How are you using AI to generate test assertions specifically — not just test structure?

Most AI test generation demos show the happy path: given a function, generate a test skeleton. I've seen this work OK. What I haven't seen good examples of: using AI to generate the **assertions** specifically. For example: given a complex API response object (nested, 40 fields), what assertion strategy do you prompt for? Do you assert every field? Just the ones most likely to regress? How do you prevent the AI from writing assertions that are too brittle or too loose? Any prompts or workflows that have worked for you at the assertion level specifically?