[javindo]

Whilst I am a huge advocate of advancing research in AI, I think we are a very long way off it being safely feasible in flight. Flight is a process so insanely complex and susceptible to random change and failure that we simply don't have the requisite provable AI in place yet to support it.

A good example of where, even the highest level of automation in an aircraft, failed to assist in an emergency was Qantas Flight 32 [1]. A physical failure on the plane caused sensor and control failures, something rendering AI almost completely useless. Admittedly, humans can also have failures and this has lead to many crashes before, however humans have also used exemplary skill, experience, and abstract problem solving to correct mechanical failures in aeroplanes countless other times.

[1] https://en.wikipedia.org/wiki/Qantas_Flight_32

[DennisP]

How about this: every crew member has a wireless button. If two of them push the button, the AI irrevocably takes over and lands the plane at the nearest airport. Pilot locks self in cockpit and heads for the mountain? Push the button.

If the AI would crash the plane one time in a thousand, and the pilot would intentionally crash the plane one time in a million, then this system reduces risk to one in a billion.

It helps against terrorists too. They have to disable almost the entire crew at once, or they'll have a nice leisurely ride to a waiting SWAT team.

[fulafel]

Airline safety culture is highly developed in the course of a long time, and a lot of it is enabled by assuming benign actors. A lot of safety principles, assumptions and practices would go out the window when you start factoring in insider adversaries and preventing the secondary plots that would pop from your threat model after adding the first couple of controls.

[aaronem]

> they'll have a nice leisurely ride to a waiting SWAT team

...and no reason not to murder the entire passsenger load on the way there, save possibly to have hostages once the aircraft's on the ground. I really don't think you have thought this all the way through.

[DennisP]

I don't see how it gives them any extra incentive to murder. I'd say it gives them less, since there's nothing to gain.

It's also a better situation by far than letting them crash the plane into a building, and if the system is known to exist it's a deterrent against hijacking planes in the first place. And if there's a struggle on board, it's less likely that the struggle will cause a crash.

[aaronem]

> I don't see how it gives them any extra incentive to murder.

Schrecklichkeit, maybe?

[frik]

Yes, good idea. But the ground control should then supervise the AI of the aircraft and takes over control if needed. Otherwise a sensor defects could have consequences, like Lufthansa Flight 1829.

[DennisP]

If you allow external control, there's also risk of an external hack. It might work out if external control is only possible after the button is pushed; that would still allow for a hack coordinated with onboard attackers, but maybe that's giving terrorists too much credit.

But I'd be happy relying on the steep odds against hardware failure happening on the same flight as human attack, as long as we can keep those variables reasonably independent.

[frik]

Launched into production during 1984, the Airbus Industries Airbus A320 became the first airliner to fly with an all-digital fly-by-wire control system. http://en.wikipedia.org/wiki/Fly-by-wire

Technically it's already possible. It's a lot easier than autonomous cars, as all the technology is already in place and less other human-beings that could interfere the system. In a few years, commercial airplanes will probably be like the US combat drones that fly autonomously in Middle East/Africa while being supervised from Las Vegas.

But you are right, that with AI other issues surface. An accident of the A320 in late 2014:

On 5 November 2014, Lufthansa Flight 1829, an Airbus A321 was flying from Bilbao to Munich when the aircraft, while on autopilot, lowered the nose into a descent reaching 4000 fpm. The uncommanded pitch-down was caused by two angle of attack sensors that were jammed in their positions, causing the fly by wire protection to believe the aircraft entered a stall while it climbed through FL310. The Alpha Protection activated, forcing the aircraft to pitch down, which could not be corrected even by full stick input. The crew disconnected the related Air Data Units and were able to recover the aircraft. The event was also reported in the German press several days before the Germanwings crash. The German BFU (Aircraft Accident Investigation Bureau) reported on the incident on 17 March 2015 in a Bulletin publishing the flight data recorder and pitch control data in English and German. As a result of this incident an Airworthiness Directive made mandatory the Aircraft Flight Manual amended by the procedure the manufacturer had described in the FOT and the OEB and a subsequent information of flight crews prior to the next flight. EASA issued a similar Airworthiness Directive for the aircraft types A330/340. -- http://en.wikipedia.org/wiki/Accidents_and_incidents_involvi...

At first the Germanwings Flight 9525 sounded very similary to the Lufthansa Flight 1829. There were speculation that the same issue happened again, but then they found the voice recorder (co-pilot...). Germanwings Flight 9525 was an Airbus A320-200: http://en.wikipedia.org/wiki/Germanwings_Flight_9525