Japan has the lowest rate of generative AI use among major economies, yet little is known about how the technology has spread within the country. This study uses the sixth wave of the Japan COVID-19 and Society Internet Survey (JACSIS, N = 27,630), which asks respondents when they first started using generative AI, to reconstruct cumulative adoption curves for Japanese municipalities from November 2022 to January 2026. Because every respondent became able to adopt at the same moment, the release of ChatGPT, the retrospective cohorts support a discrete-time hazard analysis without left truncation. Linking postal codes to municipal statistics, we document four findings. First, adoption follows the urban hierarchy: by January 2026, 50% of residents of the quintile of municipalities with the highest population density had adopted, compared with 35% in the quintile with the lowest population density, and the ratio between the two rose from 0.55 to 0.71 over the four periods, indicating relative convergence alongside a widening absolute gap. Second, most of the geographic gradient reflects who lives where rather than where they live: individual characteristics explain 64% of the population-density coefficient, and the coefficients on population density and on distance to Tokyo become statistically indistinguishable from zero once individual controls are included. Third, municipal population aging is negatively associated with adoption beyond its compositional effect, but only in the first two years of diffusion. Fourth, the job openings-to-applicants ratio in the respondent's workplace area is positively associated with work-related adoption and unrelated to purely private adoption, a pattern consistent with labor scarcity pulling generative AI into workplaces. The association is statistically significant but economically small: a one-point increase in the ratio is associated with a work-adoption probability less than one percentage point higher. Regional differences are concentrated in whether people adopt, not in how intensively adopters use the technology or whether they stop using it: one in six ever-users has lapsed, and lapsing follows age, education, and job content rather than place. The results are correlational, but they suggest that Japan's low national adoption rate is a consequence of its demographic and occupational composition rather than of a distinct rural disadvantage.