The Discursive Construction of Urban Consumption Spaces: An Analysis Based on AI Hallucinations

In the era of Artificial Intelligence (AI), Large Language Models (LLMs) have become important informational mediators through which the public perceives urban spaces, and AI discourse has emerged as a powerful force in the construction of urban spaces. Using Guangzhou as a case study, we categorized urban consumption spaces into four types: shopping spaces, catering and entertainment spaces, tourism and leisure spaces, and residential and commercial housing. We constructed an evaluation question set for hallucinations in urban consumption spaces within a discourse-power framework and used hallucination tests to examine the commonalities and differences between Chinese and international AI models, namely, DeepSeek and ChatGPT, in the production of spatial discourse, thereby explaining how AI hallucination discourse constructs urban consumption spaces. The main findings of this study are as follows. (1) In the hallucination tests of urban consumption spaces, ChatGPT exhibited lower hallucination rates than DeepSeek at both the overall level and across individual categories. Residential and commercial housing emerged as high-incidence domains of hallucinations for both AI models, whereas the most pronounced divergence in hallucination rates between the two models occurred in tourism and leisure spaces. The primary sources of AI-generated content, in descending order, were news media, individual or commercial institutions, government agencies, and online encyclopedias. Both models tend to respond to mainstream spatial discourses, demonstrating a limited capacity for revealing the complex, diverse, and contradictory realities of the city. Specifically, ChatGPT favors generalized frameworks in its depiction of urban consumption spaces, whereas DeepSeek's spatial narratives display a planning-oriented logic aligned with urban development strategies. (2) By integrating and reproducing specific discourses originating from governments, news media, and commercial institutions, AI discourse operates as a novel power subject that constructs multiple "realities" and promotes the production of meanings attached to consumption centers, symbolization of architectural landscapes, and technologization of consumption spaces and also adjudicates spatial value, allowing its power to operate in a "rational" manner. (3) The AI hallucination discourse constructs space by producing subject positions tailored to users, such as "supporters of urban development," "experience-oriented consumers," "beneficiaries of technological progress," and "astute investors." As users identify with and accept these positions, they enact specific consumption-space practices grounded in particular forms of knowledge, generating new data that are subsequently mobilized to reproduce the same discursive system. In this process, a specific knowledge regime is sustained, and power continues to operate. From a discourse-power perspective, this study elucidates the pathways through which urban consumption spaces are constructed by AI in the era of artificial intelligence. Although, it advances our understanding of the modes and impacts of urban knowledge circulation amid the rise of generative AI, critical reflection on the discursive and power relations embedded in technological products contributes to ethical scrutiny of smart city practices.