Pith. sign in

REVIEW 4 major objections 6 minor 61 references

A visual AI model, prompted with heritage checklists and street-view photos, can produce heritage scores that Swedish authorities will use in EU renovation planning.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Zero-shot GPT-4o scoring of street-view façades can flag likely heritage buildings at national scale, but the threshold choices and validation are too weak to support use without expert oversight.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection A candid, well-scoped case study of zero-shot VLM façade scoring for national heritage screening, with real governance uptake — but the 5.0 Mm² addition has no below-threshold validation, so recall and bias remain unmeasured. the 4 major comments →

arxiv 2601.06056 v2 pith:VCCM6H3Q submitted 2025-12-22 cs.CY cs.AIcs.CV

Using street view images and visual LLMs to predict heritage values for governance support: Risks, ethics, and policy implications

Tim Johansson , Mikael Mangold , Kristina Dabrock , Anna Donarelli , Ingrid Campo-Ruiz This is my paper
classification cs.CY cs.AIcs.CV
keywords heritage valuezero-shot promptingvisual large language modelsstreet view imagerybuilding stockEnergy Performance of Buildings Directivepolicy decision support
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that zero-shot predictions from a multimodal LLM — a single street-view façade image plus a prompt built from official heritage assessment checklists — are useful enough for real governance decisions about the building stock. The test case is Sweden's implementation of the EU Energy Performance of Buildings Directive, where no national heritage register exists. Out of 154,710 energy performance certificates, the model produced a 1–100 heritage-value score; the authorities took the top slice (score ≥ 50, about 1–2% of the stock, 5.0 million m² of floor area) and added it to the heritage categories feeding the national renovation plan. The paper does not claim the scores are unbiased — it documents that the model gives higher scores in Stockholm and in affluent areas, and lower scores to wooden buildings — but argues that with parameterized output, threshold calibration, and expert oversight, the method is a viable decision-support tool. That matters because most EU member states lack heritage registers, and this workflow can be reused wherever street-view imagery exists.

Core claim

The paper's central claim, stated explicitly in the conclusion, is that zero-shot predictions by multimodal LLMs can have value for understanding the building stock and in decision-making, specifically in Sweden's EPBD process. The discovery on the paper's own terms is that a checklist-built prompt — the authorities' own ocular-survey checklist turned into a structured JSON template — makes the LLM's numeric heritage score trustworthy enough at the top of the ranking that heritage experts accepted a cutoff of 50 out of 100, incorporating roughly 5.0 million square metres of previously unregistered floor area into the national renovation plan. The supporting evidence is cross-tabulation: offi

What carries the argument

The carrying mechanism is the zero-shot prompt together with a threshold rule. The prompt assigns the model the persona of a heritage expert and demands a JSON object with detail fields (construction year, style, roof shape, facade material, ornaments, etc.) plus a 1–100 'predicted heritage value', instructing it to return N/A when image evidence is insufficient. The load-bearing decision step is converting this ordinal score into a binary heritage category by taking everything ≥50; that cutoff was calibrated so that the added floor area is only about 1–2% of the stock, and the paper shows the model avoids extreme values, so the cutoff is sensitive to perturbation. Supporting machinery is th

Load-bearing premise

The method assumes that heritage value shows up in what is visible in a single street-view photograph of the façade, so buildings whose heritage value lives in interiors, social history, or hidden sides are systematically missed.

What would settle it

Take a random sample of, say, 200 buildings with predicted heritage values below 50 that are not already in official registers, have heritage professionals conduct on-site ocular surveys using the same checklists, and count how many meet the checklist criteria; if that share is roughly as high as in the above-50 group, the threshold adds no discriminative information and the 5.0 million m² designation is arbitrary.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Other EU member states that must produce national building renovation plans can replicate the workflow using their own street-view coverage and a checklist prompt, producing a first-pass heritage layer where registers are missing.
  • Because the threshold converts an ordinal score into a category, small shifts in the cutoff change inclusion rates substantially, so authorities must recalibrate the cutoff against local registers rather than reusing 50.
  • The documented biases toward Stockholm, affluent areas, and stone or plaster facades imply that blind deployment would skew heritage protection; the paper's counter is a deliberately small top slice plus expert oversight.
  • Parameterized, non-prose output makes errors auditable against registers and reduces sycophancy, offering a pattern for trustworthy LLM use in government beyond heritage assessment.
  • Since the EPBD requires updated renovation plans every two years, the method makes heritage-data updates cheap enough to repeat at that cadence.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural extension would be inversion: use the same prompt to search for high-scoring buildings in underrepresented materials such as wood, and feed those to local surveyors, counteracting the model's conservatism instead of just trimming the top.
  • Because inclusion is threshold-sensitive and the model's score distribution shifts between versions, a governance deployment should pin the model version and re-establish the cutoff against a fixed validation set before each refresh; otherwise the 5.0 million m² figure is not reproducible over time.
  • The reusable asset is not the heritage label but the calibrated score-and-threshold workflow; the same pipeline could be aimed at renovation state, material reuse, or hazardous-substance screening, as the paper hints.
  • For the Global South applications the paper mentions, the absence of energy performance certificates means the spatial-linkage step needs a substitute; OpenStreetMap footprints combined with the same street-view prompt is a testable alternative.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper reports a national-scale application of zero-shot GPT-4o prompts to Google Street View façade images in Sweden, covering 154,710 Energy Performance Certificates. The model outputs a 1–100 predicted heritage value together with building features. A threshold of 50, chosen so that the added group would be roughly 1–2% of the stock, was used by the Swedish National Heritage Board to add about 5.0 million m² of heated floor area to the heritage categories feeding the Swedish Building Renovation Plan under the EPBD. The paper presents cross-checks against EPC registers and protected-building datasets, analyzes income and façade-material biases, and discusses governance risks such as sycophancy, automation bias, and legal/ethical concerns. The central conclusion is that zero-shot multimodal LLM predictions 'can have value' for building-stock governance, specifically in the Swedish EPBD implementation.

Significance. If the central claim is accepted, the paper describes a scalable workflow for a real policy need: identifying potential heritage buildings in a national stock where no complete register exists. The paper is unusually transparent about limitations, publishes the full prompt in Appendix 1, provides a GitHub module, and grounds feature selection in the authorities' own heritage checklists. The bias analysis in Tables 6–8 is a useful contribution, and the discussion of sycophancy and human oversight is relevant for AI-assisted governance. However, the empirical support for the load-bearing 'can have value' claim is currently thin: the only expert validation is ten positive cases, the threshold is set by an inclusion-rate target rather than a recall target, and no random-sample check of below-threshold buildings is reported. Because the paper operationalizes 'value' as adding the 5.0 Mm² group to a governance dataset, the missing false-negative analysis is not a peripheral concern.

major comments (4)
  1. [§3.4, §3.2] The load-bearing validation is incomplete. §3.4 reports that ten buildings with predicted heritage value above 50 were inspected and confirmed; §3.2 explicitly acknowledges that 'there remains a risk that potential visible features representing heritage values in buildings with lower scores were overlooked.' A random or stratified sample of below-threshold buildings, including known protected buildings, is needed to estimate recall and false-negative bias. Without it, the conclusion that the workflow 'can have value' for EPBD governance is not established: a screening tool that misses systematic categories may create a biased national inventory. A recall estimate or a clear reframing of the output as a deliberately conservative, incomplete supplement is required.
  2. [§3.3, Tables 7 and 8] The bias results directly undermine the completeness of the 5.0 Mm² addition. Table 7 shows that the share of buildings scoring above 50 rises from 3.2% in the lowest-income group to 36–41% in the highest-income groups; Table 8 shows wooden façades receive above-threshold scores far less often than plaster or stone (3.2% vs 8.6% and 55.6%). Since wooden and low-income-area buildings are also categories the paper says are typical of Räkna Q 'everyday heritage,' the false-negative problem is likely socially patterned. This is not a generic caveat: it means the specific group added to the EPBD dataset is probably unrepresentative in ways that matter for distributional fairness. The paper should either provide a quantitative false-negative analysis for these subgroups or explicitly state that the tool is unsuited to identify heritage value in these categories.
  3. [§3.2, Appendix 1] The construct-validity assumption is untested. The prompt asks only for image-visible attributes (façade material, roof, windows, ornaments, and similar features), and the paper limits its remit to 'visible features that could be indicative of heritage values' (§1). Heritage value in Swedish legislation also includes social and cultural associations, and interiors or façades obscured by vegetation are invisible by construction. The paper claims the model can assist in assigning heritage values for EPBD, but the only evidence that visible façade character is a sufficient proxy is the aggregate correlations for Räkna Q and Byggnadsminnen in Table 5. A concrete test would be to examine the below-threshold buildings that are nevertheless protected in Räkna Q and characterize why they are missed. Without such a test, the 5.0 Mm² group is defined by a visible-feature proxy whose relationship t
  4. [Table 5, §3.4] The threshold choice is fragile and partly circular. The paper itself notes that LLMs avoid extreme values and that 'minor perturbations around the cutoff may lead to substantial shifts in inclusion rates.' The threshold of 50 was chosen to achieve a 1–2% inclusion target, and the same threshold defines the 'predicted heritage value' group whose validity is then argued from ten positive cases. A sensitivity analysis around the cutoff (e.g., 40, 45, 55, 60) showing the resulting floor-area additions and the expert-check pass rates would establish whether the operational result is robust or an artifact of the chosen threshold. This is necessary for the paper's claim that the specific 5.0 Mm² figure is a meaningful governance output.
minor comments (6)
  1. [Table 3 caption] The caption says 'Construction period' but the table compares number of floors. Please correct.
  2. [§3.3, text before Table 6] The sentence 'Räkna Q buildings are outside of Stockholm' is unclear; the table compares in- and outside Stockholm, and Räkna Q is used as a separate benchmark. Rephrase.
  3. [Tables 7 and 8] Typos: 'if buildings' should be 'of buildings' in the heading text preceding Table 8; also make the income-group definitions and the comparison base for percentages explicit in the table notes.
  4. [Appendix 1] The prompt is described as having color-coded segments (Blue/Yellow/Green), but the colors are not visible in a monochrome print version. Consider labeling segments with text markers instead of relying on color.
  5. [Abstract] Minor wording: 'there is a lack of a national register' should probably be 'there is no comprehensive national register', consistent with the Introduction.
  6. [References] Reference 37 is cited as 'arXiv preprint arXiv:2508.11486v3 (2026)'; verify the year and version, and clarify its status relative to the technical validation claims.

Circularity Check

0 steps flagged

No significant circularity: LLM heritage scores are external model outputs, the threshold is a policy cut-off, and the paper explicitly acknowledges its validation limits.

full rationale

The paper's central workflow applies a zero-shot GPT-4o prompt to street-view images and uses the resulting 1-100 heritage score to define a 'predicted heritage value' category above 50. The threshold is not fitted to the outcome or derived from the LLM scores in a way that makes the predictions tautological; it is an explicitly policy-driven cut-off chosen to keep the added floor area at 1-2% of the stock ('The threshold was intentionally chosen to only include an additional 1-2% of the building stock'). The predicted scores themselves vary independently of this cut-off, and the paper reports multiple patterns (wooden facades, lower-income areas, Räkna Q buildings) that show the score is not simply the threshold restated. The ten-building expert check is a positive-case check on above-threshold buildings, which is weak evidence but not circular because it does not determine the scores. The paper also explicitly concedes the key limitation: 'there remains a risk that potential visible features representing heritage values in buildings with lower scores were overlooked.' The technical validation in ref. 37 is a separately reported companion with overlapping authorship, but the governance-focused claim here does not reduce to that citation; it is additionally supported by internal consistency checks, transparent bias analyses, and the authority's decision process. No equation or fitted parameter equates inputs with the reported 5.0 Mm2 output, and no self-citation chain is used to forbid alternative methods. Under the requested standard, this is a non-circular application paper with acknowledged external-validity limitations rather than a circular derivation.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical or formal entities. Its uncharged inputs are hand-set operational thresholds (score cutoffs, sightline parameters, floor-height constant) and domain assumptions about what the images and registers actually represent. The most consequential of these is the assumption that façade visibility tracks heritage value, which is also the most fragile.

free parameters (4)
  • Predicted heritage value threshold = 50
    Chosen in §3.4 to add only an additional 1–2% of the building stock to the heritage group; the paper itself notes that small perturbations around the cutoff substantially change inclusion rates.
  • Visibility score cutoff = 50
    Buildings with a model-assigned visibility score below 50 were excluded (§2.3), reducing 186,038 observed buildings to 154,710 EPCs; the cutoff affects which buildings enter all subsequent analyses.
  • Sightline distance and angle limits = ≤50 m; <3° from perpendicular
    Adjusted from 100 m to 50 m and perpendicular deviation limited to 3° (§2.1) to improve image quality; this changes coverage and can bias the sample toward road-facing façades.
  • Floor-height multiplier = 3.0 m/story
    Used in §2.2 to estimate camera pitch from EPC-reported number of floors; a different multiplier would change the framing of the façade and hence the image input to the LLM.
axioms (4)
  • domain assumption Heritage value can be meaningfully screened from visible façade features alone.
    §1 states the assessments were limited to 'visible features that could be indicative of heritage values', and Appendix 1 asks only about image-visible attributes. If this is false, the entire screening approach is invalid.
  • domain assumption The official heritage registers (Byggnadsminne, Räkna Q) are valid benchmarks for heritage value.
    Used as ground truth in §3.2 despite the paper noting the registers are incomplete and coverage varies. A benchmark that misses most heritage buildings cannot fully support claims about the model's validity.
  • domain assumption GPT-4o's 1–100 heritage score is a comparable and stable measure across buildings and prompt invocations.
    The thresholding procedure in §3.4 treats the scalar score as comparable across 281,018 images; no repeated prompting, temperature/seed information, or inter-run variance analysis is reported.
  • domain assumption Google Street View coverage and the image filtering process yield a representative subset of the Swedish EPC stock.
    Only 72% of buildings had a valid camera position, and visibility filtering further reduces the sample (§2.1, Table 1); the paper acknowledges coverage bias but still draws stock-level conclusions.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Using street view images and visual LLMs to predict heritage values for governance support: Risks, ethics, and policy implications." pith.science (2026). https://pith.science/paper/VCCM6H3Q

@misc{pith2026260106056,
  author       = {Pith},
  title        = {Pith review of: Using street view images and visual LLMs to predict heritage values for governance support: Risks, ethics, and policy implications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VCCM6H3Q}},
  note         = {Machine review of arXiv:2601.06056}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

During 2025 and 2026, the Energy Performance of Buildings Directive is being implemented in the European Union member states, requiring all member states to have National Building Renovation Plans. In Sweden, there is no comprehensive national register of buildings with heritage values. This is seen as a barrier for the analyses underlying the development of Building Renovation Plans by the involved Swedish authorities. The purpose of this research was to assist Swedish authorities in developing information on heritage values in the Swedish building stock. Buildings in street view images from all over Sweden (N=154 710) have been analysed using multimodal Large Language Models (LLM) to assess visible aspects indicative of heritage value. Zero-shot predictions by LLMs were used as a basis for identifying buildings with potential heritage values for 5.0 million square meters of heated floor area. In this paper, the results of the predictions and lessons learned are presented and related to the development of the Swedish Building Renovation Plan as part of governance. The problems with the method and potential improvements are discussed. Risks with authorities use of LLM-based data are addressed, with a focus on issues of transparency, error detection and sycophancy.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

61 extracted references · 4 linked inside Pith

  1. [1]

    & Low, J

    Cornago, S., Ramakrishna, S. & Low, J. S. C. How can Transformers and large language models like ChatGPT help LCA practitioners? Resour. Conserv. Recycl 196, 107062 (2023)

    2023

  2. [2]

    Zou, X. et al. Deep learning for cross-domain data fusion in urban computing: Taxonomy, advances, and outlook. Information Fusion 113, 102606 (2025)

    2025

  3. [3]

    & Talen, E

    Salazar-Miranda, A. & Talen, E. An AI-based analysis of zoning reforms in US cities. Nature Cities 1–12 (2025)

    2025

  4. [4]

    Wang, S. et al. GPT, large language models (LLMs) and generative artificial intelligence (GAI) models in geospatial science: a systematic review. International Journal of Digital Earth 17, 2353122 (2024)

    2024

  5. [5]

    Ramalingam, S. P. & Kumar, V. Building usage prediction in complex urban scenes by fusing text and facade features from street view images using deep learning. Building and Environment 267, 112174 (2025)

    2025

  6. [6]

    Hou, S. et al. Chain-of-programming (CoP): empowering large language models for geospatial code generation task. International Journal of Digital Earth 18, 2509812 (2025)

    2025

  7. [7]

    Yan, Y. et al. Towards vertical urban geometry extraction: occlusion-reduced estimation from street view images using diffusion models. International Journal of Digital Earth 18, 2520475 (2025)

    2025

  8. [8]

    & Biljecki, F

    Fan, Z., Feng, C.-C. & Biljecki, F. Coverage and bias of street view imagery in mapping the urban environment. Computers, Environment and Urban Systems 117, 102253 (2025)

    2025

  9. [9]

    Umar, F. et al. On the potential of Google Street View for environmental waste quantification in urban Africa: An assessment of bias in spatial coverage. Sustainable Environment 9, 2251799 (2023)

    2023

  10. [10]

    X., Chen, S., Zhang, F., Shi, Y

    Zhu, X. X., Chen, S., Zhang, F., Shi, Y. & Wang, Y. GlobalBuildingAtlas: an open global and complete dataset of building polygons, heights and LoD1 3D models. Earth System Science Data Discussions 2025, 1–31 (2025)

    2025

  11. [11]

    J., Yigitcanlar, T

    Fu, X., Li, C., Quan, S. J., Yigitcanlar, T. & Wasserman, D. Large language models in urban planning. Nature Cities 1–8 (2025)

    2025

  12. [12]

    Lian, X. et al. Tree shortages in informal settlements across African cities. Nature Cities 2, 1049–1059 (2025)

    2025

  13. [13]

    Plattform Kulturhistorisk värdering och urval: grundläggande förhållningssätt för arbete med att definiera, värdera, prioritera och utveckla kulturarvet

    Riksantikvarieämbetet. Plattform Kulturhistorisk värdering och urval: grundläggande förhållningssätt för arbete med att definiera, värdera, prioritera och utveckla kulturarvet. (2015). 23

    2015

  14. [14]

    Preservation, Conservation and Heritage: Approaches to the Past in the Present through the Built Environment

    Ashworth, G. Preservation, Conservation and Heritage: Approaches to the Past in the Present through the Built Environment. Asian Anthropology 10, 1–18 (2011)

    2011

  15. [15]

    Balancing Building Conservation with Energy Conservation - Towards differentiated energy renovation strategies in historic building stocks

    Eriksson, P. Balancing Building Conservation with Energy Conservation - Towards differentiated energy renovation strategies in historic building stocks. (GU, 2021)

    2021

  16. [16]

    & Johansson, T

    Eriksson, P. & Johansson, T. Towards Differentiated Energy Renovation Strategies for Heritage -Designated Multifamily Building Stocks. Heritage 4, 4318–4334 (2021)

    2021

  17. [17]

    von Platten, J. et al. Using Machine Learning to Enrich Building Databases —Methods for Tailored Energy Retrofits. Energies 13, 2574 (2020)

    2020

  18. [18]

    & Mangold, M

    Johansson, T., Olofsson, T. & Mangold, M. Development of an energy atlas for renovation of the multifamily building stock in Sweden. Applied Energy 203, 723–736 (2017)

    2017

  19. [19]

    & Mjörnell, K

    von Platten, J., Holmberg, C., Mangold, M., Johansson, T. & Mjörnell, K. The renewing of Energy Performance Certificates—Reaching comparability between decade-apart energy records. Applied Energy 255, 113902 (2019)

    2019

  20. [20]

    & Mjörnell, K

    von Platten, J., de Fine Licht, K., Mangold, M. & Mjörnell, K. Renovating on Unequal Premises: A Normative Framework for a Just Renovation Wave in Swedish Multifamily Housing. Energies 14, 6054 (2021)

    2021

  21. [21]

    & Johansson, T

    Mangold, M. & Johansson, T. Aspekter Av Kulturvärden: Från Datorseende Och Gatubilder . (Riksantikvarieämbetet, 2025)

    2025

  22. [22]

    Zhang, J., Xiang, R., Kuang, Z., Wang, B. & Li, Y. ArchGPT: harnessing large language models for supporting renovation and conservation of traditional architectural heritage. Heritage Science 12, 220 (2024)

    2024

  23. [23]

    Zeng, Z. et al. Zero-Shot Building Age Classification from Facade Image Using GPT -4. arXiv preprint arXiv:2404.09921 (2024)

    Pith/arXiv arXiv 2024

  24. [24]

    & Zhao, B

    Lin, Y. & Zhao, B. Posthuman cartography? Rethinking artificial intelligence, cartographic practices, and reflexivity. Annals of the American Association of Geographers 115, 499–512 (2025)

    2025

  25. [25]

    Attention economics, artificial intelligence, and the future of the planning profession

    Metzger, J. Attention economics, artificial intelligence, and the future of the planning profession. Planning Theory 14730952251360224 (2025)

    2025

  26. [26]

    Compositional Intelligence: Architectural Typology Through Generative AI

    Koehler, D. Compositional Intelligence: Architectural Typology Through Generative AI. (Taylor & Francis, 2025). 24

    2025

  27. [27]

    Milojevic-Dupont, N. et al. EUBUCCO v0. 1: European building stock characteristics in a common and open database for 200+ million individual buildings. Scientific data 10, 147 (2023)

    2023

  28. [28]

    Jang, K. M. & Kim, J. Multimodal large language models as built environment auditing tools. The Professional Geographer 77, 84–90 (2025)

    2025

  29. [29]

    J., Milojevic-Dupont, N., Creutzig, F., Repke, T

    Hintz, M. J., Milojevic-Dupont, N., Creutzig, F., Repke, T. & Kaack, L. H. A systematic map of machine learning for urban climate change mitigation. Nature Cities 1–13 (2025)

    2025

  30. [30]

    & Wang, D

    Zhao, Z. & Wang, D. Evaluation of large language models for the intangible cultural heritage domain. npj Heritage Science 13, 439 (2025)

    2025

  31. [31]

    & Chen, Y

    Zhou, Z., Xi, Y., Xing, S. & Chen, Y. Cultural bias mitigation in vision -language models for digital heritage documentation: A comparative analysis of debiasing techniques. Artificial Intelligence and Machine Learning Review 5, 28–40 (2024)

    2024

  32. [32]

    & Biljecki, F

    Liang, X., Xie, J., Zhao, T., Stouffs, R. & Biljecki, F. OpenFACADES: An Open Framework for Architectural Caption and Attribute Data Enrichment via Street View Imagery. arXiv preprint arXiv:2504.02866 (2025)

    arXiv 2025

  33. [33]

    Sycophancy in large language models: Causes and mitigations

    Malmqvist, L. Sycophancy in large language models: Causes and mitigations. in 61–74 (Springer, 2025)

    2025

  34. [34]

    AI-Kommissionens Färdplan För Sverige

    AI-kommissionen. AI-Kommissionens Färdplan För Sverige . https://www.regeringen.se/contentassets/11431c781ff746b3b7a49312fa4de624/ai-kommissionens-fardplan- for-sverige.pdf (2024)

    2024

  35. [35]

    Artificial intelligence may affect diversity: architecture and cultural context reflected through ChatGPT, Midjourney, and Google Maps

    Campo-Ruiz, I. Artificial intelligence may affect diversity: architecture and cultural context reflected through ChatGPT, Midjourney, and Google Maps. Humanities and Social Sciences Communications 12, 1–13 (2025)

    2025

  36. [36]

    Helbich, M., Danish, M., Labib, S. M. & Ricker, B. To use or not to use proprietary street view images in (health and place) research? That is the question. Health & Place 87, 103244 (2024)

    2024

  37. [37]

    Dabrock, K. et al. Automated Building Heritage Assessment Using Street -Level Imagery. arXiv preprint arXiv:2508.11486v3 (2026)

    arXiv 2026

  38. [38]

    & Willard, J

    Liu, Q., Wang, W. & Willard, J. Effects of prompt length on domain -specific tasks for large language models. arXiv preprint arXiv:2502.14255 (2025). 25

    Pith/arXiv arXiv 2025

  39. [39]

    White, J. et al. A prompt pattern catalog to enhance prompt engineering with chatgpt. arXiv preprint arXiv:2302.11382 (2023)

    Pith/arXiv arXiv 2023

  40. [40]

    Prompt engineering for ChatGPT: a quick guide to techniques, tips, and best practices

    Ekin, S. Prompt engineering for ChatGPT: a quick guide to techniques, tips, and best practices. Authorea Preprints (2023)

    2023

  41. [41]

    Checklista karaktärsdrag

    Boverket. Checklista karaktärsdrag. (2023)

    2023

  42. [42]

    Så Byggdes Husen 1880-2000 : Arkitektur, Konstruktion Och Material i Våra Flerbostadshus under 120 År

    Cecilia Björk, Per Kallstenius, & Laila Reppen. Så Byggdes Husen 1880-2000 : Arkitektur, Konstruktion Och Material i Våra Flerbostadshus under 120 År. (2008)

    2000

  43. [43]

    & Bansal, M

    Pothiraj, A., Stengel -Eskin, E., Cho, J. & Bansal, M. Capture: Evaluating spatial reasoning in vision language models via occluded object counting. arXiv preprint arXiv:2504.15485 (2025)

    Pith/arXiv arXiv 2025

  44. [44]

    Räkna q i Södermanland - Kulturhistoriskt Värdefull Bebyggelse Skyddad Enligt Plan- Och Bygglagen

    Gillbrand, E. Räkna q i Södermanland - Kulturhistoriskt Värdefull Bebyggelse Skyddad Enligt Plan- Och Bygglagen. (2015)

    2015

  45. [45]

    & Palm, P

    Andersson, M., Kopsch, F. & Palm, P. How cultural values are reflected on the housing market–direct effects and the cultural spillover. International Journal of Housing Markets and Analysis 12, 405–423 (2019)

    2019

  46. [46]

    & Lessani, M

    Ning, H., Li, Z., Akinboyewa, T. & Lessani, M. N. An autonomous GIS agent framework for geospatial data retrieval. International Journal of Digital Earth 18, 2458688 (2025)

    2025

  47. [47]

    W., Corrêa, N

    Pasetti, M., Santos, J. W., Corrêa, N. K., de Oliveira, N. & Barbosa, C. P. Technical, legal, and ethical challenges of generative artificial intelligence: an analysis of the governance of training data and copyrights. Discover Artificial Intelligence 5, 193 (2025)

    2025

  48. [48]

    Quintana, M. et al. Global urban visual perception varies across demographics and personalities. Nature Cities 1– 15 (2025)

    2025

  49. [49]

    & Myers, D

    Avrami, E., Macdonald, S., Mason, R. & Myers, D. Values in heritage management: Emerging approaches and research directions. (2019)

    2019

  50. [50]

    G., Arrese, M

    Agudo, U., Liberal, K. G., Arrese, M. & Matute, H. The impact of AI errors in a human -in-the-loop process. Cognitive Research: Principles and Implications 9, 1 (2024)

    2024

  51. [51]

    Article 14: Human Oversight

    EU IA Act. Article 14: Human Oversight. (2026)

    2026

  52. [52]

    AI Playbook for the UK Government

    UK Government Digital Services. AI Playbook for the UK Government . https://www.gov.uk/government/publications/ai-playbook-for-the-uk-government (2025). 26

    2025

  53. [53]

    The Evolving Role of Human-in-the-Loop Evaluations in Advanced AI Systems, European Journal of Computer Science and Information Technology

    Kapoor, M. The Evolving Role of Human-in-the-Loop Evaluations in Advanced AI Systems, European Journal of Computer Science and Information Technology. European Journal of Computer Science and Information Technology 13 (9), (2025)

    2025

  54. [54]

    & You, S

    Jiang, P., Sonne, C., Li, W., You, F. & You, S. Preventing the immense increase in the life -cycle energy and carbon footprints of LLM-powered intelligent chatbots. Engineering 40, 202–210 (2024)

    2024

  55. [55]

    & Detyniecki, M

    Ruf, B. & Detyniecki, M. Mitigating the Carbon Footprint. Trends in Sustainable Computing and Machine Intelligence: Proceedings of ICTSM 2024 15 (2025)

    2024

  56. [56]

    & Jia, P

    Dai, S., Li, Y., Stein, A., Yang, S. & Jia, P. Street view imagery -based built environment auditing tools: a systematic review. International Journal of Geographical Information Science 38, 1136–1157 (2024)

    2024

  57. [57]

    Uppdrag Att Främja En Cirkulär Ekonomi i Byggoch Fastighetssektorn

    Boverket. Uppdrag Att Främja En Cirkulär Ekonomi i Byggoch Fastighetssektorn. (2024)

    2024

  58. [58]

    Data-driven Approaches for Predicting Hazardous Substances in the Building Stock

    Wu, P.-Y. Data-driven Approaches for Predicting Hazardous Substances in the Building Stock. (LTH, 2024)

    2024

  59. [59]

    EU Building Stock Observatory

    European Climate, Infrastructure and Environment Executive Agency. EU Building Stock Observatory. (2025)

    2025

  60. [60]

    Dabrock, K., Pflugradt, N., Weinand, J. M. & Stolten, D. Leveraging machine learning to generate a unified and complete building height dataset for Germany. Energy and AI 17, 100408 (2024). 27 Appendix 1 The prompt used As a cultural heritage value expert, assess the building in the image located at {{Adress}}. Valid values are provided in square brackets...

    2024

  61. [61]

    It completely fills the image both verti cally and horizontally. 4. No obstructions such as vehicles, vegetation, or scaffolding are covering significant parts of it. 5. Only one building is shown. [1-100]} Appendix 1 The prompt has the following segments: Blue: Context, information, task; Yellow: Format, permitted values; Green: Details 28 Appendix 2 Aux...

    1900

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.