The study, authored by Sangho Lee of the Graduate School of Logistics at Incheon National University, formalizes what it calls the IFEZ AI City Framework, a four-pillar architecture designed to carry the zone beyond conventional smart city practice. The timing is not accidental. In December 2025, IFEZ formally declared its AI CITY INITIATIVEs at the IFEZ Smart City Summit, and the announcement was deliberately aligned with the January 2026 enforcement of South Korea’s Framework Act on Artificial Intelligence, widely known as the AI Basic Act. The study treats that coincidence of infrastructure maturity and new national law as a defining institutional moment, one that could make IFEZ a testbed for city-scale AI deployment and validation.

The intellectual core of the paper is a distinction that urban scholars have been sharpening for several years: the difference between a smart city and an AI city. Smart cities, as they evolved through the 2000s and 2010s, were fundamentally about connectivity and efficiency, using IoT sensor networks, cloud computing, and centralized data platforms to optimize existing processes such as traffic management, energy distribution, and public safety. AI cities, by contrast, are conceived as systems that learn continuously, predict future conditions, and derive solutions autonomously. The paper draws on recent theoretical work on urban artificial intelligence and AI urbanism, including concepts such as City AI and Agentic Urban AI, in which AI systems move from recommending actions to reasoning, planning, and executing multi-step decisions on their own.

What makes IFEZ a plausible candidate for this transition is the sheer density of its existing digital nervous system. The zone manages approximately 9,000 on-site urban facilities, CCTV units, traffic devices, and emergency response terminals, all feeding a centralized Integrated Operations Center supported by high-performance GPUs. That infrastructure generates continuous, high-volume operational data suitable for AI model training and real-time inference. Yet the study is candid about the gaps: its SWOT analysis identifies fragmented data infrastructure and insufficient GPU and high-performance computing capacity for large-scale model training as current constraints, meaning that further integration and computing expansion are prerequisites rather than afterthoughts.

The first pillar of the framework is the AI-based urban model architecture, the technical backbone of the proposed AI city. Its centerpiece is a proposed AI City Brain platform, a centralized AI engine that would integrate urban data, simulation models, and decision processes through an ontology-based architecture, acting as a semantic layer that connects heterogeneous data streams from traffic sensors, environmental monitors, and administrative databases. This sits alongside a proposed unified data lakehouse combining the flexibility of data lakes with the performance of data warehouses, and digital twin models that would dynamically represent physical urban assets and allow scenario simulation before any real-world policy or operational change is made. Predictive and AI-assisted operations, from traffic flow and energy consumption forecasting to predictive maintenance, would run on top of this foundation, with selected functions progressing toward higher automation only where technically reliable, legally permissible, and subject to defined human oversight.

The second and third pillars are economic rather than infrastructural, and they reveal the framework’s most distinctive bet: that an AI city must be an AI economy. Pillar two, the AI-bio industry enablement system, builds on Songdo’s biopharmaceutical cluster, home to Samsung Biologics and Celltrion, and proposes an AI bio data hub integrating biomedical, clinical, and genomic data for precision medicine and drug discovery, federated data sharing that preserves data sovereignty, AI-driven laboratory automation, and interdisciplinary talent programs drawing on the five international universities at the Incheon Global Campus, including SUNY Stony Brook, George Mason, Ghent University, the University of Utah, and FIT. Pillar three proposes an AI startup ecosystem built around a real-city sandbox, an expanded Incheon Startup Park, and a Startup-City Co-Growth Model in which startup solutions that pass technical, legal, ethical, and operational validation could be progressively integrated into city services.

The fourth pillar translates all of this into citizen-facing services: AI-assisted analysis of predefined safety events such as crowd congestion, falls, and fire indicators; multilingual AI chatbots for administrative services suited to IFEZ’s internationally diverse population; and an AI city operation center that would integrate all IoT and sensor data streams to automate maintenance scheduling and monitor KPIs aligned with ESG and UN Sustainable Development Goal indicators. Crucially, the framework insists that AI-generated alerts in public safety be treated as decision-support inputs rather than as evidence of unlawful or dangerous intent, with consequential actions remaining subject to human review.

Implementation is structured as a three-phase roadmap running from 2026 to 2030. Phase one, in 2026, focuses on what the study calls data assetization, the deliberate transformation of raw smart city data into AI-ready assets that are preprocessed, labeled, and standardized so that AI companies can use them immediately; its flagship initiative proposes de-identification and event labeling of CCTV footage to create safety-oriented training datasets, subject to privacy and data-protection compliance. Phase two, spanning 2027 and 2028, would expand the AI City Sandbox across the three districts and deploy edge AI infrastructure to process analytics closer to camera and sensor endpoints, reducing latency and privacy exposure. Phase three, from 2029 to 2030, aims to institutionalize validated AI services within the Integrated Operations Center and to standardize and internationally disseminate a K-AI City model, though the study is careful to frame export as a future objective rather than an accomplished fact.

Governance receives as much attention as technology, reflecting the constraints of the AI Basic Act, which defines High-Impact AI systems that materially affect citizens’ lives, safety, and fundamental rights and imposes obligations for prior notification, risk management, and algorithmic explainability. The framework’s Responsible AI by Design approach recommends explainable AI for consequential administrative decisions, human-in-the-loop oversight, de-identification, Zero-Trust security principles, and a sovereign AI infrastructure strategy to reduce dependence on external platforms for sensitive public-sector applications. It also confronts the darker risks of AI-enabled surveillance head-on, including algorithmic bias, false positives, function creep, and disproportionate monitoring, and prescribes purpose limitation, data minimization, periodic bias audits, and public channels for review and redress.

The study is equally explicit about its own limits. The framework is conceptual, derived from qualitative synthesis of academic literature, policy documents, and IFEZ materials rather than from empirical validation, and the author acknowledges that his affiliation with the IFEZ administration may introduce institutional bias. Many components, including the AI City Brain platform and the full-scale sandbox, are proposed or planned rather than operational. Yet the paper’s transferable design principles, that data assetization should precede AI deployment, that vertical sector-specific AI and horizontal urban AI should develop together, that regulation can function as an innovation-enabling capability rather than merely a compliance burden, and that modularity and interoperability should be embedded from the outset, offer a genuinely replicable architecture. If the roadmap holds, the world’s first district-scale AI city may be less a technological moonshot than a disciplined exercise in institutional design.

Subject of Research: A strategic framework for transitioning the Incheon Free Economic Zone from a smart city to an AI city model

Article Title: A strategic framework for transitioning from smart city to AI city in the Incheon free economic zone (IFEZ)

Article References: Lee, S. (2026). A strategic framework for transitioning from smart city to AI city in the Incheon free economic zone (IFEZ). Discover Cities, 3(1), Article 198. https://doi.org/10.1007/s44327-026-00379-w

Image Credits: AI Generated

DOI: 10.1007/s44327-026-00379-w

Keywords: AI city, smart city, IFEZ, urban artificial intelligence, digital twin, AI governance, free economic zone, Songdo, AI Basic Act, data assetization, agentic AI, urban policy

The technical heart of the problem lies in LiDAR, the laser-scanning technique that fires millions of light pulses at façades, roofs and streets to capture the geometry of the built environment. Airborne or mobile LiDAR campaigns rarely see everything. Occlusions from trees, neighbouring buildings and blind angles leave point clouds riddled with holes, so researchers have turned to artificial intelligence models—geometry-aware transformers such as PoinTr and AdaPoinTr, and completion architectures like SnowflakeNet and PointAttN—to fill in the missing geometry. These models hallucinate plausible surfaces where the laser never reached, producing complete building shapes from partial captures. The trouble is that a completed point cloud is not a certified measurement. If an algorithm invents the far side of a roof, how good is that invention, and who decides when it is good enough to base renovation priorities on?

That question is where the Košice team makes its sharpest intervention. In the point-cloud completion literature, quality is typically reported as an F-score computed after normalising the data, and it is tempting to treat a threshold on that score as a physical accuracy guarantee. The authors show, through a careful unit-chain analysis, that this temptation is a trap. A nominal acceptance threshold of two centimetres, read off a normalized completion F-score, in fact certifies a median physical tolerance of 7.6 centimetres—and that tolerance drifts with the size of the building. In other words, a number that looks like a length is not a length at all, and certifying building data on that basis would be an error silently baked into policy. Instead, the team re-anchors acceptance tiers to physical point-to-mesh deviation bands drawn from scan-to-BIM accuracy specifications, including the Level of Accuracy guidance used by the building-documentation industry, so that every quality class corresponds to a real, size-independent geometric tolerance that a surveyor or regulator can understand.

With the acceptance rule fixed, the remaining challenge was delivery. Field deployment means commodity mobile hardware, intermittent connectivity and campaign budgets that cannot absorb cloud-computing costs or round trips to a distant server for every query. The researchers therefore built a mobile-edge dashboard that carries the entire decision-support apparatus to the browser. They demonstrated it on 9,998 AI-completed buildings drawn from the TU Delft Building-PCC 50k benchmark, a large-scale dataset of building point clouds designed for completion research. The primary view of the dashboard ships in a 251 KB payload—comfortably under the 300 KB target in the study’s title—which includes enough pre-computed statistical structure that the heavy lifting of quality triage happens locally, on the device, without any server-side computation.

The interaction model is deliberately simple and deliberately fast. Live sliders let an analyst move the physical accuracy bounds—the point-to-mesh deviation bands—up and down, and the dashboard recomputes the distribution of buildings across acceptance tiers in linear time, using plain JavaScript with no framework dependencies. The same sliders project the campaign-level consequences: because better-verified digital models mean fewer repeat site visits, the tool estimates the survey-travel carbon savings, amounting to 35.6 tonnes of CO₂-equivalent in the study’s mid bracket for the demonstrated building set. That figure is not a hypothetical for the whole EU stock; it is a measured projection for the benchmark buildings under way, illustrating how the choice of an acceptance threshold propagates directly into diesel burned and carbon emitted by inspection vehicles.

Performance numbers are reported with unusual rigour for a dashboard paper. Under throttled network emulation corresponding to a 1.6 Mbps 3G link, the cold-cache time-to-interactive was measured at 3.0 seconds; on 4G it dropped to 1.0 second. Those figures matter because the entire premise of the system is that it works where infrastructure is worst—on a rural survey route, in a basement archive, at the edge of a network cell. The engineering contract that makes this possible is pinned by schema-lock unit tests, which guarantee that the pre-computed data shipped from the edge always matches what the browser code expects, so a silent schema change cannot corrupt the triage logic in the field.

Governance receives as much attention as speed. The system is designed to be auditable against the OWASP Top 10:2021 catalogue of critical web application security risks and against WCAG 2.2, the current Web Content Accessibility Guidelines, so that security reviewers and accessibility auditors both have a defined checklist rather than a black box. It follows GDPR data minimisation by design: because the analytics run client-side on pre-computed summaries, no building-level personal or location data needs to leave the device, and there is no back-end to breach. The architecture also embodies the Industry 5.0 principle of human oversight, positioning the dashboard not as an automated gatekeeper but as a transparent instrument that keeps a human expert in command of every acceptance decision. The team frames this within a broader line of human-centric digitalisation work, arguing that sustainability triage of buildings is precisely the kind of consequential, norm-laden judgement that should not be fully delegated to a model.

The broader significance of the work goes beyond one dashboard. Cities worldwide are investing in digital twins—living digital replicas of urban infrastructure—yet the field has repeatedly stumbled on the question of data quality assurance: what does it mean for a machine-completed model to be trustworthy enough for energy simulation, renovation subsidy allocation or regulatory compliance? By tying AI completion quality to normative, physically meaningful tolerance bands, the Košice study offers a template for converting benchmark scores into regulatory language. It also demonstrates a deployment philosophy that runs counter to the current fashion for ever-heavier client applications: a quarter-megabyte of carefully pre-computed data and dependency-free JavaScript can outperform a bloated framework app where connectivity and hardware budgets are thinnest.

For the 2030 deadline itself, the implications are practical. Energy agencies do not need to wait for nationwide 5G coverage or for every municipality to procure server infrastructure; a decision tool that runs on a mid-range phone over 3G can be distributed to surveyors almost as easily as a PDF. The triage framing is equally consequential: rather than demanding perfect digital models of every building, the system sorts the stock into tolerance tiers, allowing renovation programmes to prioritise buildings whose digital representations are confidently reliable while flagging those that need re-surveying or better completion models. That kind of tiered honesty about uncertainty is rare in smart-city marketing and essential in energy policy, where a misclassified building can mean misdirected subsidies and unmet carbon targets.

The study, partially funded through Slovak research programmes including KEGA and APVV projects on digital technologies and circular construction, arrives at a moment when the gap between AI capability and regulatory usability has become the central bottleneck of urban digitalisation. Point-cloud completion transformers will keep improving, and benchmarks like Building-PCC will keep expanding. What the Košice team adds is the connective tissue: a unit-sound acceptance rule, a browser-native decision tool that fits in 251 KB, measured performance on the worst networks a surveyor is likely to meet, and an audit trail aligned with security, accessibility and privacy norms. It is a reminder that the path from a laboratory benchmark to a policy instrument is paved with unglamorous engineering—and that sometimes the most futuristic thing a smart city can deploy is a very small, very fast, very honest web page.

Subject of Research: A mobile-edge smart-city dashboard that certifies AI-completed LiDAR building point clouds against physical accuracy tolerances for EU building-stock sustainability triage.

Article Title: A Mobile-edge Smart-city Dashboard for AI-completed LiDAR Building Stocks: Normative-anchored Sustainability Triage Under a 300 KB Payload

Article References: Behúnová, A., Vavrák, V., Pohorenec, M., & Behún, M. (2026). A Mobile-edge Smart-city Dashboard for AI-completed LiDAR Building Stocks: Normative-anchored Sustainability Triage Under a 300 KB Payload. Mobile Networks and Applications. https://doi.org/10.1007/s11036-026-02534-4

Image Credits: AI Generated

DOI: 10.1007/s11036-026-02534-4

Keywords: smart city, mobile-edge computing, LiDAR, point cloud completion, building energy performance, digital twins, scan-to-BIM accuracy, Industry 5.0, GDPR data minimisation, WCAG 2.2, sustainability triage, TU Delft Building-PCC benchmark