Two researchers from Northeastern have reviewed more than a hundred studies in *Nature*: AI predicts floods and simplifies public documents, but inherits the biases in its data.
Esteban Moro, a professor of physics, and Ryan Wang, a professor of civil and environmental engineering, both at Northeastern University's Network Science Institute, have published in Nature an analysis of more than a hundred studies that combine urban planning, social computing and geospatial data with artificial intelligence.
Their review distinguishes between two uses. Generative distribution models are used to anticipate flood scenarios or to produce, from satellite images, renderings of future neighbourhoods. Foundation models, on the other hand, intervene upstream of the public debate: they simplify urban policy documents that are deemed illegible, and, in Beijing, agents have taken part in urban planning meetings to bring the point of view of residents, to the point of influencing the development plans chosen.
The authors point to three recurring weaknesses: the resumption and amplification of biases contained in training data, the production of false or misleading information, and the lack of explanation of the decisions produced. "In critical contexts such as disaster response, these limitations become governance risks," summarizes Esteban Moro. They propose an ethical framework based on five requirements: equity, confidentiality, explainability, controllability and participation of residents.
Read the full article on the original source
Go to the source →