Urban planning is a complex task. The question of whether Artificial Intelligence (AI) could be employed to assist urban planning is not new and recent developments suggest exciting new opportunities on the horizon. In this article, we examine how Generative Pre-Trained Transformer (GPT)-4, OpenAI’s latest large language model launched in April 2023, could be applied to urban planning tasks. GPT is a machine learning algorithm that responds to input (e.g. text prompts, image inputs) with human-like text as output. GPT-4 was trained on massive amounts of text and data, and functions like a chatbot. Through “conversations” with GPT-4, we studied its capabilities in accessing planning concepts and strategies, and in utilising technical-oriented data and tools. In particular, we explored GPT-4's abilities across five themes: A. macro strategies and policies, B. the Master Plan and local planning, C. urban design, D. development control, and E. infrastructure planning.
The full working paper with technical details can be found here.
A. Macro Planning Concepts & Strategies GPT-4 demonstrated its ability to interpret and respond to complex questions in the context of urban planning and design in Singapore. When prompted with a newspaper commentary discussing meritocracy, it could connect such a topic with urban planning and policy-making in Singapore’s context to provide well-formed textbook answers. It could also suggest relevant quantitative analytical methods and engage in meaningful discussions. However, it was unable to provide working links to web publications cited in its answers. (To refer to Figures 1 & 2 in the working paper) B. Master Plan and Local Planning A conversation on the URA Master Plan and local planning was initiated with text from a local news report relating to the en-bloc sale and potential redevelopment of a mixed-use development in Singapore. GPT-4 was able to provide sound answers on general and conceptual planning issues, as well as local knowledge of the district. However, it was unable to provide accurate responses relating to the Master Plan’s land uses and planning processes, and provided incorrect information in some instances. Thus, at this point in time, GPT-4 is not yet a good source of information on the Master Plan and other geospatial-related matters. (To refer to Figures 3 & 4 in the working paper) C. Urban Design On urban design, GPT-4 was able to respond well when quizzed on general concepts and strategies, citing elements such as connectivity, diversity and streetscape design. It could also generate code for simple sketches and 3D visualisations of scenarios when given physical descriptions. However, it remained poor in its ability to surface working links to web publications, and was unable to provide accurate urban design guideline parameters such as building setbacks and heights. (To refer to Figures 5, 6 & 7 in the working paper) D. Development Control When text from a regulatory circular pertaining to definitions of floor area were fed to GPT-4's chatbot platform, GPT-4 recognised and brought up general development control concepts. Nevertheless, it was unable to provide accurate replies relating to specific regulatory guidelines in Singapore as some of its responses appeared to be formulated based on information or practices from other cities. (To refer to Figure 8 in the working paper) E. Infrastructure Planning GPT-4 was queried on a news report regarding an upcoming major road infrastructure in Singapore. It was able to respond well on general infrastructure planning and transport modelling concepts, including the interaction between infrastructure and land-use planning, which again seemed to be formulated based on information or practices from other cities. It was however limited in spatial abilities as it faced challenges in deconflicting two underground pipes in relation to their depths and diameters. (To refer to Figure 9 in the working paper)
Overall, GPT-4 has significant potential in assisting planners to outline planning strategies, check for ‘blind spots’ in conceptual frameworks, draft concepts or talking points, and outline analytical approaches, amongst its other abilities. However, it is currently not a reliable source of information if one were to seek advice on the Master Plan and other geospatial-related matters. While it is able to cite/make references to studies and publications in its responses, its inability to provide working links to these information sources remains a limitation. The data and information which GPT-4 was trained on likely determines the quality of its response. This points to the need to better understand the source of its knowledge, and the scope of specialised domain information it should be provided with to function effectively. In conclusion, we find that GPT-4 does have the potential to revolutionise urban planning by assisting and informing planners in their planning processes. However, further research must be conducted to improve its capabilities and mitigate its limitations. Given GPT-4's ability to confidently and coherently present inaccurate information, human planners still play a fundamental role and should be held responsible to fact check, evaluate, think critically and apply knowledge while working alongside such AI tools. We are excited by the prospects of GPT-4, and AI’s continued evolution, and their potential to enhance/advance urban planning in Singapore if incorporated into our existing planning processes and workflows.
Additional Notes: