I'm curious about this, and have been trying to find more information, but it's scarce and contradictory right now. Time points out that Airbuses have had problems in the past with the autopilot throwing the plane into a dive. They did have backups, but for some reason, they didn't work.

The system is intended to eliminate the possibility of electronic error: the flight computer, which is always monitoring the trio, can disregard one ADIRU if it begins relaying information that conflicts with the other two.

But that's not what happened when one of them went awry on Oct. 7 and began sending erroneous data spikes on the plane's angle of attack (AOA) — the angle between its wings and the air flowing over them — to the flight-control computer. "For some reason, the damn computer disregarded the healthy channels," says Hans Weber, an aviation expert who heads Tecop International, an aviation-consulting firm in San Diego. "Instead, it acted upon the information from the rogue channel."

Looking back on all the data, it can be clear to the human observer which channels were the healthy ones. But the "decisions" were made by a computer in real-time, programmed ahead of time by some other distant human removed from the actual environment. Whatever inference engine or heuristics were programmed in, they got confused by the chaos of the erroneous data, and "decided" that the rogue channel was the only healthy channel.