The research, led by Associate Professor Mohammad Javad Koohsari of the Urban Design Science for Health Laboratory at the Japan Advanced Institute of Science and Technology (JAIST) together with Professor Koichiro Oka of the Faculty of Sport Sciences at Waseda University, examined the ethical properties of advice generated by ChatGPT for modifying built environments to support human health. The study was made available online on August 10, 2026, and will be published in Volume 27 of the journal Developments in the Built Environment on October 1, 2026. According to the researchers, it is the first investigation anywhere in the world to subject LLM-generated advisory text on urban design and health to a structured ethical audit.

The methodological design was deliberately rigorous. The team framed the evaluation around six well-established pathways through which the built environment shapes health: physical activity, dietary intake, social interaction, air pollution, traffic safety and crime, and noise. They then constructed 18 distinct prompts. Twelve of these described scenarios in higher-income and lower-income neighborhoods, while six described mixed-income neighborhoods operating under an explicit budget constraint, a condition that forces trade-offs and therefore exposes how a model distributes scarce resources. Each prompt was run ten times, yielding 180 separate responses for analysis.

Every response was then scored against four ethical criteria drawn from established frameworks in professional ethics: non-maleficence, meaning the avoidance of harm; distributive justice, meaning fair treatment of disadvantaged neighborhoods; collective participation, meaning recognition that communities should be involved in decisions affecting them; and transparent oversight, meaning acknowledgment that human judgment and institutional processes must govern the final outcome. Two co-authors independently coded all 180 answers, providing a check on subjective interpretation. This content-analysis approach converts the vague intuition that an AI answer “seems fine” into measurable, comparable evidence.

The headline result concerns harm avoidance. Non-maleficence was satisfied in all 180 responses, meaning that across every health pathway, income context, and budget condition, the model never clearly recommended a harmful or unsafe built environment change. It did not, for example, propose designs that would plausibly increase pedestrian injury risk or degrade air quality in ways that endanger residents. For a technology often criticized for confident errors, this baseline of safety in the urban design domain is a meaningful finding, suggesting that widely available models have absorbed at least the most fundamental professional conventions about what constitutes dangerous advice.

Distributive justice performed nearly as well. Of the 120 evaluable units, 110, or 91.7 percent, satisfied the criterion, indicating that lower-income neighborhood contexts were rarely handed systematically weaker proposals than their wealthier counterparts. In other words, when the model was asked how to make a disadvantaged neighborhood healthier through changes to walkability, land use, or public space, it generally produced recommendations of comparable quality and ambition. This matters because algorithmic systems have repeatedly been shown elsewhere to encode or amplify socioeconomic bias, and a failure of distributive fairness in urban planning advice could quietly entrench health inequalities.

Where the model faltered was on the procedural dimensions of ethical practice. Collective participation appeared in only 109 of 180 answers, or 60.6 percent, meaning that in nearly four out of ten cases the advice failed to mention that residents and communities should have a voice in shaping their own environments. Transparent oversight fared somewhat better but still fell short, appearing in 136 of 180 answers, or 75.6 percent. These gaps are significant because urban design for health is not merely a technical optimization problem; the legitimacy of a plan depends on how decisions are made, who is consulted, and how uncertainty and accountability are handled.

The differences became especially pronounced when the model faced constrained resources. In the mixed-income prompts that included a budget limitation, collective participation appeared in just 21 of 60 answers, or 35.0 percent, and transparent oversight in 28 of 60 answers, or 46.7 percent. By comparison, the same criteria were present in 73.3 percent and 90.0 percent of answers respectively in the prompt set without budget constraints. The pattern suggests that when forced to prioritize, the model defaults to a technocratic mode, allocating funds and selecting interventions while dropping the language of community engagement and human oversight. This is precisely the scenario in which procedural ethics matters most, since budget-driven prioritization determines which neighborhoods receive health-supporting investment and which are left behind.

“Our findings show that ethical urban design for health depends not only on what physical changes are proposed, but also on how decisions are made, who is involved, and how uncertainty and human oversight are addressed,” Dr. Koohsari said. The statement captures the study’s central conceptual contribution: evaluating AI advice solely on the plausibility of its physical recommendations misses half the ethical picture. A model can propose technically sound bus lanes, parks, and pedestrian improvements while remaining silent on the participatory and institutional processes that determine whether those proposals serve residents equitably.

The practical implications are nuanced rather than dismissive. The authors suggest that LLM-generated advice can serve as a useful initial input for urban designers, planners, and public health professionals brainstorming health-supportive changes to streets, public spaces, transportation infrastructure, land use, and greenspaces. The consistency of harm avoidance and the near-uniform distributive fairness indicate that such tools are unlikely to steer users toward obviously dangerous or inequitable proposals. However, the study is emphatic that these outputs should not replace professional judgment or genuine community participation, particularly when limited budgets require prioritization among competing needs. The procedural gaps identified in the research are exactly the areas where human deliberation, local knowledge, and democratic accountability cannot be delegated to a text generator.

“With appropriate safeguards, LLMs could support more health-informed urban design while ensuring that important decisions remain grounded in professional expertise, community participation, and institutional processes,” Dr. Koohsari concludes. As cities worldwide grapple with aging infrastructure, health inequities, and the pressure to adopt AI tools for efficiency, this study provides a template for what responsible adoption should look like: audit the outputs against explicit ethical criteria, watch closely for procedural failures under resource constraints, and treat the model as a drafting assistant rather than an ethical adviser. The technology, it turns out, has learned what not to propose; whether it has learned how decisions ought to be made is another question entirely, and for now the answer must remain with people.

Subject of Research: Ethical assessment of large language model-generated advice for health-oriented urban design

Article Title: AI and urban design for health: Are large language models ethical advisers?

Article References: AI and urban design for health: Are large language models ethical advisers?. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: large language models, urban design, public health, ChatGPT, built environment, research ethics, distributive justice, community participation, urban planning, AI ethics, health equity, Japan