AI Tool Maps Climate Equity Plans Across US Cities

Artificial intelligence can now read hundreds of American city climate plans, extract their policies, and match cities with similar approaches, according to a new study published in Computational Urban Science. Researchers developed a framework called the Retrieval-Augmented Policy Analysis Framework, or RAPAF, which uses large language models combined with retrieval-augmented generation to systematically analyze climate equity and action plans across the United States. The work, led by Seung Jun Choi and Sangyup Lee, demonstrates how planners could use AI to compare policy practices across jurisdictions without manually reviewing thousands of pages of documents.

The study focused on US cities with populations exceeding 100,000 that have adopted climate equity plans. Drawing on a list compiled from Wikipedia with estimated 2024 population figures, the researchers identified 346 candidate cities, of which 192, or 55 percent, had regional climate equity or action plans. These plans carry different names in different cities: New York City calls its effort PlaNYC: Getting Sustainability Done, Houston’s is the Climate Action Plan: Resilient Houston, and Denver’s is the 80 × 50 Climate Action Plan. The team collected the most recent version of each plan from publicly available repositories in PDF format. In total, the review process involved examining 21,489 pages comprising 5,834,348 words, with documents spanning 2009 to 2025. The majority of plans, 129 documents representing 67 percent, were published between 2020 and 2025.

RAPAF operates in three stages: indexing, retrieval, and generation. During indexing, policy documents were segmented using a recursive character-based text splitter with a chunk size of 1,000 tokens and an overlap of 200 tokens to preserve semantic continuity across segments. Each chunk was embedded into a dense vector representation using OpenAI’s text-embedding-3-small model, and the resulting embeddings were indexed in a FAISS vector store, which supports approximate nearest-neighbor search over large corpora. During retrieval, each structured prompt was embedded and used to query the FAISS index. Rather than standard top-k retrieval, the team applied Maximal Marginal Relevance, which iteratively selects passages that maximize relevance to the query while minimizing redundancy among selected passages. They set the number of retrieved chunks at five, the initial candidate pool at twenty, and the relevance weighting parameter at 0.7. Random sampling was additionally applied over the candidate pool to mitigate retrieval bias toward topically similar but non-representative excerpts.

In the generation stage, retrieved passages were concatenated with a structured system prompt and passed to GPT-4o-mini as context. The prompt explicitly constrained the model to reason only from the provided passages and required citation of source documents and page numbers, creating an attribution chain designed to mitigate hallucination and support auditability. The temperature parameter was fixed at zero, which collapses the output distribution to deterministic decoding and ensures reproducible outputs across repeated runs. The pipeline was built using LangChain, an open-source framework launched in October 2022 that provides modular abstractions for document loading, chunking, embedding, vector storage, retrieval, and generation. Retrieval-augmented generation itself was introduced by researchers in 2020 to address limitations of large language models, including their tendency to generate factually incorrect or fabricated outputs that may appear superficially plausible, by combining pre-trained parametric language models with a non-parametric external retrieval memory.

To validate the extraction performance, the researchers compared the policy relevance extracted and evaluated by ChatGPT against human-labeled ground truth across the 192 climate equity and action plans. The ground truth was generated through a structured manual coding procedure by a single expert coder certified by the American Institute of Certified Planners. The overall average accuracy, defined as the overall proportion of correctly classified instances, was 70 percent, with relatively higher performance observed in category-level classification compared to section-level classification. Section-level accuracy for policy, strategy, and action was 64, 63, and 68 percent respectively, while category-level accuracy for the same three elements reached 64, 79, and 81 percent. The average precision, recall, and F1-score across all tasks were 69, 88, and 77 percent respectively. The results indicate that the model tends to achieve high recall across most tasks but comparatively lower precision in certain cases, particularly at the section level.

The findings reveal what American climate equity plans actually contain. No city achieved a full score of 20 out of 20 for the transportation or energy criteria evaluated. On average, transportation themes scored 7 for policy, 6 for strategy, and 8 for action across the evaluated cities. Land use and transportation integration emerged as the most common transportation policy theme, with transportation resilience receiving the least attention. For transportation strategies, electric vehicle charging infrastructure was the most common theme, whereas smart mobility and integrated transportation planning were the least represented. Across the energy themes, the average scores were 9 for policy, 8 for strategy, and 6 for action. Energy efficiency was the most frequently addressed theme in energy policy, while grid modernization was the least addressed. Renewable energy was the most common focus for energy strategies, and implementing energy efficiency measures was the most common action item, though conducting formal energy impact assessments was observed in only one case.

The plans commonly address transportation, environmental, and energy-related measures aimed at reducing greenhouse gas emissions, and they predominantly employ affirmative language. Using a BERT-based classifier fine-tuned for binary sentiment classification, the researchers scored the extracted responses for semantic polarity. Among documents that did not include uncertain responses, negative classifications were identified in only 19 policy-related phrases, 19 strategy-related phrases, and 17 action-related phrases at the section level. Affirmative cases typically include language supporting future directions and planned actions, while negative cases often reflect policy language acknowledging limitations, challenges, or unresolved issues. One phrase from Brownsville explicitly limits the potential impact of the policy, South Bend highlights explicit cost constraints, and Tampa, Florida has a dedicated section that lacks substantive detail. The researchers note this polarity analysis was not intended to evaluate policy effectiveness but to diagnose how explicitly or assertively climate or equity measures were articulated.

Built on these extracted elements, the team developed a content-based recommendation system that identifies the top five most similar city pairs using cosine similarity. The similarity reflects patterns in recorded action items rather than geographic proximity. When Las Vegas was used as the reference city, the system identified Chico, Ann Arbor, Berkeley, Richmond, and Long Beach as having analogous transportation action patterns. Common practices across these cities include promoting shared mobility options, facilitating community workshops, and launching transportation awareness campaigns, while the documentation appears to lack action items explicitly focused on enhancing transit accessibility, supporting vanpool services, and conducting needs assessments. For energy-related actions, Fort Lauderdale shared similarities with Saint Paul, Des Moines, Sterling Heights, Columbia, and Buffalo, with commonly observed actions including supporting energy transition efforts and facilitating stakeholder engagement. The researchers observed that policy practices often exhibit geographically patterned similarities, though they emphasize this clustering does not imply causality.

The study also examined how the planning profession itself is changing in the era of AI by analyzing job postings scraped from LinkedIn and Indeed on October 23, 2025. Of 69,116 US job postings initially collected, manual review and deduplication left only 83 postings directly responsible for the governance and physical planning of urban systems. Explicit references to AI were largely absent from the analyzed postings despite frequent requirements for GIS and data analysis skills, and domain-specific planning areas such as transportation, environmental planning, land use, and housing remained central across positions. Of the 83 job postings, 48, or 58 percent, placed notable emphasis on communication-oriented responsibilities, including internal and external stakeholder communication, strong verbal and written communication skills, cross-functional coordination, and public-facing communication. The topic modeling results suggest planning jobs are evolving toward a hybrid professional model combining coordination-oriented generalist roles with specialized managerial and technical functions.

The researchers position AI not as a substitute for planners’ normative judgment but as infrastructure that augments cross-city learning. The system achieved a 94 percent recall rate for category-action classification, demonstrating exceptional capability in exhaustive retrieval at the granular level, though the precise selection of higher-level policies remains a nuanced challenge. The authors recommend future implementations prioritize a human-in-the-loop framework that would enable planners to audit AI-generated insights. They also acknowledge limitations: responses may still be sensitive to design choices such as text chunking strategies and chunk size despite deterministic parameters, future model updates may influence semantic interpretations, and policy identification relies on LLM-driven semantic interpretation rather than manual human coding. The demonstrated recommendation system adopts a content-based approach without incorporating explicit geographic or spatial information, and alternative approaches such as collaborative filtering or spatially informed recommendation models may surface different patterns of policy convergence. The study concludes by envisioning localized and personalized AI-assisted systems designed to support professional planning judgment within urban governance contexts.

Subject of Research: Using retrieval-augmented large language models to extract and compare climate equity policies across US cities

Article Title: Mapping and comparing climate equity policy practices using RAG LLM-based semantic analysis and recommendation systems

Article References: Choi, S. J., & Lee, S. (2026). Mapping and comparing climate equity policy practices using RAG LLM-based semantic analysis and recommendation systems. Computational Urban Science, 6(1), Article 49. https://doi.org/10.1007/s43762-026-00279-0

Image Credits: AI Generated

DOI: 10.1007/s43762-026-00279-0

Keywords: climate equity, large language models, retrieval-augmented generation, urban planning, policy analysis, recommendation systems, ChatGPT, semantic analysis, climate action plans, artificial intelligence, planning practice, cross-city comparison