4 ms·

Would a sufficiently good track record convince you? Say, a million flights with 80% lower accident rates than human pilots? A strict utilitarian should only c

by tlb 8y ago

Would a sufficiently good track record convince you? Say, a million flights with 80% lower accident rates than human pilots? A strict utilitarian should only care about accident rates. But as Greene [0] and others show, most people don't make moral judgements in a strictly utilitarian way. Thus, we judge homicide as worse than choosing not to save someone from death, even though in both cases someone dies as a result of a decision. One of the lessons from probing moral calculus is that nobody can really articulate reasons for their actual moral decisions. People can articulate utilitarianism, but not what people actually do. But, if you can articulate why you might prefer a human pilot over an autopilot, even given sufficient statistics that the autopilot is safer, it'd be interesting to hear. [0] http://www.joshua-greene.net/ http://www.joshua-greene.net/

benmowa 8y ago
what about the one-in-a-million events that pilots continuously train for? Is one million flights a large enough sample?
- tlb 8y ago
  The current helicopter accident rate for emergency medical service is 3.9 / 100,000 flight hours [0], so a million flights is probably large enough. But feel free to answer the question based on some larger number that would convince you it's statistically safer. [0] https://www.aea.net/events/rotorcraft/files/US_Rotorcraft_Accident_Data_And_Statistics.pdf https://www.aea.net/events/rotorcraft/files/US_Rotorcraft_Ac...
oliveshell 8y ago
I think I could be convinced, eventually— but it would take quite a lot of evidence. A consistently lower accident rate than human pilots would be a good start, but it would be even better if those flights were conducted in the most punishing weather conditions possible and in all sorts of dangerous scenarios where pilot error is likely. One of the biggest factors is that it is terrifying to imagine falling out of the sky in an aircraft under the command of a software-based pilot with which I cannot communicate and that I cannot see doing everything it possibly can to avert disaster. It may not be a rational fear, but we humans are not rational creatures. It’s an issue of trust, which the folks interviewed in the article do acknowledge, to their credit. We can instinctively judge whether or not to trust another human being, because we have a lifetime of practice interacting with them. The burden of proof of safety for an uncommunicative mechanical system is much higher, and lots of testing (that I can observe on video) will go some way towards to building that trust. Furthermore, things like avoiding thunderstorms, the motion of which we still cannot predict with computers (turbulence being an unsolved problem in physics), could be very tricky for an automated system to handle. As autonomous technology progresses and we accumulate data, I might begin to feel differently. But for right now, without knowing all of the details of the system, I have no way of knowing how it is reasoning and perceiving the situation (I’m very familiar with the human consciousness, as I possess one myself) and this leads to a fundamental sort of uneasiness in my lizard brain. I’m still bullish on technology, though, and this is an area in which I would be enthusiastic to be proven wrong.

by tlb 8y ago

benmowa 8y ago

what about the one-in-a-million events that pilots continuously train for? Is one million flights a large enough sample?

tlb 8y ago
The current helicopter accident rate for emergency medical service is 3.9 / 100,000 flight hours [0], so a million flights is probably large enough. But feel free to answer the question based on some larger number that would convince you it's statistically safer. [0] https://www.aea.net/events/rotorcraft/files/US_Rotorcraft_Accident_Data_And_Statistics.pdf https://www.aea.net/events/rotorcraft/files/US_Rotorcraft_Ac...

oliveshell 8y ago

I think I could be convinced, eventually— but it would take quite a lot of evidence. A consistently lower accident rate than human pilots would be a good start, but it would be even better if those flights were conducted in the most punishing weather conditions possible and in all sorts of dangerous scenarios where pilot error is likely. One of the biggest factors is that it is terrifying to imagine falling out of the sky in an aircraft under the command of a software-based pilot with which I cannot communicate and that I cannot see doing everything it possibly can to avert disaster. It may not be a rational fear, but we humans are not rational creatures. It’s an issue of trust, which the folks interviewed in the article do acknowledge, to their credit. We can instinctively judge whether or not to trust another human being, because we have a lifetime of practice interacting with them. The burden of proof of safety for an uncommunicative mechanical system is much higher, and lots of testing (that I can observe on video) will go some way towards to building that trust. Furthermore, things like avoiding thunderstorms, the motion of which we still cannot predict with computers (turbulence being an unsolved problem in physics), could be very tricky for an automated system to handle. As autonomous technology progresses and we accumulate data, I might begin to feel differently. But for right now, without knowing all of the details of the system, I have no way of knowing how it is reasoning and perceiving the situation (I’m very familiar with the human consciousness, as I possess one myself) and this leads to a fundamental sort of uneasiness in my lizard brain. I’m still bullish on technology, though, and this is an area in which I would be enthusiastic to be proven wrong.