The Autopilot's Autonomous disconnect has some instructive considerations. Long ago, in a discussion with PJ2 and Will Fraser, it was determined that excursion limits AP had programmed into it would cause a potentially unexpected disconnect, and a degradation of Flight Law.
Now this is known. Let us add to that possibility the existence of severe turbulence, with its downside of allowing "Out of Phase" reactions of the airframe to rapid excursions in Pitch, Roll and Yaw. The AP program allows 45 degrees in Roll and 13u, 9d in Pitch. At this point, there should be a discussion of whether it is wise to continue into turbulence in Autoflight. Not now.
How in essence shall we differentiate between out of synch airframe response and controls position? The AutoPilot is programmed by control position and airframe aspect. So is it correct to eliminate the possibility of Upset whilst in AutoPilot? No. Assume the 330 was close to upset, or at least Unusual attitude at drop out of AP.
Instead of poor stability with the AP engaged, the Pilots are confronted with no stability when it quits. Law is Alternate, and hand flying in turbulence with an airframe that is aerodynamically, busy, is the handful of Enchilada Will Fraser posted. Getting the "Feel" under such conditions is not duplicable in the sim. Whilst attempting to regain stable and level flight, sufficient disturbance of airflow about the sensors is certainly possible to create duff reads.
Perhaps the "Upset" was not an identifiable moment, but an in and out thing that was mitigated for a time by the pilots, until perhaps air loads passed the ultimates for a control load, or for aero flight. "Out of Phase" has an interesting relationship to AA587.