PaperlayerView exact revision on arXiv

Source-linked AI summary

From Community Values to AI Design Constraints: A Mixed-Methods Study of a Proposed AI-Driven Wildfire Risk Assessment Tool in Los Angeles County

Sanaz Sadat Hosseini, Mona Azarbayjani, Mohammad Pourhomayoun, Hamed Tabkhi

arXiv:2610.03724v1cs.HCcs.CY

TL;DR

Residents may lack clear property-specific wildfire information, and the reviewed research leaves unclear how early community evidence can translate into design requirements. This study elicited responses to a conceptual wildfire-risk tool and mapped them to provisional requirements; participants valued property-level information but also wanted understandable explanations, transparent data practices, and feasible recommendations.

  • Problem

    The literature leaves unclear how early community evidence can be documented as specific design requirements for resident-facing AI hazard tools before model and interface choices are fixed.

    Source: §2.5 · p. 4

  • Method

    The study examined resident responses to a proposed tool and translated participant findings into provisional design requirements through the AI Value Map.

    Source: §7.1 · p. 19, §9 · p. 20

  • Results

    Participants valued property-level risk information and mitigation guidance, while also wanting clear explanations, transparent data practices, feasible recommendations, and attention to neighborhood conditions.

    Source: §9 · p. 20

  • Takeaways & Limitations

    The AI Value Map connects community input to early design decisions and offers provisional requirements to guide later design and evaluation.

    Source: §9 · p. 20

    The convenience sample of 30 survey respondents and seven town hall participants supports neither neighborhood-specific comparisons nor conclusions about how common these perspectives are county-wide.

    Source: §8 · p. 19

Abstract

from arXiv · show

AI-driven hazard tools are increasingly proposed for resident-facing risk communication. However, residents are often asked for feedback only after decisions about data use, privacy, and explanation have been made. This study takes an earlier approach by examining residents' values, concerns, and practical expectations for a proposed AI-based risk assessment tool before a predictive model or functional interface is developed. The proposed smartphone application would use property photographs to estimate parcel-level wildfire risk and provide an interpretable risk score, uncertainty information, and mitigation recommendations. Using a convergent mixed-methods design, we surveyed 30 Los Angeles County residents, seven of whom also participated in a virtual town hall. We analyzed the data using descriptive statistics, exploratory FDR-adjusted Spearman correlations, and inductive thematic coding of open-ended and town hall responses. Participants were cautiously receptive: 66.7% said they would be very likely or likely to use the tool. They linked fair risk assessment to whether the tool considered relevant property and neighborhood conditions. Privacy and data security were the most common concerns (64.3%), while cost was the main barrier to acting on recommendations (65.5%). Likelihood of using the tool was strongly associated with likelihood of saving or sharing risk results ($\rho = 0.752$, $p_{\mathrm{FDR}} < .001$). We translated the findings into the AI Value Map, a community-grounded artifact with nine value dimensions. The Map connects participant findings to provisional design requirements and possible consequences if they are not addressed. This study provides exploratory evidence on residents' responses to a proposed AI wildfire tool and offers a pre-model procedure for translating early community input into traceable design constraints for resident-facing AI hazard tools.

1 Introduction and Problem Context

Nearby properties can experience sharply different wildfire damage, while existing public tools often lack locally calibrated, resident-facing guidance at the property scale. This study examines residents’ expectations for a proposed photo-based tool and translates their input into provisional design constraints.

1 Introduction and Problem Context

  • Wildfire can spread between properties through connected landscaping, exterior materials, and structure spacing, making neighborhood awareness and coordinated mitigation relevant to resilience.The passage notes that nearby homes can have contrasting damage outcomes and points to defensible space, vegetation, building materials, and parcel layout as related conditions.Source: §1 · p. 2
  • Existing public tools may not provide locally calibrated, neighborhood- or home-scale assessments or resident-facing prevention guidance, leaving residents without clear property-specific information.The proposed tool would use property photographs alongside other data to provide parcel-level risk scores, uncertainty information, and mitigation recommendations; no predictive model had yet been developed.Source: §1 · p. 2
  • AI risk assessment, XAI, ethical AI, and participatory design offer resources, but provide limited guidance for translating residents’ values into constraints before technical and interface choices are fixed.The unresolved design questions include data use, risk and uncertainty explanations, privacy, fairness, resident interaction, and recommendation feasibility.Source: §1 · p. 2
  • The study combines a survey and virtual town hall to examine residents’ expectations and organize findings into the AI Value Map’s nine value dimensions.The Map connects participant evidence to provisional requirements and possible consequences if they are not addressed; the authors present it as a pre-model procedure that may be adapted to other hazard settings.Source: §1 · p. 3

2 Background and Research Gap

Prior research addresses wildfire communication and vulnerability, technical risk prediction and interpretability, and participatory design. The integrated gap is how to document a path from early community evidence to concrete design requirements for a hazard tool.

2.1 Wildfire risk communication and the equity dimension

  • Wildfire communication research links protective action to environmental and social cues, perceived threat, trust, ability to act, and local conditions; vulnerability and adaptive capacity also vary across communities.Clear, specific, timely information connected to local conditions is emphasized, while cited studies report greater vulnerability in some racial and ethnic communities and gaps in emergency response and recovery.Source: §2.1 · pp. 3–4

2.2 Wildfire risk prediction advances and the resident comprehension gap

  • Wildfire prediction advances expand risk-mapping capacity, but technical performance alone does not establish whether resident-facing information will be understandable or useful.The reviewed literature offers limited empirical guidance on presenting wildfire risk to residents or shaping explanations and design choices before model development.Source: §2.2 · p. 4

2.3 Explainable AI and the limits of technical interpretability for residents

  • XAI methods support examining model behavior and predictions, but technical interpretability alone does not determine how residents’ risk information should be explained.The literature calls for explanations designed around users’ questions and contexts and notes the lack of a case-specific traceability procedure from early resident evidence to hazard-tool requirements.Source: §2.3 · p. 4

2.4 Participatory and value-sensitive design for AI tools

  • Value Sensitive Design and participatory design offer ways to consider values and involve intended users, while participatory AI warns that consultation can be symbolic without influence over design decisions.The study therefore focuses on whether residents are involved early enough to shape choices about data, explanations, privacy, fairness, and recommendations.Source: §2.4 · p. 4

2.5 The integrated gap

  • The integrated gap is how to document the path from early community evidence to specific design requirements before key model and interface decisions become unchangeable.Wildfire communication, AI/XAI, and participatory and value-sensitive design each offer relevant foundations, but the reviewed literature leaves this translation process less clear.Source: §2.5 · p. 4

3 Study Framing and Case Setting

The study examines community expectations while a proposed wildfire tool’s design choices remain open, grounding that work in a specific application and Los Angeles County context. Participants reviewed a conceptual smartphone workflow rather than a functioning model.

3 Study Framing and Case Setting

  • The study commits to eliciting residents’ expectations before core design choices are fixed, grounding input in a specific tool and context, and making links from input to decisions explicit.The framing treats a structured artifact, rather than informal researcher interpretation, as the means of documenting those links.Source: §3 · p. 5

3.1 Pre-model engagement as a methodological commitment

  • Engaging residents before a model or prototype exists leaves concerns able to inform provisional design requirements, whereas later engagement may have less influence on underlying choices.During data collection the tool remained wireframes and conceptual flows, and participants responded to simulated examples rather than a functioning AI system.Source: §3.1 · p. 5

3.2 Proposed AI-driven wildfire risk assessment tool

  • The proposed smartphone app would use exterior property images to provide parcel-level feedback for residents’ everyday mitigation decisions.Its conceptual workflow showed a simulated score, uncertainty range, explanatory factors, recommendations, and a save-or-share report; participants responded to illustrative outputs, not model-generated results.Source: §3.2 · p. 5, Figure 2 · p. 6

3.3 Case selection: Los Angeles County, CA

  • Los Angeles County was selected for recurring wildfire exposure and social and housing diversity; recruitment was county-wide, with no comparison among individual communities.The county-wide scope highlights the need to accommodate different housing situations, property types, and levels of authority to make property changes.Source: §3.3 · pp. 5–6

4 Research Methods

The study combines survey and town hall evidence about the same proposed tool, using qualitative discussion to help interpret survey responses. Recruitment and sample size support exploratory, within-sample analysis rather than population-level claims.

4.1 Research design

  • A convergent mixed-methods design combined survey responses from 30 Los Angeles County residents with town hall discussion from seven participants, then brought the separately analyzed data together during interpretation.Both data sources used a shared engagement protocol and core tool concept and visual materials.Source: §4.1 · pp. 6–7

4.2 Recruitment and participants

  • Targeted convenience recruitment sought adult county residents, including homeowners and renters, and encouraged participation from Altadena, Pasadena, and Pacific Palisades.Three of 30 respondents reported living in Pasadena; none reported living in Altadena or Pacific Palisades, and no neighborhood-level comparisons were conducted.Source: §4.2 · p. 7

4.3 Sample size justification

  • The sample was chosen for exploratory mixed-methods analysis; the survey supports descriptive summaries and exploratory associations, not population-level inference.The seven-person town hall and open-ended responses were used to identify design-relevant concerns, not to claim representativeness or thematic saturation.Source: §4.3 · p. 7

4.4 Data collection and town hall procedure

  • A 90-minute virtual town hall presented wildfire context, the proposed workflow, scenario cards, and wireframes, with survey sections between presentation and discussion.The fixed agenda included facilitated discussions and online survey segments; participants raised experiences, reactions, missing risk factors, and expectations, and the town hall was recorded and transcribed with consent. The survey link later allowed 23 respondents to review recorded materials before responding.Source: §4.4 · p. 7

4.5 Survey instrument

The study used a purpose-built survey with varied question formats and Likert items for attitudinal analyses; its 30-person sample was mostly working-age homeowners with near-universal smartphone access.

4.5 Survey instrument

  • The survey was designed for non-expert participants and exploratory analysis, with item wording informed by survey-design guidance and reviewed before IRB submission.Source: §4.5 · p. 7
  • The 31-question survey covered five sections and used binary, nominal, ordinal, Likert, multi-select, and open-ended formats.Five-point Likert items supported descriptive summaries and exploratory association analyses.Source: §4.5 · p. 8
  • The town hall agenda was planned for 90 minutes, with segment durations adjusted in real time to accommodate participant engagement.Source: Table 1 · p. 8, §4.5 · p. 8

4.6 Sample demographics

  • The sample included 46.7% aged 30–44, 63.3% homeowners, and 56.7% single-family occupants; 96.7% had smartphone access.Although 80.0% reported property insurance, 30.0% could not confirm wildfire coverage.Source: §4.6 · p. 8

4.7 Data preparation

Survey records were screened and retained without imputing missing values; item responses were coded for analysis, while open-ended material was prepared separately for thematic analysis.

4.7 Data preparation

  • All 30 respondents provided consent and were retained after screening; missing values were kept as missing rather than imputed.Q11 and Q12 each had one missing response; Q22 had two blank responses, while Q5a was answered by five eligible respondents.Source: §4.7 · p. 8
  • Multi-select percentages used each item's respondents as the denominator, and categorical, Likert, and selected multi-select responses were coded according to their measurement types.Likert items remained ordinal, and selected multi-select options became binary indicator variables.Source: §4.7 · pp. 8–9
  • Open-ended answers to Q25 and Q31 and transcripts from four town hall discussion segments were prepared for thematic analysis.Source: §4.7 · p. 9
  • The final survey items included sharing recipients, app-use likelihood, community-safety sharing willingness, and factors encouraging regular use.Variable names followed conventions used in later analyses.Source: §4.7 · p. 9

4.8 Quantitative analysis

The quantitative analysis combined descriptive summaries with exploratory Spearman correlations across ordinal survey items, applying false-discovery-rate adjustment and interpreting associations cautiously within the sample.

4.8 Quantitative analysis

  • Descriptive statistics summarized participant characteristics, response frequencies, and Likert items; ordinal Likert items were summarized with means and standard deviations.Source: §4.8 · p. 10
  • The analysis evaluated all 66 pairwise combinations among 12 ordinal survey items using Spearman correlation, with pairwise sample sizes from n=28 to n=30.Missing responses were removed separately for each item pair.Source: §4.8 · p. 10
  • 21 associations remained significant at pFDR < .05, and 10 were reported because they were directly relevant to integrated interpretation.The study treated these findings as exploratory within-sample patterns, not population-level evidence.Source: §4.8 · p. 10
  • Benjamini–Hochberg adjustment was applied across the 66 tested associations, and the small, non-probability sample constrained interpretation to exploratory patterns.Source: §4.8 · p. 10

4.9 Qualitative analysis

Inductive thematic analysis of survey comments and town hall transcripts was compared with quantitative findings, then integrated to develop the AI Value Map while retaining divergent views.

4.9 Qualitative analysis

  • Inductive coding covered Q25 responses (n=22), Q31 responses (n=24), and four town hall discussion segments, with codes grouped into candidate themes and checked against excerpts.Thematic boundaries were revised, and comments that did not fit emerging themes were retained.Source: §4.9 · pp. 10–11

4.10 Mixed-methods integration

  • Survey and town hall themes were compared for convergence and divergence, and integrated with quantitative findings to develop the nine AI Value Map dimensions.Differences were retained to represent the range and complexity of residents' views.Source: §4.10 · p. 11

4.11 Validity, reliability, and ethical considerations

  • The team supported trustworthiness by documenting core study materials and checking themes and quotations against original comments and transcripts.The same core materials were used across the town hall and post-event survey.Source: §4.11 · p. 11, §4.9 · p. 11
  • Survey responses were anonymous, and incentive contact information was stored separately so it could not be linked to responses.Participants consented electronically or verbally before recorded discussion segments.Source: §4.11 · p. 11

5 Integrated Findings

Across the integrated findings, residents weighed the usefulness of parcel-level information against questions about how the assessment works, whether it is fair, and whether recommendations are feasible. Adoption interest was present, but participants’ responses point to needs for understandable results, locally relevant assessment, privacy protections, and practical support; exploratory associations describe this sample only.

5.1 Wildfire risk perception and property-level information gap

  • Participants were unsure about property-level wildfire conditions and often relied on broad information sources, while town hall participants described wanting parcel-level detail.One-third did not know which prevention steps had been taken at their property; over half reported using no wildfire-specific apps, websites, or alert systems.Source: §5.1 · pp. 11–12

5.2 Comfort, usefulness, and willingness to act on visual assessment

  • Risk scores and recommendations were rated useful, but comfort with visual assessment was lower and varied more; participants wanted explanations tied to visible property conditions.Greater comfort and perceived usefulness were associated with higher intended app use, but these exploratory associations do not establish causation.Source: §5.2 · p. 12

5.3 Risk-score explanation preferences and feedback on conceptual screens

  • Participants rated the conceptual layout highly and most preferred color bands, yet open-ended feedback asked for clearer score meaning, photo guidance, visible reasoning, and community context.Color bands were selected by 53.3%, numeric scores by 26.7%, and descriptive labels by 20.0%.Source: §5.3 · pp. 12–13, Figure 3 · p. 13

5.4 Fairness, privacy, and equity concerns

  • 64.3% selected privacy or data security as a concern, while participants linked fairness to whether relevant local conditions—including those not visible in photographs—were considered.Participants also called for transparent, user-controlled rules on sharing information with third parties; 76.7% viewed the proposed concept as somewhat or very fair, before a working tool existed.Source: §5.4 · p. 13

5.5 Adoption intent and community sharing

  • 66.7% said they were likely or very likely to use the app; likelihood of use was associated with saving or sharing results (Q27 ↔ Q29, ρ = 0.752, pFDR < .001).Participants also tended to connect intended use with acting on guidance and sharing results for community safety.Source: §5.5 · pp. 13–14

5.6 Practical barriers and conditions for continued use

  • Cost was the most common barrier to acting on app insights (65.5%), and practical constraints often overlapped with time or property restrictions.Participants also sought current local updates, tracking of risk and improvements over time, and connections between photographs, scores, recommendations, feasibility, and neighborhood context.Source: §5.6 · pp. 14–15

5.7 Summary of findings

  • Residents’ interest depended on more than receiving a score: trust required understandable results, relevant factors, and clarity about data use, while action depended on residents’ resources and authority.The findings point to supporting people with different levels of wildfire knowledge and control over property decisions.Source: §5.7 · p. 15

6 The AI Value Map

The AI Value Map translates participant findings into provisional design requirements and possible consequences if they are not addressed. Its nine dimensions capture the proposed wildfire tool’s design tensions, while the mapping structure may be adapted to other settings.

6 The AI Value Map

  • The AI Value Map links participant findings to provisional design requirements and possible consequences, covering data, fairness, explanation, privacy, and residents’ ability to act.The requirements have not been tested in a functioning system and are intended to guide later development and evaluation.Source: §6 · p. 16

6.1 How the AI Value Map was derived

  • The nine dimensions were developed by identifying and grouping quantitative and qualitative patterns, summarizing findings, translating them into requirements, and pairing requirements with possible risks.The Map organizes each dimension across participant findings, design requirements, and what may happen if requirements are unaddressed.Source: §6.1 · p. 16, Figure 6 · p. 16
  • Privacy and data security concerned 64.3% of Q22 respondents; the privacy-boundaries example translates this concern and an insurance-related comment into a pre-upload policy and controls to view or delete stored information.The paper notes residents may be less willing to upload images or return without safeguards.Source: §6.1 · p. 16
  • At the concept stage, 76.7% rated the tool somewhat or very fair across neighborhoods, linking fairness to whether shared conditions such as nearby vegetation and vacant lots were considered.The Map proposes explaining relevant property and neighborhood conditions and evaluating the tool across varied settings.Source: §6.1 · p. 18

6.2 Key design tensions and considerations

  • Insurance appeared as both a possible reason to use the tool and a privacy concern; the proposed approach keeps it disconnected from insurers by default and lets residents control sharing.Some participants valued records for insurance-related needs, while others worried the information could affect coverage or be used against them.Source: §6.2 · p. 18, §6.1 · p. 18
  • Despite modest perceived risk, participants prioritized household preparedness; the paper proposes readiness-focused framing, while treating the nine dimensions as local and the mapping structure as transferable.Different tools should develop dimensions and requirements based on their own users, setting, and purpose.Source: §6.2 · p. 18

7 Discussion and Implications

The discussion presents the AI Value Map as a pre-model way to make the path from community evidence to design choices visible. It connects that approach to implications for explanation, feasible action, fairness, and data governance, while noting that the requirements remain provisional.

7.1 The AI Value Map as a pre-model design method

  • The Map records how participant evidence becomes a provisional design constraint before model and interface development, making visible what may happen if the constraint is not addressed.Its role is to make community input traceable to proposed system changes, rather than merely documenting it.Source: §7.1 · p. 19
  • The requirements remain provisional because participants responded to a conceptual tool and no model-generated results were tested.The paper positions the Map as a starting point for later co-design, model development, and prototype evaluation.Source: §7.1 · p. 19

7.2 Implications for explanation, actionability, fairness, and governance

  • Residents wanted explanations of what the score means, which visible conditions contribute, and what action to take; divided format preferences suggest connected presentation forms rather than one fixed format.The paper also identifies plain-language interpretation, photo highlights, uncertainty information, and links to recommended actions as needs.Source: §7.2 · p. 19
  • Whether residents can act on recommendations depends on cost, time, housing type, restrictions, physical ability, and decision authority, so later testing should include varied housing and financial circumstances.The paper describes the tool’s potential usefulness as connecting property conditions to steps feasible for a resident’s circumstances.Source: §7.2 · p. 19
  • Participants tied fairness to considering shared conditions such as wind, neighboring parcels, and vacant lots; clear data policies and limits on insurer access may also affect willingness to submit information.The discussion treats explanation, feasibility, fairness, and data governance as connected parts of the resident experience.Source: §7.2 · p. 19

8 Future Directions

Future work will test and refine the proposed requirements through community engagement and prototype evaluation, while addressing sample, procedure, and deployment-governance boundaries.

8 Future Directions

  • The county-wide convenience sample of 30 survey respondents and seven town hall participants cannot support neighborhood-specific comparisons or estimates of how common perspectives are across Los Angeles County.Of 66 tested associations, 21 remained significant after FDR adjustment; these exploratory within-sample patterns warrant examination in a larger, more representative sample.Source: §8 · p. 19
  • Engagement in Altadena, Pasadena, and Pacific Palisades, including residents affected by recent fires, can refine the Value Map against community-specific priorities and constraints.A broader, more diverse sample can test whether values and requirements hold across differing wildfire exposure, housing, socioeconomic contexts, and mitigation capacities.Source: §8 · p. 20
  • Different survey formats may have influenced responses: seven town hall participants completed the survey live, while 23 others responded after recorded materials and the presentation.Future studies can use standardized engagement protocols and larger samples to examine differences across participant groups and communities more systematically.Source: §8 · p. 20
  • Prototype testing should assess realistic resident interactions and whether explanations, privacy controls, and feasibility-oriented recommendations improve understanding, trust, and willingness to act.The current findings reflect anticipated needs from wireframes and conceptual materials, not actual usability or adoption.Source: §8 · p. 20
  • Before larger-scale deployment, future work should define data-sharing boundaries, institutional partnerships, and governance mechanisms that preserve resident control over uploaded images and stored assessments.The proposed tool may interact with insurers, fire agencies, planners, community organizations, and other third parties.Source: §8 · p. 20

9 Conclusion

Residents responding to a conceptual wildfire-risk tool were cautiously receptive but raised concerns about explanations, privacy, and practical ability to act. The study's AI Value Map turns their input into provisional design requirements, not evidence that a future tool will be usable, accurate, fair, or widely adopted.

9 Conclusion

  • Residents valued property-level risk information and mitigation guidance but wanted clear explanations, transparent data practices, feasible recommendations, and attention to neighborhood conditions.Risk-score presentation preferences varied; privacy and data security were common concerns, and cost, time, housing situation, or decision authority could limit action.Source: §9 · p. 20
  • The AI Value Map connects participant findings to provisional decisions about assessment outputs, resident interactions, and practical conditions affecting whether recommendations can be carried out.It gives community input a defined role in early design.Source: §9 · p. 20
  • The findings do not establish that a future tool will be usable, accurate, fair, or widely adopted; further community engagement, prototype testing, and governance work remain necessary.Future work should engage residents in wildfire-affected communities and establish data-sharing boundaries and institutional partnerships.Source: §9 · p. 20

Related papers

Loading 2610.03724v1…