error (RMSE), or Hausdorff distance, is rooted in the requirement for a ‘No-Reference’ quality assessment. Most established geospatial metrics are extrinsic, meaning they require a high-accuracy ‘ground truth’ dataset to measure spatial deviation. In the context of real-time VGI ingestion, such a reference is rarely available. SQM, by contrast, provides an intrinsic assessment of geometric ‘plausibility’ by analysing the inherent geometric properties of a polygon (building footprint). This allows our GeoAI-Gatekeeper to recognise AI slop, such as overly rounded corners or redundant vertices, simply by comparing the feature’s statistical signature against the known benchmarks of human-mapped buildings, regardless of spatial overlap with other data sources.

In our previous work, SQM was used primarily as a descriptive tool—a way to compare the output of different AI models or to benchmark OSM against authoritative datasets (such as OSI) [18]. In this paper, we propose a novel shift to this approach: by framing SQM as a decision instrument. By establishing an empirical Acceptable Quality Threshold (AQT), we move from saying “this shape is too complex” to “this shape is safe for OSM.” This transition aligns with international standards such as ISO 19157 (Geographic Information—Data Quality), which advocates for objective, quantitative quality measures [31]. By using SQM as a quantitative gating mechanism, a governance framework can be created where AI models are granted a “License to Upload” only if their output matches the morphological signature of high-quality, human-verified benchmarks.

2.4. Governance Through Triage: The Three-Tier Model

Recent trends in HITL systems suggest that total automation is rarely the goal for high-stakes public data. Instead, the literature points toward triage models to automatically handle low-risk tasks while flagging high-risk cases for human intervention [32,33,34].

The GeoAI-Gatekeeper presented in this paper continues this trend by introducing a Tiered Governance Pipeline for VGI mapping. For example, unlike previous “binary” filters that either accept or reject a polygon feature, a tiered approach respects the subjective nature of mapping diverse built environment features. In Tier 2 (Manual Review), the framework acknowledges that architectural complexity (e.g., an historic cathedral) may look like “slop” to a simplistic algorithm but is valid to the human eye. This ensures that the Gatekeeper preserves the Editorial Control of the OSM community while significantly reducing the “noise” of irregular residential footprints that currently dominate AI-generated datasets.

3. Methodology

This section details the operational logic and technical implementation of the proposed GeoAI-Gatekeeper framework. The methodology is structured to transition the Shape Quality Metric (SQM) from a descriptive comparison tool to a prescriptive governance instrument.

The proposed GeoAI-Gatekeeper is organized into three interconnected phases: (1) Empirical Calibration; (2) Automated Governance; and (3) Human-in-the-Loop verification. The end-to-end Gatekeeper system architecture and dataflow between its functional modules is shown in Figure 2.

3.1. Study Area and Baseline Calibration

The experimental validation was conducted in Dublin, Ireland, focusing on a dense urban-to-suburban transect defined by the bounding box

$$53.{330}^{\circ }\mathrm{N},-6.{296}^{\circ }\mathrm{W}\wedge 53.{363}^{\circ }\mathrm{N},-6.{224}^{\circ }\mathrm{W}$$

Reference Dataset Profile

To build the AQT benchmarks, we utilized 52,487 OSM building footprints (126 MB GeoJSON) extracted from the live OSM database.

Table 1. Morphological calibration and AQT scores by building type.

Geometric Signature	P99 Threshold	Mean SQM (AQT)	Count	Building Type
Standard orthogonal/rectangular	1.26	0.52	31,204	Residential
High compactness, large area	1.05	0.48	8913	Commercial
Variable elongation	1.89	0.63	1247	Industrial
Multi-unit clusters	1.34	0.55	2856	Apartments
Tightly bound, compact	0.98	0.46	987	Retail
Highly elongated, large perimeter	2.14	0.71	342	Warehouse

illustrates the distribution of these buildings and their resulting SQM benchmarks.

A notable finding during calibration was that warehouses and industrial buildings exhibit significantly higher baseline SQM scores (0.71 and 0.63 respectively). This is not an indication of lower quality, but a reflection of their elongated forms and larger perimeters, which naturally increase the SQM complexity driver. Conversely, retail buildings showed the tightest distribution (AQT = 0.46), indicating a highly standardized, compact morphology in the Dublin region.

3.2. Data and Resources

A robust governance framework requires two distinct data pipelines: a high-fidelity reference set to define community map quality standards and a diverse set of AI-generated outputs to test the Gatekeeper’s diagnostic precision [35].

3.2.1. Reference VGI Data

To establish the Acceptable Quality Threshold (AQT), a comprehensive reference dataset was extracted from OSM covering the Dublin metropolitan area. To ensure this data represented the “best-in-class” of human mapping, we applied a rigorous filtering protocol:

• Temporal filter: Only features with a last_edit timestamp after 1 January 2023, were included. This ensures the data was mapped or validated against the most recent high-resolution aerial imagery available in the region.
• Integrity filter: Features flagged with active “OSM Notes” or those lacking a primary building tag were excluded.
• Volume: This resulted in a reference set of 52,487 building footprints, providing a statistically significant sample size for urban morphology analysis.

3.2.2. GeoAI Experimental Data

The AI-generated footprints used to test the Gatekeeper were acquired from our previous benchmark studies, utilizing two distinct architectural approaches to building extraction:

• OSM-GAN and Poly-GAN: A Generative Adversarial Network optimized for generating sharp, rectilinear building edges by learning from OSM structural patterns [36].

3.3. Scope and Geometric Focus

It is essential to define the Gatekeeper’s operational boundaries prior to detailing the triage logic. A building feature in OSM comprises two distinct components: the geometry (the nodes (vertices) and ways (edges) defining the footprint) and the semantics/attributes (tags such as building:levels or addr:street).

The GeoAI-Gatekeeper is exclusively a geometric validator. We posit that while AI can efficiently outline features from pixels, semantic attribution frequently requires multi-modal data or local human knowledge [37]. By focusing on object position, orientation, and shape, the Gatekeeper addresses the most labour-intensive aspect of VGI: the accurate digitization/representation of building outlines. Consequently, the framework is designed to protect the “cartographic container,” ensuring it is topologically and geometrically sound, thereby allowing the VGI community to focus their efforts on high-value semantic tagging.

It is important to re-emphasize that the “Auto-Accept” tier functions strictly as a geometric validator. Until GeoAI can reliably automate semantic profiling, it remains the responsibility of the OSM mapper to add relevant tags and attributes to these accepted features, maintaining the usual human-centric workflow for entering non-geometric data.

3.4. Empirical Threshold Derivation (AQT Discovery)

Rather than imposing SQM thresholds universally, the Gatekeeper is calibrated against the benchmark reference data acquired in Section 3.2.1. This process ensures the AQT is not an arbitrary number but a reflection of local OSM community mapping standards.

3.4.1. Stratification by Typology

Building footprints were grouped by their building tag (e.g., residential, commercial, industrial). This accounts for legitimate geometric variation; for instance, large-scale industrial warehouses are inherently more elongated and possess different area-to-perimeter ratios than detached suburban houses [38].

3.4.2. Statistical Calibration

Recognizing that urban morphology varies, where a simple rectangular house and a complex industrial warehouse have inherently different geometric “signatures”, the reference data was stratified by typology (e.g., residential, commercial, industrial). For each building type T, the 95th percentile of SQM scores was computed

$${\mathit{AQT}}_T=P_{95}\left(\left\{\mathit{SQM}\left(b\right)\vee b\in {\mathit{Buildings}}_T\right\}\right)$$

(2)

This choice invokes the ‘95% Confidence’ principle—establishing that an AI-generated footprint is eligible for automated upload to OSM only if its geometric quality aligns with the standards maintained by 95% of human-verified data. This approach facilitates the rapid integration of GeoAI footprints with minimal human verification. The remaining 5% of reference data typically contains valid but highly irregular “edge cases” (e.g., heritage sites) which the framework intentionally routes for manual review.

3.5. The Three-Tier Triage System

The core of the governance framework is a triage pipeline that routes each AI-generated feature (f) based on its computed SQM(f) relative to the empirical $AQ{T}_T$ of its type.

• Tier 1: Auto-Accept ($\mathit{SQM}\left(\mathit{f}\right)>{\mathit{AQT}}_{\mathit{T}}$)
  Features in this tier are geometrically indistinguishable from high-quality human mapping. These are accepted without human intervention, significantly reducing reviewer burden.
• Tier 2: Manual Review (${\mathit{AQT}}_{\mathit{T}}\le \mathit{SQM}\left(\mathit{f}\right)<\mathbf{1.5}\times {\mathit{AQT}}_{\mathit{T}}$)
  Features exhibiting borderline quality are flagged for mapper inspection. This tier captures features that are either valid complex architectures or “near-miss” AI artifacts. The context-sensitive nature of these shapes requires the “eye-test” of a human editor to determine their suitability for upload to the online database.
• Tier 3: Reject/AI Slop ($\mathit{SQM}\left(\mathit{f}\right)\ge \mathbf{1.5}\times {\mathit{AQT}}_{\mathit{T}}$)
  Automatic rejection is triggered for features that deviate so significantly from the community norm that their ingestion would degrade the integrity of the online map. This tier identifies systemic ‘AI slop’ such as buildings with irregular perimeters or shadow-hallucinated extensions.

The 1.5 × buffer for Tier 2 was determined through iterative empirical testing; tighter thresholds (e.g., 1.2×)resulted in excessive false rejections of valid complex shapes, while looser thresholds (e.g., 2.0×) admitted obvious geometric artifacts into the manual review queue, wasting volunteer time.

3.6. Human-in-the-Loop: The Web-GIS Decision Interface

The Gatekeeper framework is not designed as a “black box” automation tool, but as a decision-support system. To facilitate the Tier 2 (Manual Review) workflow, a Web-GIS interface was developed to serve as the critical bridge between algorithmic filtering and human editorial judgment (See Figure 3).

• Visualizing uncertainty—Features flagged as Tier 2 are presented to the user in a specialized validation environment. Unlike standard GIS editors, this interface overlays the Geometric Signature (vertex distribution) of the GeoAI footprint against the underlying high-resolution imagery. Visual cues, such as color-coded vertices, highlight the areas where the $\mathit{SQM}$ detected “anomalies” or “non-convexity,” allowing VGI mappers to quickly identify the failure mode (e.g., a shadow-distorted edge).
• Adjustable confidence governance—A core feature of the HITL interface is the Confidence Slider. Recognizing that different mapping campaigns have different risk tolerances, mappers can dynamically adjust the $\mathit{AQT}$ percentile (e.g., from the conservative $P_{99}$ to a more permissive $P_{80}$). As the slider moves, the map updates in real-time, re-classifying features between Auto-Accept and Manual Review. This empowers local OSM chapters to “tune” the gatekeeper based on the specific urban morphology or the intended use-case of the data.
• Feedback loop for semantic enrichment—As users validate Tier 2 geometries, the interface facilitates the transition from Geometry to Semantics. Once a shape is “Accepted” by the user, the interface prompts for basic semantic tagging (e.g., building=yes), ensuring that while the geometry was AI-generated, the final feature enters the OSM database with a “human-verified” seal of approval.

4. Experimental Results

4.1. Processing Results and Triage Performance

The Gatekeeper framework processed 16,118 AI footprints in 14.2 s on a standard M1 Mac, demonstrating that the SQM-based governance layer introduces negligible latency into the data ingestion pipeline.

Table 2. Final triage results (AI-generated dataset for Dublin).

Percentage	Count	Action	Triage Tier
93.6%	15,080	Auto-Accept	Tier 1
4.8%	775	Manual Review	Tier 2
1.6%	263	Rejected (Slop)	Tier 3

illustrates the final distribution of AI-generated footprints across the three-tier triage system, highlighting the Gatekeeper’s capacity to automatically validate most of the incoming data stream.

Experimental results (Table 2) indicate that 93.6% of the AI-generated features were geometrically indistinguishable from the OSM reference community data. Under the Gatekeeper framework, these would then be available for auto-upload into OSM with zero human intervention while still flagged for further manual semantic edits. Only 4.8% were routed to the Web-GIS interface for human oversight of their geometric properties, while 1.6% were immediately discarded as ‘AI Slop’.

4.2. Diagnostic Accuracy and Rejection Case Studies

To validate the Gatekeeper’s rejections, a manual audit was conducted on a random sample of 50 building features from Tier 3 (Rejected). The audit identified three distinct failure modes:

1.

Irregular edges (76%): 38 cases featured simple rectangular structures characterized by geometric artifacts. These edges were corrupted by over 100 redundant vertices resulting from high-frequency spatial noise (“stair-like” edges). The $\mathit{SQM}$ effectively identified these errors by penalizing the resulting high perimeter-to-area ratio, which deviates significantly from the expected form of a standard (rectangular) building footprint.

2.

Structural merging (18%): 9 cases showed extreme elongation (aspect ratios > 10:1) where the “AI” (simulated jitter/overlap) appeared to merge adjacent buildings into a single distorted polygon.

3.

False rejections (6%): 3 cases were identified as legitimate, highly irregular courtyard apartments. Their natural convexity deviations caused them to exceed the $\mathit{AQT}$.

4.3. Threshold Sensitivity Analysis

The confidence percentile used in the AQT calculation was varied to determine the optimal “Community Risk” setting (

Table 3. Sensitivity analysis of confidence threshold. (the $P_{95}$ baseline is highlighted as the optimal balance between automation volume and geometric quality control).

Tier 3 (Reject)	Tier 2 (Review)	Tier 1 (Accept)	AQT (Avg)	Confidence
0.7%	2.1%	97.2%	0.42	$P_{80}$
1.1%	3.8%	95.1%	0.48	$P_{90}$
1.6%	4.8%	93.6%	0.52	$P_{95}$ (Baseline)
3.5%	8.2%	88.3%	0.64	$P_{99}$

).

As the threshold moves from $P_{80}$ to $P_{99}$, the system becomes increasingly “strict.” While $P_{99}$ ensures almost zero “slop” enters the database, it nearly doubles the manual review burden (from 4.8% to 8.2%). We suggest that $P_{95}$ (highlighted in Table 3) represents the “Goldilocks Zone” for urban data (in Dublin)—balancing high-volume automation with a rigorous defence against geometric degradation of the resulting online map.

This ‘Sensitivity-by-Design’ approach allows the Gatekeeper to be tuned according to a community’s risk tolerance. While the $P_{95}$ baseline is empirically optimal for our study area, a full cross-regional sensitivity analysis to determine how these thresholds shift in different urban morphologies remains a subject for further research. In different contexts, such as informal settlements versus planned suburbs, the distribution of ‘plausible’ human mapping may vary, yet the theoretical logic of using a high-percentile statistical norm as an automated data quality gate remains. For this initial framework, prioritizing precision by minimizing false accepts over recall by maximizing automation ensures that the integrity of the OSM database is protected regardless of the specific threshold value chosen.

4.4. Comparison to Existing Workflows

The VGI mapping potential of the proposed GeoAI-Gatekeeper is best understood through a comparison to human labour requirements.

• Scenario A (total manual review): Inspecting all 16,118 features at 30 s/feature would require 134 h (3+ weeks) of volunteer labour.
• Scenario B (one-click confirmation—e.g., RapiD [39]): Confirming features at 5 s/feature still requires 22+ h (~3 days) of active manual attention.
• Scenario C (GeoAI-Gatekeeper): With 93.6% auto-accepted, humans only need review 775 features at 30 s/feature. Total time: ~6.5 h.

The GeoAI-Gatekeeper provides a 95% reduction in labour compared to traditional manual review and a 50% reduction compared to current state-of-the-art “AI-assisted” tools. This efficiency allows the VGI community to pivot from the tedious task of ”fixing shapes" to the high-value task of semantic enrichment and local knowledge verification.

5. Discussion and Future Research Directions

5.1. Governance Implications for OSM

The OSM Automated Edits Code of Conduct mandates that edits occur only after informed community consultation. The GeoAI-Gatekeeper presented in this paper provides a technical basis to enforce this mandate computationally. We propose that by shifting the “burden of proof” from the volunteer mapper to the algorithm, OSM can confidently move from a reactive audit model to a proactive governance model.

• Empirical standards: The calibration phase encodes local community norms into statistical distributions ($P_{95}$), ensuring AI data is judged by the same standard as human mappers.
• Graduated response: The three-tier triage replaces binary “all-or-nothing” imports with a nuanced workflow that defaults to acceptance only for high-confidence features.
• Algorithmic accountability: Through the audit log, every automated decision is transparent and reproducible, satisfying the need for accountability in VGI mapping ecosystems.

5.2. Limitations and Morphological Edge Cases

An analysis of the 6% false rejection rate in Tier 3 reveals that purely geometric metrics, while robust, cannot perfectly capture all architectural expressions.

• Complexity vs. slop: Highly irregular heritage buildings, courtyard structures, and L-shaped apartments often exhibit convexity defects similar to AI slop.
• Semantic orthogonality: We suggest that geometric governance can be set apart from semantic validation. While SQM can confidently measure that the “container” is cartographically sound, mappers retain the essential responsibility for thematic accuracy (e.g., distinguishing a church from a house).
• Regional generalization: Dublin’s relatively homogeneous building stock (e.g., Victorian rows) allows for stable AQT benchmarks. Applying this framework to informal settlements or rural traditional architectures will require finer-grained stratification and local re-calibration.

5.3. Future Work: Toward a Multi-Modal Gatekeeper

To address current imposed restrictions on the functionality, future iterations of the Gatekeeper could evolve from a geometry-only data quality tool to a multi-modal assessment engine:

1.

Integration into OSM Production: Beyond technical enhancements, a key next step involves investigating the practical integration of the GeoAI-Gatekeeper into the online OSM mapping ecosystem. While a full integration roadmap is beyond the scope of this paper and would require close collaboration with OSM core developers, future research should explore the socio-technical challenges of such a rollout. This includes analysing the trade-offs between centralized quality gating and the decentralized nature of OSM, as well as gauging the reactions of the diverse contributor base. For instance, corporate contributors, who often handle large-scale data imports, might welcome the efficiency of an automated gatekeeper, whereas individual ’craft’ mappers may harbour concerns regarding the erosion of manual editorial control. Evaluating how to implement this framework as a standard ’pre-ingest’ plugin for editors like iD or JOSM remains an open problem for future study.

2.

Pixel-level fusion [40]: Integrating AI-model confidence maps (probability scores) with SQM metrics to identify “hallucinated” features that might otherwise appear geometrically plausible.

3.

Adaptive learning [41]: Transitioning from static percentiles to an active learning loop. By tracking mapper acceptance rates in Tier 2, the system could dynamically adjust the AQT by lowering thresholds if manual reviews consistently result in acceptance.

4.

Cross-domain expansion: Extending the SQM logic to other OSM feature types, such as road/cycle networks and land-use boundaries (e.g., parks), to provide a more comprehensive “Quality Gate” for all automated map feature imports.

5.

Semantic validation via Multimodal LLMs (MLLMs) [37,42,43]: While the current framework excels at geometric gatekeeping, it remains “thematically blind.” A future iteration could integrate Multimodal Large Language Models (MLLMs) to perform cross-modal semantic validation. By feeding the model both the AI-generated polygon and corresponding high-resolution or street-level imagery, the system could verify if the assigned tags (e.g., building = retail) are visually consistent with the structure’s appearance (e.g., presence of signage, storefronts, or parking lots). This would help to address the Semantic attribution gap identified in this study, ensuring that a building is not only the right shape but also carries the correct functional attributes.

6.

Threshold Sensitivity and Transferability: Another key area for future investigation is the empirical sensitivity analysis of the tiered pipeline across diverse geographic contexts. While the current thresholds were derived from the morphological signatures of a European city (Dublin), further research is required to determine if this holds true for informal settlements, high-density megacities, or traditional rural architectures. Future studies could focus on ‘Cross-Context Calibration,’ testing how thresholds must be adjusted to accommodate different urban forms and whether a global universal threshold is even achievable or if local re-calibration remains a prerequisite for deployment in new regions.

6. Conclusions

This research addresses the critical “Scale Asymmetry” between today’s rapid production of GeoAI building footprints and the time-consuming manual validation capacity of the VGI mapping community. Guided by the OpenStreetMap Code of Conduct for Automated Edits, this study delivers four primary contributions to the field of crowdsource mapping and geospatial data governance:

1.

A geometry-first Gatekeeper framework: We developed and implemented an AI model-agnostic validation pipeline that successfully decouples geometric fidelity from semantic attribution. This framework ensures that the “cartographic container” of a building meets local VGI community mapping standards before human mappers invest effort in thematic enrichment.

2.

Empirical AQT derivation via the 95% confidence principle: This study moved beyond arbitrary quality metrics by introducing a reference-driven calibration method. Using a “benchmark” dataset of over 52,000 human-verified (OSM) footprints in Dublin, we demonstrated that an automated system could derive local, culturally relevant Acceptable Quality Thresholds (AQT) that reflect actual community driven map accuracy standards.

3.

The three-tier triage system: We propose and validate an automated triage logic approach—comprising Auto-Accept, Manual Review, and Rejection tiers—that successfully distinguishes between legitimate architectural complexity and ‘AI slop’. This mechanism provides a scalable solution for managing the influx of huge GeoAI generated datasets while still maintaining a high level of individual feature and overall map integrity.

4.

A Human-in-the-Loop (HITL) governance bridge: Through the development of a Web-GIS mapping interface, we demonstrate that automated gatekeeping does not necessitate the removal of human agency. Instead, it empowers the community by providing visual “quality signatures,” allowing mappers to act as high-level auditors rather than manual digitizers.

The empirical results from our Dublin study are significant. With a 93.6% auto-acceptance rate and a 1.6% rejection rate for geometric slop, the GeoAI-Gatekeeper achieved a 95% reduction in human labour hours compared to traditional manual review/edit workflows. These findings confirm that the framework can process thousands of features in seconds, matching the speed of GeoAI inference with a machine-enforceable governance (data quality) layer.

In summary, adopting a GeoAI-Gatekeeper into the online mapping workflow represents a fundamental shift in contemporary VGI methodology: from reactive manual editing to proactive automated governance. While limitations remain regarding the capture of highly irregular heritage structures, the framework provides the necessary technical foundation for a new type of collaborative mapping—one where AI handles the labour of digitization and humans retain the authority of validation. As global mapping initiatives increasingly rely on automated feature extraction, our Gatekeeper framework offers a repeatable, transparent, and community-centric blueprint for ensuring that the global map remains a true representation of our real-world built environment.