In an opinion paper published in Computational Urban Science, researchers Christophe Claramunt, Lars De Sloover, and Nico Van de Weghe warn that generative artificial intelligence systems in urban planning and computational geography systematically mistake linguistic fluency for genuine spatial understanding. Revisiting Drew McDermott’s 1976 critique, Artificial Intelligence Meets Natural Stupidity, the authors demonstrate that early artificial intelligence shortcuts and rhetorical inflation have resurfaced within modern large language models. The paper examines computational urban science as a domain where physical structure, topological consistency, and metric precision are essential. The authors document that language models struggle with spatial reasoning, frequently hallucinating nonexistent geographical features, distorting topological relationships, and failing to maintain metric consistency. By treating surface text generation as domain comprehension, practitioners risk misinforming municipal policy, distorting land-use planning, and introducing severe vulnerabilities into spatial decisions.
McDermott’s 1976 essay targeted artificial intelligence’s tolerance of sloppy thinking and tendency to let rhetorical enthusiasm outpace empirical validation. A central focus was wishful mnemonics, naming program modules, data structures, or relations with cognitive terms such as UNDERSTAND, GOAL, or IS-A before systems earned those descriptors. These labels imported the complex semantic richness of human thought without providing theoretical or algorithmic substance, creating illusions of competence through lexical inflation. To expose this trap, McDermott suggested replacing evocative names with arbitrary programmatic labels, such as G0034, to determine whether underlying operations still justified a cognitive interpretation. The authors show that when contemporary claims attributing spatial reasoning or understanding to generative models are translated into neutral mechanistic terms, their apparent intelligence largely dissipates, demonstrating that perceived competence resides primarily in suggestive phrasing rather than computational architecture.
This rhetorical pattern persists across generative systems like GPT, Claude, Copilot, DeepSeek, and Gemini, which are described as understanding users or reasoning about prompts despite lacking formal computational theories of those terms. The same tendency appears in alignment and safety frameworks that model agentic traits like beliefs, intentions, and goals within next-token prediction systems. Citing arguments by Bender and Mitchell, the authors note that these metaphors anthropomorphize statistical algorithms, creating unjustified expectations while obscuring their mathematical foundation. Rather than clarifying system behavior, such language encourages practitioners to treat token predictors as though they possess human mental faculties. In geographic information science, where language models are increasingly used for code synthesis, conversational geoprocessing, and location-aware inference, conflating grammatical coherence with spatial cognition conceals structural limitations and impedes objective evaluations of operational capabilities and safety risks.
To ground McDermott’s critique empirically, the authors examine failures occurring when large language models handle geographic analysis and cartography. When prompted to generate maps or give directions, models produce fluent outputs containing severe spatial errors. They place landmarks in impossible locations, calculate irrational travel times, and invent non-existent cities. In land-use planning, models have fabricated fictional zoning categories, while in emergency response, hallucinated evacuation routes misrepresent road connectivity and flood extents, endangering populations. These breakdowns highlight the divide between language models and geographic information systems. While geographic software rests on decades of spatial theory, coordinate systems, and topological constraints, language models learn solely from textual patterns. Consequently, they lack foundational geographic knowledge, such as knowing that rivers flow downhill, roads connect continuously, or north corresponds to upward orientation.
The authors explain that spatial hallucinations are structural outcomes of mathematical optimization processes governing large language models. Citing Kalai and colleagues, the paper notes that models are trained to emit the most probable continuation rather than verify factual reality, producing confident yet erroneous statements when certainty is missing. In geospatial contexts, interactions operate on a prompt-first basis, where users supply natural-language descriptions of constraints and models generate responses through probabilistic decoding. This mechanism introduces high prompt sensitivity: minor shifts in phrasing, syntax, or ordering silently alter spatial objectives without user awareness. For example, modifying a prompt requesting proximity can silently pivot an output from straight-line Euclidean distance to road travel time or cost-weighted routing. Even when augmented by retrieval pipelines, the workflow remains a stochastic process whose reliability depends on retrieval coverage and generation faithfulness, areas where standardized evaluation protocols remain unsettled.
Causal reasoning in language models presents further acute challenges. Recent evaluations show that models rely heavily on surface associative cues rather than formal causal identification, exhibiting brittleness in novel contexts. When queried about whether introducing a dedicated bus lane decreased urban travel times, models frequently treat narrative sequence as proof of causation. This vulnerability directly reflects McDermott’s warning against substituting theoretical narrative for engineering verification. Today, that issue re-emerges through demo-driven development and publication by implication, where models are showcased using cherry-picked prompts in controlled settings that hide performance volatility. Conversely, established geographic information software, grounded in decades of engineering, often fails to integrate conceptual advances from geographic information science. This disconnect highlights an epistemological split: while generative systems risk ungrounded rhetorical inflation, conventional geographic platforms risk stagnation by underutilizing spatial theory, leaving both approaches ill-equipped for reliable urban decision-making.
To operationalize McDermott’s critique into a structured methodology, the authors introduce a four-layer analytical framework entitled Rhetoric to Reliability. The first layer addresses discursive and conceptual risks, requiring researchers to establish strict terminological discipline by distinguishing linguistic fluency from spatial competence. The authors emphasize that concepts like navigation and cartographic design embody formal reasoning about topology, metric relationships, orientation, scale, and uncertainty, none of which reduce to textual pattern matching. When an artificial intelligence system generates plausible directions or map descriptions, verbal coherence must not be mistaken for spatial cognition. For example, a model might describe a pedestrian route across a highway or place a public facility in an inaccessible spot while sounding authoritative. To protect scientific integrity, naming conventions must avoid anthropomorphic language and wishful mnemonics, ensuring that advertised capabilities reflect verified empirical performance rather than aspirational vocabulary.
The framework’s second layer focuses on modeling and reasoning issues, examining how language models fail when decoupled from grounded spatial ontologies. Generative systems frequently apply deterministic, binary relations like IS-IN or CONNECTED-TO that fail to capture the probabilistic vagueness inherent in geographic space. In contrast, geographic information science utilizes fuzzy, context-dependent relations such as MAYBE-LOCATED-IN or LIKELY-NEAR, which accurately model real-world ambiguities like informal settlements or culturally defined neighborhoods lacking sharp administrative boundaries. Without topological consistency and metric precision, models assert proximity based merely on textual co-occurrence, ignoring actual transit accessibility or physical barriers. Additionally, models fall into attentional learning traps where early feedback reinforces oversimplified spatial assumptions. For instance, an urban mobility model might assume a location two kilometers away is more accessible than one five kilometers away, ignoring terrain constraints like steep hills or socio-cultural divisions like gated communities and segregation.
The third layer articulates methodological safeguards, mandating that partial model failures be documented systematically as opportunities for diagnostic learning and pipeline refinement rather than concealed. As Rich and Gureckis established, machine learning systems underweight rare events and construct illusory correlations when trained on sparse or unevenly sampled datasets, a condition prevalent across geographic space. To counteract these distortions, geospatial pipelines must implement coverage-aware monitoring to identify where data scarcity skews inferences. Methodologically, the authors advocate for hybrid architectures that integrate large language models with geographic information systems and formal geospatial ontologies. In these systems, language models translate prompts into code while delegating spatial computation to validated geographic components and external resources like OpenStreetMap. By embedding automated topological and validity checks into model execution paths, hybrid pipelines prevent recurrent errors, constrain generative outputs within geometric rules, and ground inferences within verified spatial representations.
The fourth layer governs reporting and evaluation, addressing severe reproducibility challenges confronting artificial intelligence in geospatial research. The authors observe that while open-source software is celebrated for transparency, its practical reuse across geographic domains is frequently obstructed by complex system dependencies, rapid model release cycles, and diverse geographical contexts. Many implementations lack standardized evaluation baselines, explicit versioning, or benchmark datasets, making independent replication across environments difficult. To ensure responsible deployment, the framework calls for transparent governance protocols, including full provenance tracking of input prompts, retrieved documents, and tool calls. Furthermore, developers must publish coverage dashboards and structured failure logs delineating where model performance drops across regional, rural, or informal environments. Without explicit reporting of performance disparities, deploying ungrounded models risks amplifying spatial inequalities, generating inaccurate planning outputs, or fostering misplaced confidence in resource-poor areas that lack digital representation.
Ultimately, Claramunt, De Sloover, and Van de Weghe argue that technical safeguards must be guided by epistemic humility, defined as recognizing the boundaries of computational knowledge and exercising restraint when communicating capabilities. Citing Gellers, the paper highlights that epistemic humility counters the temptation to treat polished language as cognitive depth, an essential corrective for ethical artificial intelligence deployment. The authors note that the risks of mistaking linguistic fluency for domain competence extend beyond urban planning into critical fields like medical diagnostics, climate modeling, and economic forecasting, where ungrounded assertions carry severe societal consequences. Rather than dismissing generative artificial intelligence, the authors advocate for hybrid systems combining language models, geographic information frameworks, and human oversight. Grounding artificial intelligence in McDermott’s epistemic discipline ensures that computational tools respect real-world complexity, transforming historical methodological critique into a modern design principle and reminding researchers that self-criticism remains vital for building trustworthy scientific technologies.
Subject of Research: Enduring relevance of Drew McDermott’s critique for large language models in computational urban science
Article Title: Computational and urban science amid generative AI: the enduring relevance of McDermott’s critique (opinion paper)
Article References: Claramunt, C., De Sloover, L., & Van de Weghe, N. (2026). Computational and urban science amid generative AI: the enduring relevance of McDermott’s critique (opinion paper). Computational Urban Science, 6(1), Article 29. https://doi.org/10.1007/s43762-026-00264-7
Image Credits: AI Generated
DOI: 10.1007/s43762-026-00264-7
Keywords: computational urban science, generative AI, large language models, spatial reasoning, Drew McDermott, wishful mnemonics, geographic information systems, cartographic hallucination, topological consistency, urban geography, prompt sensitivity, spatial ontologies, epistemic discipline, epistemic humility, causal reasoning, open-source GeoAI, hybrid spatial pipelines
News Source: Denise Maddox. (November 12, 2026). Urban Science Opinion Warns Generative AI Models Lack True Spatial Understanding. Scienmag.
