you really have to try hard to make a model hallucinate when asked to summarize an email. I think they didn't mention it because they can't guarantee 100%, but it's virtually on non-issue for such task.