The autopilot that handed back a plane nobody could fly

On the night of 31 May 2009, Air France Flight 447 left Rio de Janeiro for Paris. The Airbus A330 had 216 passengers and 12 crew on board. Around 2:02 the captain left the cockpit for a planned rest, leaving the two co-pilots at the controls.

At 2:10, ice crystals most likely blocked the plane's speed sensors. The speed readings became wrong, and the autopilot and automatic engine control switched off. The plane also moved into a mode with fewer automatic protections. The wrong speed readings lasted only about half a minute to a minute, depending on the display.

The pilot flying pulled the nose up. The plane climbed, lost speed and stalled, meaning its wings stopped giving enough lift. A stall warning sounded repeatedly. The captain came back about a minute and a half after the autopilot cut out, but by then the plane was already stalled. It hit the sea at 2:14, about four and a half minutes after the autopilot disconnected.

The wreck was found almost 4,000 metres down in April 2011, and the flight recorders were recovered soon after. In its final report in July 2012, the French investigators, the BEA, found that the crew did not link the lost speed readings to the right procedure and never recognised the stall. They pointed to a lack of training in flying by hand at high altitude, the shock of the sudden alarms, and a cockpit that did not show clearly what the computers had detected.

Automation flew the plane almost all the time, so the pilots rarely flew by hand at cruising height. The investigators said training did not make up for this. When the computers gave up, they handed the hardest moment of the flight to people who had the least recent practice at it, in the dark, with alarms going off.

This is a real risk with AI at work. If AI drafts every report, writes every formula and answers every customer, people stop practising those skills. The tool will still fail sometimes, and usually on the unusual, high-stakes case. Someone has to be able to take over calmly and understand what has gone wrong. That ability fades if it is never used.

The idea behind this caseLesson: Will AI take my job? →

1. 1Do some important tasks without AI on a regular schedule, so the skill stays fresh.
2. 2Write down what to do if the AI tool is down or clearly wrong, and practise it once in a while.
3. 3Make sure new staff learn the task by hand before they learn it with AI.
4. 4Ask AI tools to show their working, so you can follow along and spot when something is off.
5. 5Choose tools that warn you clearly when they are unsure, rather than failing silently.