The question is at what point flight software can be demonstrated to have a small fraction of the failure rate that humans have. We could certainly have softwa

The question is at what point flight software can be demonstrated to have a small fraction of the failure rate that humans have. We could certainly have software flying a plane tomorrow if we wanted to, at the cost of failure rates several orders of magnitude higher. Humans have the capacity look around, intuit what's going on and what doesn't make sense, and work around it. Software that can detect and understand what to do when 3 things that each occur only one in 1,000 flights, happen simultaneously while 3 out of 10 of the sensor networks are down for maintenance and rare weather patterns are occurring on the route. We have 100,000 flights per day to cover. Right now, we have lethal accidents roughly monthly, but we have contingency conditions that might plausibly cause a lethal accident a thousand times more often, conditions that are caught, troubleshooted, and successfully mitigated by a human.