With the widespread application of large language models, AI-assisted reading is increasingly reshaping reading practices and cognitive modes. Against this background, this study examines the mechanism through which AI-assisted reading influences deep reading competence by adopting cognitive pathway reconstruction as its theoretical perspective and by drawing on questionnaire data from 1,025 users. Rather than treating AI-assisted reading as either simply beneficial or simply harmful, the study focuses on how AI tools reorganize the cognitive pathways through which readers acquire information, construct knowledge, and sustain attentional engagement, and how such reorganization may generate both enabling and constraining effects on deep reading. The analysis is based on a questionnaire survey conducted in April 2025 among users who regularly engage in AI-assisted reading. The sample is characterized by a relatively young, highly educated, and urban profile: 64.4% of respondents are aged 18 to 34, 87.7% hold a bachelor's degree or higher, and 55.0% live in first-tier cities. On this basis, the study uses descriptive statistics and comparative analysis to investigate patterns of AI tool use in reading, perceived changes in reading-related behavior, and self-reported shifts in capacities associated with deep reading. Through this mechanism-oriented design, the study seeks to reveal not merely whether AI changes reading, but how the organization and sequence of cognition in reading are being reconfigured. The findings show that AI-assisted reading reconstructs deep-reading pathways along three dimensions. First, cognitive processing shifts from gradual exploration to more immediate acquisition. A substantial proportion of respondents' report using AI for terminology clarification (77.9%), content summarization (67.6%), and assistance in understanding complex texts (54.2%). These practices indicate that AI compresses the path of information access and reduces some of the cognitive burden traditionally associated with exploratory reading. At the same time, such efficiency gains may weaken readers' motivation to undertake the generative cognitive effort required for sustained and self-directed deep reading. Second, knowledge construction shifts from independent autonomy to human–AI collaboration. Although 84.2% of respondents report improved knowledge integration ability, the study suggests that this improvement should not be understood simply as a direct enhancement of intrinsic cognitive competence. Instead, AI-assisted reading introduces a collaborative mode in which interpretation, summarization, and perspective generation are partially externalized to AI systems, thereby reconfiguring the locus and process of meaning construction. Third, the form of attentional engagement shows structural differentiation. Unlike the first two dimensions, which display relatively clear directional tendencies, changes in sustained attention are more heterogeneous: 48.9% of respondents report improvement, whereas 25.7% report decline. This suggests that AI-assisted reading does not exert a uniform effect on concentration, but instead reorganizes attentional patterns in differentiated ways. Taken together, the study argues that cognitive pathway reconstruction produces a complex dual effect. While AI-assisted reading can enhance information-processing efficiency and broaden access to knowledge, it may also exert a profound influence on cognitive autonomy, knowledge internalization, and the conditions necessary for sustained deep reading. These findings highlight the need for continued attention to how deep reading competence can be preserved and developed in AI-mediated reading environments.