3 ms·

> because that includes accidents like https://en.wikipedia.org/wiki/Germanwings_Flight_9525 https://en.wikipedia.org/wiki/Germanwings_Flight_9525 > He had set

by ejolto 6y ago

> because that includes accidents like https://en.wikipedia.org/wiki/Germanwings_Flight_9525 https://en.wikipedia.org/wiki/Germanwings_Flight_9525 > He had set the autopilot to descend to 100 ft (30 m) and accelerated the speed of the descending aircraft several times thereafter. That could have been avoided if the autopilot refused to crash the air plane. To be fair, I bet Boeing's autopilot also accepts commands to crash the plane. However this accident could have been avoided if the autopilot had refused to comply with an obviously deadly commands.

macintux 6y ago
I'm skeptical that an AI for airplanes exists that can distinguish between "obviously deadly commands" and "trying to rescue the plane during a mechanical emergency".
- notahacker 6y ago
  A hypothetical system would actually function almost exactly the same as the much derided MCAS. If angle of attack sensors indicate an inappropriate angle of ascent/descent, overcompensate for or override the pilot inputs. As we've learned at great cost, angle of attack sensors are wrong more often than pilots are homicidal though, which is part of the reason why the form of intelligence used to deter and attempt to prevent obviously deadly pilot actions is a lump of meat called a copilot.
  - ejolto 6y ago
    > a lump of meat called a copilot. Looks like that's how they fixed the issue. > aviation authorities in some countries implemented new regulations that require two authorized personnel to be present in the cockpit of large passenger aircraft https://en.wikipedia.org/wiki/Germanwings_Flight_9525#Regulatory https://en.wikipedia.org/wiki/Germanwings_Flight_9525#Regula...
vultour 6y ago
An A320 already does much more crash-prevention than a 737 because of its FBW system. It’s impossible to stop someone who’s determined to crash the plane unless you make it fully autonomous.
bipson 6y ago
You're onto something, keep on thinking about it... https://en.wikipedia.org/wiki/Lufthansa_Flight_2904 https://en.wikipedia.org/wiki/Lufthansa_Flight_2904 Please excuse the snark, but there is more to flight safety than a Hacker's "AI's the best"-mentality. We actually have a lot of experience gathered from accidents and non-accidents how these systems should behave in a general sense, and "overriding pilot-input" has proven not to be a good idea. The discussion above was about details and UX.

by ejolto 6y ago

macintux 6y ago

I'm skeptical that an AI for airplanes exists that can distinguish between "obviously deadly commands" and "trying to rescue the plane during a mechanical emergency".

notahacker 6y ago
A hypothetical system would actually function almost exactly the same as the much derided MCAS. If angle of attack sensors indicate an inappropriate angle of ascent/descent, overcompensate for or override the pilot inputs. As we've learned at great cost, angle of attack sensors are wrong more often than pilots are homicidal though, which is part of the reason why the form of intelligence used to deter and attempt to prevent obviously deadly pilot actions is a lump of meat called a copilot.
- ejolto 6y ago
  > a lump of meat called a copilot. Looks like that's how they fixed the issue. > aviation authorities in some countries implemented new regulations that require two authorized personnel to be present in the cockpit of large passenger aircraft https://en.wikipedia.org/wiki/Germanwings_Flight_9525#Regulatory https://en.wikipedia.org/wiki/Germanwings_Flight_9525#Regula...

vultour 6y ago

An A320 already does much more crash-prevention than a 737 because of its FBW system. It’s impossible to stop someone who’s determined to crash the plane unless you make it fully autonomous.

bipson 6y ago

You're onto something, keep on thinking about it... https://en.wikipedia.org/wiki/Lufthansa_Flight_2904 https://en.wikipedia.org/wiki/Lufthansa_Flight_2904 Please excuse the snark, but there is more to flight safety than a Hacker's "AI's the best"-mentality. We actually have a lot of experience gathered from accidents and non-accidents how these systems should behave in a general sense, and "overriding pilot-input" has proven not to be a good idea. The discussion above was about details and UX.