PaperlayerView exact revision on arXiv

Source-linked AI summary

Toward Governance-Aware Autonomous GIS: A Narrative Review of Ethical and Privacy Risks in LLM-Enabled GeoAI

Maya Subramanian, Devika Jain

arXiv:2609.16232v1cs.AI

TL;DR

LLM-enabled GeoAI raises governance risks including hallucinated spatial facts and passive location inference, while spatially specific verification and grounding methods are largely absent from the reviewed literature. This narrative review synthesizes eight recurring issues and maps them to proposed controls and auditable artifacts across a geospatial data-lifecycle architecture. The review finds uneven response maturity and limited field testing of proposed controls.

  • Problem

    Spatially specific verification and grounding methods for generated spatial claims are largely absent from the reviewed literature, alongside risks such as passive location inference and hallucinated routes.

    Source: Illustrative Example (hypothetical… · pp. 12, 25, b. Spatial Privacy and Location… · p. 13

  • Method

    The review synthesizes recurring governance issues and maps them to implementable controls, enforcement points, and auditable artifacts in a proposed LLM-enabled GeoAI architecture.

    Source: §IV · p. 37, §VI · pp. 53–54

  • Results

    Response maturity varies: spatial mechanisms have established GIScience diagnostic practices, while LLM-specific technical risks are barely recognized as a distinct problem and largely unaddressed.

    Source: §VI · pp. 53–54

  • Takeaways & Limitations

    The proposed architecture explicitly links spatially specific risk mechanisms to auditable controls across the geospatial data lifecycle.

    Source: §IV · p. 37, §VI · p. 54

    Proposed controls remain largely unvalidated or unfield-tested, and spatially specific verification and grounding methods are largely absent from the reviewed literature.

    Source: Illustrative Example (hypothetical… · p. 25, §IV · p. 37

Abstract

from arXiv · show

Geospatial artificial intelligence (GeoAI) powered by large language models (LLMs) is expanding the capacity to query, generate, and interpret spatial information through natural-language interfaces and agentic autonomous GIS workflows. This capability creates governance challenges that general AI ethics discussions do not fully capture, including passive location inference from mobility traces, spatially structured bias amplification driven by spatial autocorrelation and scale effects, hallucinated spatial facts, and uncertainty compounding across multimodal geospatial inputs. This narrative review identifies eight recurring issues in LLM-enabled GeoAI: data provenance and consent, spatial privacy and inference risk, algorithmic bias and spatial inequity, spatial mechanisms as structural risk (spatial autocorrelation, the modifiable areal unit problem, and scale effects), LLM-specific technical risks, explainability, policy and regulatory gaps, and public enablement and workforce development. For each issue, we characterize the underlying mechanism, ground it in an illustrative example from the literature, and assess the current state of technical or institutional responses, ranging from largely unaddressed to actively debated or subject to emerging policy. Building on this synthesis, we propose a governance-aware architecture for LLM-enabled autonomous GIS that maps each issue to enforceable controls and auditable artifacts across the geospatial data lifecycle, illustrated through a worked flood-response routing scenario. The review highlights a persistent evidence gap: proposed responses remain largely conceptual, and field-tested evaluations of governance controls for LLM-enabled GeoAI remain limited. We close by outlining a research agenda emphasizing empirical validation, spatially specific interpretability tools, and workforce training aligned with these emerging risks.

I. Introduction and Motivation

LLMs expand access to spatial analysis through natural-language interaction, while creating governance risks that include location inference, spatial inequity, consent, and uncertainty. This narrative review synthesizes governance issues and responses and proposes controls for autonomous GIS workflows.

I. Introduction and Motivation

  • Natural-language interfaces let non-technical users query and interpret heterogeneous spatial data, potentially democratizing access to spatial intelligence.LLMs can interpret spatial queries, synthesize information across datasets, and produce narrative explanations.Source: §I · pp. 2–3
  • LLM-enabled GeoAI can intensify location-inference privacy threats, spatial inequity, consent challenges, and difficulty communicating uncertainty and decision logic.These risks arise as large spatial datasets are combined with generative or agentic AI.Source: §I · p. 3
  • The review organizes eight recurring governance issues by mechanism, illustrative literature example, and current technical or institutional response.The issues span provenance and consent, privacy, bias, spatial mechanisms, LLM-specific risks, explainability, policy, and workforce development.Source: §I · p. 4

II. Methodological Approach

  • The narrative review covers literature from January 2020 to January 2026 and synthesizes representative sources rather than claiming exhaustive coverage.Its search used targeted keyword searches, citation chaining, and selected policy and institutional documents.Source: §II · pp. 6–8
  • The review may underrepresent English-language and higher-income-region perspectives, miss recent grey literature, and lacks extensive real-world evaluations of proposed controls.The authors note that many governance proposals remain conceptual and their real-world effectiveness evidence is limited.Source: §II · p. 9

d. Spatial Mechanisms as Structural Risk

The material identifies MAUP and narrative distortion as structural spatial-risk scenarios. It also names examples and topics spanning flood routing, land-use classification, and inferred-location policy.

d. Spatial Mechanisms as Structural Risk

e. LLM-Specific Technical Risks

III. Mapping the Governance Landscape of LLM-Enabled GeoAI

Data provenance gaps can carry inaccurate spatial inputs into LLM-enabled recommendations, while location traces can reveal sensitive facts beyond users’ intended disclosures. Proposed privacy approaches have spatial fidelity and auditability trade-offs.

a. Data Provenance and Consent

  • Public geospatial datasets often lack metadata on consent, origin, and intent, making data tracing increasingly problematic as LLMs generalize heterogeneous sources.The review connects weak grounding and inaccurate spatial data with potentially misleading zoning or disaster-response recommendations.Source: a. Data Provenance and Consent · pp. 11–12

Illustrative Example (hypothetical scenario, grounded in documented LLM routing failures; see

  • GPT-4V produced an incorrect hallucinated route in each of approximately 25 OpenStreetMap-image trials, illustrating risks from inaccurate spatial inputs.The hypothetical flood-response system’s outdated map overlays could suggest routes on non-existent roads; Roberts et al. (2023) reported the trial result.Source: Illustrative Example (hypothetical… · p. 12

b. Spatial Privacy and Location Inference

  • GPS traces can reveal home and work locations even when users consent only to fitness or navigation tracking, while vague consent may not disclose downstream use.Garroussi et al. (2025) is cited for the precision with which these locations can be identified.Source: b. Spatial Privacy and Location… · p. 13
  • Synthetic geospatial data can reduce direct privacy exposure, but it must preserve topology and adjacency to avoid misrepresenting rare or sparse spatial patterns.Fidelity verification remains an open challenge, and ungrounded narratives may result when synthetic outputs are fed to LLMs.Source: b. Spatial Privacy and Location… · pp. 14–15
  • Federated learning keeps raw data local, but regional differences in spatial autocorrelation can complicate convergence and wash out local structure.Audit logs and provenance records support verifiability, yet query logs can reconstruct sensitive trajectories; the literature has no settled resolution or spatial adversary evaluation of proposed compromises.Source: b. Spatial Privacy and Location… · p. 15

c. Algorithmic Bias and Spatial Inequity

Spatial bias in GeoAI can reproduce historical inequities, while spatial autocorrelation and scale make these harms geographically clustered and difficult to detect with general audits. The review describes spatially explicit bias auditing as a diagnostic gap, with proposed controls not yet validated.

c. Algorithmic Bias and Spatial Inequity

  • Historical and unrepresentative data, along with selection, exclusion, or detection bias, can reproduce existing disparities in spatial AI systems.The review also describes spatial AI as reflecting dominant social power structures across urban-rural, racial, and economic lines.Source: c. Algorithmic Bias and Spatial Inequity · p. 16
  • Historically biased arrest data can lead spatial risk models to concentrate enforcement in over-policed neighborhoods, reinforcing the original disparity (Gupta et al., 2023).The review identifies predictive policing as its clearest documented example.Source: c. Algorithmic Bias and Spatial Inequity · pp. 16–17
  • Spatial autocorrelation can cluster model errors, so a system may pass a global fairness check while under-serving particular neighborhoods.The review notes that spatially explicit auditing methods accounting for clustered bias are not comparably mature.Source: c. Algorithmic Bias and Spatial Inequity · pp. 16–17
  • The reviewed field has named spatial bias and its roots in historical power structures but has not produced deployed tools to detect or mitigate it in LLM-enabled GeoAI.Spatially disaggregated evaluation is proposed as a candidate control, not a validated practice.Source: c. Algorithmic Bias and Spatial Inequity · pp. 17–18

d. Spatial Mechanisms as Structural Risk

Spatial autocorrelation, aggregation choices, scale mismatch, and multimodal uncertainty make geospatial errors structured rather than independent. The review links these mechanisms to established GIScience diagnostics and governance artifacts, while identifying translation and enforcement as the response gap.

d. Spatial Mechanisms as Structural Risk

  • Spatial autocorrelation can cluster errors and bias in underrepresented areas, producing systemic under-service that overall accuracy may conceal.The review calls for spatially disaggregated audits rather than relying only on a global performance metric.Source: d. Spatial Mechanisms as Structural Risk · pp. 18, 20
  • Results can change with spatial aggregation, and LLM narratives may present zoning-dependent patterns as underlying phenomena or mask small-area disparities.A neighborhood ranking can differ across census-tract, block-group, or grid aggregation; the review calls for declaring the areal unit and summarizing sensitivity across an alternative for high-stakes decisions.Source: d. Spatial Mechanisms as Structural Risk · pp. 18, 20
  • A model trained at one resolution but queried at another can produce confident yet misleading local explanations, obscuring variability relevant to equity decisions.The review calls for disclosing training and evaluation scales and limiting claims not validated at the requested scale.Source: d. Spatial Mechanisms as Structural Risk · pp. 20–21
  • Uncertainty in flood extent, road closures, and map recency can compound in routing outputs, making an unqualified “safe route” a high-stakes governance risk.More broadly, uncertainty can propagate as remote sensing, volunteered geographic information, and administrative records are fused.Source: d. Spatial Mechanisms as Structural Risk · pp. 19, 21
  • GIScience already offers diagnostics for these mechanisms, but the review identifies translation and enforcement—not diagnostic invention—as the response gap.The proposed governance requirements include resolution and recency reporting, aggregation documentation, spatial error or uncertainty surfaces, and sensitivity analysis.Source: d. Spatial Mechanisms as Structural Risk · pp. 19, 21

e. LLM-Specific Technical Risks

LLM-enabled GeoAI adds risks from spatial hallucination, prompt framing, multimodal grounding gaps, and agent tool use. The review finds spatially specific verification and scale-sensitive robustness controls largely absent from the reviewed literature and not field-tested.

e. LLM-Specific Technical Risks

  • LLMs can generate plausible but unsupported spatial claims about roads, adjacency, distances, or features, which non-experts may find difficult to recognize as unreliable.Unlike an obviously wrong date or name, a confidently described route or boundary has no visible marker of unreliability.Source: e. LLM-Specific Technical Risks · p. 23
  • The same spatial data can yield materially different narratives when prompts vary scale, aggregation unit, or comparison boundary, creating reproducibility risk.Prompt framing compounds MAUP-related narrative distortion.Source: e. LLM-Specific Technical Risks · p. 23
  • Fluent multimodal summaries can remain unsupported by or contradict the raster or vector layers they describe (Su et al., 2024; Mohamad et al., 2026).Alignment errors between text, imagery, and spatial layers are not resolved by fluent output.Source: e. LLM-Specific Technical Risks · p. 23
  • Agentic GIS workflows introduce control-plane risks when systems invoke GIS functions unintentionally or without appropriate access.This differs from data-quality or bias concerns: the issue is tool selection and access, not whether a tool returns correct results.Source: e. LLM-Specific Technical Risks · p. 24

Illustrative Example (hypothetical scenario, grounded in documented LLM routing failures; see

  • In the flood-routing example, outdated map overlays and unsupported generated routes can coexist, making tool-call logs important for distinguishing data faults from hallucinations (Roberts et al., 2023).The review says the distinction can be difficult without records of what the model queried and reported.Source: Illustrative Example (hypothetical… · p. 24
  • The review found spatially specific verification methods and evaluation of scale-sensitive prompt robustness largely absent, with proposed controls lacking field-tested operational evaluations.Tool-first grounding and refusal when data is insufficient are described as architectural proposals; the review calls this its most significant open gap.Source: Illustrative Example (hypothetical… · p. 25

f. Explainability and the Black-Box Problem

GeoAI opacity complicates understanding how outputs are derived, and the review identifies a shortage of spatial-specific interpretability tools for LLMs. A documented land-use case illustrates limits of general XAI methods, while current approaches remain borrowed rather than spatial-native.

f. Explainability and the Black-Box Problem

  • GeoAI opacity makes it difficult to understand how outputs are derived, and detailed explanations can overwhelm users or decision-makers.The review also notes insufficient model documentation, feedback loops, and performance metrics for stakeholder review.Source: f. Explainability and the Black-Box… · p. 26
  • In a land-use classification case, general XAI methods offered limited insight, were difficult to select appropriate reference data for, and struggled with geographic scale and map visualization (Xing and Sieber, 2023).The case illustrates that applying general-purpose XAI to a spatial task surfaced difficulties specific to geographic data.Source: f. Explainability and the Black-Box… · p. 27
  • LIME, SHAP, Integrated Gradients, and GNN Explainer have been applied to spatial models, but were not developed with spatial reasoning in mind.Such methods do not by default account for spatial structure, such as dependence on a particular aggregation unit.Source: f. Explainability and the Black-Box… · p. 27
  • The review characterizes current explainability practice as borrowed general XAI tooling rather than spatial-native methods.Wei et al. (2025) identify spatial-specific interpretability tools for LLMs as an open gap.Source: f. Explainability and the Black-Box… · pp. 27–28

g. Policy Issues and Regulatory Gaps in Spatial AI

Existing AI policies address general ethical risks but often lack spatially specific standards, while enforcement and cross-jurisdiction governance remain unresolved. Current responses include emerging policy proposals and California’s data-broker mechanism.

g. Policy Issues and Regulatory Gaps in Spatial AI

  • General frameworks, including Responsible AI principles and the OECD AI Principles, do not provide binding standards tailored to location-specific inferences such as GPS, satellite, or mobility data.Spatial risks include re-identification, population tracking, and policy uses of inferred spatial clustering [Zang & Bolot, 2011].Source: g. Policy Issues and Regulatory Gaps… · pp. 28–29
  • India’s DPDPA and Brazil’s LGPD emphasize identifiers and consent without specific standards for passive spatial inference or synthetic data; U.S. privacy regulation remains largely state-based, with CCPA/CPRA explicitly covering geolocation and derived inferences.The review notes that California amended the CCPA through the CPRA and that federal proposals remained under consideration at the time of review.Source: g. Policy Issues and Regulatory Gaps… · p. 29
  • CCPA/CPRA covers geolocation and inferred profiles, but data brokers’ fragmented ecosystem limited discoverability; California’s Delete Act created registration, disclosure, and centralized deletion requirements.The example frames the challenge as practical enforceability rather than a categorical gap in legal coverage.Source: g. Policy Issues and Regulatory Gaps… · pp. 29–30
  • Centralized rules could improve consistency and enforcement but need international buy-in, while jurisdiction-by-jurisdiction rules are more feasible near-term yet leave regulatory gray zones; neither model is demonstrated at scale for spatial data.Model registries were proposed as shared infrastructure, but remain unimplemented; a Responsible GeoAI workshop also called for traceability and attention to cross-border transfers [Rao et al., 2023].Source: g. Policy Issues and Regulatory Gaps… · pp. 30–31

h. Public Enablement and Workforce Development

Public-facing and workforce literacy are presented as necessary complements to technical and regulatory governance. Existing responses include training and transparency proposals, but LLM-specific GIS workforce standards remain undeveloped.

h. Public Enablement and Workforce Development

  • 79% of public-sector leaders reported user education as a major barrier to trustworthy AI adoption, and limited literacy can leave users unaware of privacy, bias, and spatial-inference risks.The review also links this gap to builders: standard GIS training has not incorporated LLM-specific risks such as hallucination, prompt sensitivity, and tool or agent risk.Source: h. Public Enablement and Workforce… · p. 32
  • A municipal planner may accept an underserved-neighborhood ranking without asking whether it changes under a different aggregation unit, illustrating how limited literacy can obscure MAUP sensitivity.The scenario attributes the gap to the absence of that question in general AI literacy and traditional GIS training, not to hidden information.Source: h. Public Enablement and Workforce… · p. 32
  • Proposed responses include workplace AI-literacy training, public understanding initiatives, localized consent forms, and dashboards showing which data and mechanisms influence decisions.The review reports no empirical GeoAI studies directly measuring the relationship between public understanding and trust or equitable outcomes.Source: h. Public Enablement and Workforce… · p. 33
  • The review calls for tiered public enablement and open-access materials, but reports no GIS curricula incorporating LLM-specific risks and no established spatial AI auditing competency standard.GIS diagnostic skills such as scale-mismatch recognition and auditing spatially clustered error could transfer to spatial AI auditing; the review found no evidence that professional bodies had formalized this training for LLM-enabled GeoAI.Source: h. Public Enablement and Workforce… · pp. 33–34

IV. Toward an Operational Governance Architecture

The section proposes a governance-aware GIS architecture that connects issue-level risks to enforceable controls and auditable artifacts across the spatial data lifecycle. A flood-response routing scenario illustrates how the architecture could host these controls, though the proposed responses remain mostly unvalidated.

IV. Toward an Operational Governance Architecture

  • The contribution links documented risk mechanisms to enforceable controls and verifiable artifacts across the spatial data lifecycle, rather than introducing new ethical principles.The authors contrast this mapping with principle-level frameworks that are spatially agnostic and prior GeoAI work that does not map controls to system layers.Source: §IV · pp. 34–35
  • The proposed architecture layers spatial data, GIS processing, GeoAI/LLM capabilities, and user-facing applications, with governance checkpoints spanning ingestion through decommissioning.Checkpoints include provenance and consent at ingestion, linkage and re-identification controls during storage, audit documentation during modeling, and access controls and review at deployment.Source: §IV · pp. 35–36
  • Controls map common LLM-enabled GeoAI risks to enforcement locations and auditable artifacts, distinguishing candidate controls from established GIScience practices.Candidate controls have not been empirically validated or field-tested in an operational system; established practices operationalize techniques documented or validated in GIScience.Source: §IV · p. 37

architecture (Figure 2)

  • Tool-first GIS generation is theoretically motivated to reduce hallucinations but remains unevaluated operationally, while constraining flexibility and requiring more engineering than free-form generation.The authors recommend tool-first grounding for high-stakes applications and reserve free-form generation for lower-stakes exploratory use.Source: architecture (Figure 2) · p. 41

Illustrative Operational Scenario: LLM-enabled flood response routing agent

  • The flood-response scenario illustrates governance controls and institutional accountability within a high-stakes routing workflow configured as decision support rather than autonomous decision-making.The scenario includes provenance and consent enforcement, audit logging, privacy-risk checks, and spatial-validation checks.Source: Illustrative Operational Scenario:… · pp. 42–43
  • The proposed framework organizes governance principles around a central node and translates them into enforceable practices, including machine-readable consent constraints and auditable explainability and fairness reviews.Examples include query-time enforcement and logging for consent, disclosure of data recency and spatial resolution, and retained artifacts for external privacy and spatial-fairness audits.Source: Illustrative Operational Scenario:… · pp. 45–46

V. Future Research Agenda for Responsible Spatial AI

The future agenda prioritizes empirical testing of spatially specific risks and controls, alongside interpretability, training, and cross-jurisdictional research. It emphasizes that several technical responses remain absent or largely untested in the reviewed literature.

LLM-specific technical risk validation

  • Spatially specific hallucination verification and prompt-robustness testing across spatial scales and aggregation units are priority gaps for future work.The review found no spatially specific verification approaches and no evaluation of prompt sensitivity to spatial scale or aggregation-unit framing.Source: §V · pp. 51–52

Empirical validation of privacy and bias controls

  • Future work should empirically validate privacy preservation and bias mitigation in diverse real-world settings, including audits where errors cluster geographically.Comparative assessment of proposed spatial bias-auditing approaches on genuinely spatial data remains almost entirely absent.Source: §V · p. 52

Interpretability, participation, and training

  • The agenda calls for spatial and multimodal interpretability tools, participatory and longitudinal studies, and evaluated training curricula for professionals and the public.Curricula should integrate LLM-specific risks with existing GIScience training on uncertainty and error propagation.Source: §V · pp. 52–53

Cross-jurisdictional governance

  • Comparative research should examine how technical practices, ethical frameworks, and regulatory environments interact across transnational contexts.The agenda highlights cross-border data flows, portability of accountability tools, and policy harmonization.Source: §V · p. 53

VI. Conclusions

The review maps eight governance issues in LLM-enabled GeoAI and proposes lifecycle controls, while emphasizing that the architecture remains unvalidated. It calls for validation of less mature risks and coordinated work to move governance from principle to verified practice.

VI. Conclusions

  • The review maps eight recurring governance issues and assesses how far responses have progressed, from established GIScience diagnostics to LLM-specific risks that are barely recognized.The issues span data provenance, privacy, bias, spatial mechanisms, technical risks, explainability, policy, and workforce development.Source: §VI · pp. 53–54
  • The proposed architecture links each issue to enforceable controls and auditable artifacts across the geospatial data lifecycle, rather than offering new ethical principles.The flood-routing scenario is illustrative, not deployed, and the architecture and controls are not empirically validated.Source: §VI · p. 54
  • Whether tool-first grounding reduces hallucination, spatially disaggregated audits reveal disparities missed by global metrics, or the governance layer works institutionally remains open.These are untested questions, not settled findings.Source: §VI · p. 54
  • The review prioritizes validation of the least mature issues, particularly spatially specific technical risks not yet addressed by general AI safety research.GIS practitioners, LLM developers, and policymakers are identified as needed collaborators for moving governance from principle to verified practice.Source: §VI · pp. 54–55

https://oecd.ai/en/ai-principles

The supplied section consists of references spanning spatial data quality, privacy, algorithmic bias, GeoAI, and AI governance. The bibliography also includes work on autonomous GIS, generative AI, and geospatial applications.

https://oecd.ai/en/ai-principles

age/excellence-and-trust-artificial-intelligence_en

Proceedings of the 2023 Fifteenth International Conference on Contemporary Computing

  • Listed works address geospatial AI, autonomous GIS, and LLM-based spatiotemporal analysis.The references include an autonomous GIS paper, a research agenda toward autonomous GIS, and UrbanGPT: Spatio-temporal large language models.Source: generative-ai · p. 58

generative-ai

  • The bibliography also covers GeoAI privacy and bias, geospatial health, multimodal models, and synthetic geospatial data.The cited titles include work on privacy-preserving GeoAI, bias in data-driven AI, geospatial health foundation models, multimodal LLM capabilities, and AI-derived geography.Source: generative-ai · pp. 59–60
Loading 2609.16232v1…