Capt Claret points out a very important consideration - we all like to play pilots with the sim and do interesting things in spare playtime - and these activities have a very useful potential in their training value.
However, many forget that the sim is a computer, not an aircraft, and that what we observe is the computer-generated response, not that of the real aircraft. In those regions of the envelope where the fidelity is tested and proved to be adequate, then the result is pretty much that the aircraft response is produced - otherwise the response may be quite unrepresentative, and only of a generic training value.
To highlight this consideration -
(a) some months ago I was co-opted (a disadvantage of being around at 2 in the a.m.) to fly a comparative test sequence pre- and post-software mod on my current sim. The results were quite impressive - the pre-mod responses in regard to the particular control problems were quite unlike those alleged of the aircraft, while the post-mod were very believable.
(b) in similar vein, we routinely do fairly extreme UA departures for confidence building during endorsement programs. While there are no rational acceleration cues, nor are the instrument dynamics necessarily anything like those on the aircraft - there is still a useful value to the student in building confidence to recognise the various situations and to apply the conventionally recommended recovery protocols.