Background: AI chatbots now give people with diabetes diet advice, but whether those foods can be bought near home is rarely checked. Methods: In a cross-sectional ecological study of 11,974 California and Michigan census tracts, I linked diabetes prevalence (CDC PLACES 2025) with two USDA food-access maps: all SNAP-authorized stores (2025) and large grocery stores only (2019). I identified double-burden tracts (high diabetes and low income/low access) and hidden food deserts (low access to large grocery stores, but not to any SNAP store). In a pilot, ChatGPT, Gemini, and Claude generated meal plans for the same person across 20 high-diabetes tracts in Metro Los Angeles and Metro Detroit; I scored whether advice was healthy and whether foods were sold at nearby store types. Results: Double-burden tracts were few (71 in California, 19 in Michigan), almost none in the two metros. Food-desert tracts had a 1.5 percentage-point higher diabetes prevalence (95% CI, 1.0-2.0). Counting all SNAP stores hid 567 Californian and 329 Michigan food deserts; in Metro Detroit, all 97 disappeared, including 51 high-diabetes tracts with about 19,500 adults with diabetes. Chatbot advice was healthy, but in hidden tracts 30% of recommended foods were not reliably sold nearby; the chatbots barely adjusted. Of 55 places they named, 13% had errors, and most were over a mile away. Conclusions: The current USDA map obscures food deserts, especially in Detroit, and AI diet advice barely adjusts for them. AI nutrition tools for diabetes should include a place-aware check built on public food-access data.