Airbus has recommended that upset recovery training should not be conducted in a flight simulation training device because simulator motion does not accurately represent aircraft motions in unusual attitudes and there is a lack of validated flight envelope model data at extreme conditions.
In addition, most simulator upset recovery training is conducted with the motion on, which may not increase the value of the training. Current devices can only simulate 1g flight, a fleeting condition in the early stages of an upset recovery.
Simulation may even provide strong negative training, as was the case in the crash of American Airlines flight 587 in New York in 2001. Inappropriate aggressive rudder usage to counter wake turbulence-generated rolling motions caused the Airbus A300's vertical stabiliser to snap off. At the time American Airlines had been teaching the use of rudder to help in upset recovery. While the A300 simulator may have accurately represented aircraft response to rudder inputs, it was not an engineering loads simulator and the crews would not have be alerted to the excessive structural loads their inputs were generating.
from flight global:
Getting back in the envelope with CAE/APS upset training