Autopilots are not omniscient nor omnipotent.
Remember this: computers are better than humans at routine things, humans are better than computers at the unexpected and unanticipated.
Once a flight exceeds certain parameters it is usually safe to assume something unexpected/unanticipated has happened, at which point the human should take over.
Try the following experiment:
Find an open space in the room without anything for you to lean on or against.
Stand on one foot with your arms extended out to either side.
See if you can do this for 60 seconds. You probably can.
Repeat this with your eyes closed. Virtually no one can.
As touched on earlier, you have semicircular canals in your ear , which provide information to the lower brain on rotational accelerations in the pitch, roll and yaw axes. However, prolonged rotation (beyond 15-20 s) results in a cessation of semicircular output, and cessation of rotation thereafter can even result in the perception of motion in the opposite direction."
(From Wikipedia)
So get in a banked turn to the left for half a minute, you are no longer receiving input from this system about “hey…we’re banking to the left!”, then when you roll wings level your semicircular canals sense the change and say “Holy Cow-we were wings level, now we’re banking to the right!”
The most important thing is to maintain a good scan and recognize the symptoms of spatial disorientation-that what my body feels is different than what my instruments show.
No aircraft I’m aware of have automation taking over as usedtobe suggested. The USAF has a developmental program to add that capability, with ground avoidance, into late-model fighters. But this is still an experiment, not a fielded system.
The central issue is failure modes. Current autopilots are designed under the assumption that all systems are operating normally. They are NOT designed to cope with incorrect or conflicting inputs. In the old days (1960s/70s), they’d blindly follow bad inputs into a crash. Since about the 80s in most cases they’re now smart enough to at least figure out that they’re getting conflicting inputs & shut off since they can’t tell which inputs are true and which are false.
In theory one could use a bunch more redundant sensors and artificial intelligence to build an autopilot which could continue operating through more sorts of malfunctions and unusual situations. Or instead of AI you could rely on the natural intelligence you’ve already got on board. Which also isn’t perfect, but does possess some significant strengths the current state of automation art does not.
If someone truly believes “Especially in a stall in zero visibility it seems like the auto pilot should be doing the recovery, not the human pilot.” then they actually believe they’d be better off with no human pilot at all.
I have no doubt that fully automated pilotless aircraft, including not having any data link to a human on the ground can and will eventually be built. But we’re many decades away from that being possible, much less acceptable to the traveling public or those who live below.
When I was in flight training under the hood (a device that let me see instruments but not out the windows) my instructor told me: The average amount of time a VFR pilot can maintain control of the aircraft when flying into IFR conditions is 18 SECONDS. I believe it. My time under the hood showed me how hard it is to fly with just the instruments. You have to be aware of heading, pitch, roll and yaw. I’d notice one of those was off, and I’d focus on fixing that, and one (or all) of the others would deviate.
J.
I think this was what happened with Air France 447.
AF 447 is the textbook case of pilot disorientation, as I understand it. Some sensor had iced up and started returning weird readings, so even though the autopilot was flying fine, the pilot disengaged it because he thought there was a problem. He then got so confused by the inputs he was seeing that he put the plane into a stall so freakishly severe that the stall warning shut off, because the computer thought the speed and position readings it was getting had to be the result of computer error. This had the perverse result that when the pilot did let go of the stick, the plane would fall back into a more normal stall configuration, and the warning would come back on. Ultimately, rather than diving to come out of the stall, he just kept pulling the stick all the way back, because that made the warning stop.
Essentially, the instruments confused the pilot, and then the pilot confused the flight computer.
I installed Google Vertigo, then I fell down a flight of stairs while I was texting. Not sure what I was thinking.
Interestingly enough, the Soyuz capsule (still, IIRC) uses a lil’ doll attached to a string as a ‘directional’/accelerometer indicator. Here is a decent pic of it from a quick Googly search. When doll free-floated, they were in true ‘freefall’ (ie orbit). Low-tech, but very effective from what I’ve heard. Not too hard to tell what is happening with doll right there 
I’ve had lots of time ‘cloud-busting’ during aerobatic flying (usually an open-cockpit Stearman w/ basic instrumentation), and it is really, really hard to keep things oriented by senses. Takes VERY close attention to what instruments are telling you, for sure. My Instructor made me practice recovering from purposed stalls of various power in clouds routinely (higher-deck clouds, of course), to recover if/when I failed to get control in-cloud) and it made a huge difference in how I reacted compared to not having such experience(s). First few times scared the crap out of me, but it got much easier with experiences. Still dangerous no matter what. We always had ‘chutes on, just in case 
And always done out over Gulf for others’ safety, etc…
If you want a long, sad, read, look up
Pinnacle Airlines Flight 3701
on NTSB’s Accident/Incident database
http://www.ntsb.gov/_layouts/ntsb.aviation/index.aspx
The idiots wanted to take the plane up faster then it wanted to go.
At one point (a couple, actually) the plane pushed the stick forward to lower the nose.
This was a small twin jet - don’t know what all the heavy iron carries.
Big thanks to everyone who has responded with their explanations, analogies, and experiments I can do at home. I find all of this fascinating. As I learn more about this, I wonder if part of the reason the banks feel more obvious to me as a passenger is because I’m always in a window seat, which is going to experience more change during a bank, whereas the pilots are in the exact center of the plane, so the force of a bank on their bodies is less exaggerated. Do you think this could help explain it?
More likely, being able to look out a window makes it more likely you’ll have a visual input regarding the bank.
I wouldn’t expect a computer to determine that the best course of action for an emergency is to make a landing in the Hudson river. Human pilots are needed to make complicated decisions. I still think that for a stall an auto-pilot could be programed to use the instruments to determine attitude and put the plane nose down without the problem of human spatial disorientation.
No spatial disorientation in that crash, but some truly astonishing behavior from (nominally) professional pilots.
I doubt your self-professed ability to detect these things. I don’t see how you could ever know if you were right or wrong. You think you’re guessing right, but you probably aren’t.
My instructor also did the “try to fly blind for a while” test. I thought I was straight and level, but I was in a dive to the right. I tried again, and I was equally far off. The fact is you really can’t tell what the plane is doing just by feeling it.
If you think your butt or your ears can tell up from down, you are wrong.
And there are a whole bunch of dead pilots and passengers (including Buddy Holly, Big Bopper, and Richie Valens) who say so.
That is a stall identification device, a stick pusher in this case.
In a traditional light aircraft with straight wings that we all learn to fly in, the stall is accompanied by two things. First, just prior to the actual stall, a buffet is felt through the airframe and controls, it is caused by the air becoming turbulent over the wings. Second, the actual stall is identified by the nose dropping all by itself. It does that because of where the lift vector acts in relation to the C of G. That then is the ideal stall characteristics of a docile training aircraft. Light aircraft also have a buzzer that serves as a stall warning because the buffet can be subtle.
As planes get bigger and faster, the design requirements to make them bigger and faster mean that the buffet (stall warning) and nose drop (stall identification) become either masked or just aren’t present at all. In order to simulate those two classic stall characteristics so that a pilot’s initial training still translates to the bigger plane a stick shaker is installed and a stick pusher. The stick shaker is a device that rattles the control column as an analogue to the airframe buffet and warns of the approaching stall. The stick pusher pushes the control column forward abruptly as an analogue to the traditional nose drop and serves to identify the stall itself. The stick pusher may push just before the true stall occurs because aircraft that require a pusher may have such poor stall characteristics that the true stall needs to be avoided altogether.
So the Pinnacle Airlines jet had a stick pusher which is typical for swept wing aircraft especially large ones.
Neither the stick shaker or stick pusher is an example of the autopilot taking control though, they are both simple mechanical devices that serve to warn of the approach stall and identify that the stall has occurred.
Reading a bit more about stick pushers I’m reminded that they are more common on T-tail aircraft. Low tail jets regardless of size, are less likely to need one.
A famous VFR into IFR demo… 178 seconds to live
I remember reading an article on the development of the F-35 or F-22; the pilot was giving a demo and entered a stall or a spin - he then took his hands off the stick and the autopilot recovered the aircraft.
