← back to the reports
MY-010 Airbus A330-203 · Air France/France 2009

Air France Flight 447 — A Stall Held to the Ocean in the Dark

Killed
228
Aircraft
Airbus A330-203
Operator
Air France
Status
Pilot

Summary

In the early hours of 1 June 2009, Air France Flight 447, an Airbus A330-203 flying overnight from Rio de Janeiro to Paris, fell into the equatorial Atlantic with 228 people aboard. None survived. The aircraft had been cruising normally at 35,000 feet when its three pitot tubes — the small forward-facing probes that measure airspeed — iced over inside a band of high-altitude convective weather. The airspeed readings became briefly unreliable, the autopilot and autothrust disconnected as designed, and control of a perfectly airworthy jet passed abruptly to two pilots who did not understand what was happening to it. Within about four and a half minutes the A330 had stalled and descended, nose high and wings roughly level, into the sea.

The aircraft was almost new and the icing event was transient: the probes cleared within about a minute, and the airframe never suffered any failure that would have prevented continued flight. The accident sequence was instead an unrecognized aerodynamic stall. The pilot flying, the most junior of the three crew, made and then sustained nose-up control inputs that pulled the aircraft into a steep climb, bled off its speed, and held it stalled all the way down. A stall warning sounded almost continuously, yet the crew never identified the condition or applied the standard recovery — nose down, reduce angle of attack. The captain, resting at the moment the trouble began, returned to the cockpit too late to diagnose the situation before impact.

France's Bureau d'Enquêtes et d'Analyses pour la sécurité de l'aviation civile (BEA) led the investigation under ICAO Annex 13. Its task was extraordinary: the wreckage and recorders lay nearly 4,000 metres deep, and it took almost two years and four search campaigns to locate them. The BEA published its final report on 5 July 2012. The report's analysis centred on the crew: the loss of airspeed information triggered a chain of inappropriate manual inputs and a failure to recognize the stall, set against deeper deficiencies in high-altitude manual-flying training, crew coordination, and the ergonomics of the warnings the crew received.

The legal aftermath ran far longer than the technical one. A 2022–2023 criminal trial in Paris ended in March 2023 with the acquittal of both Air France and Airbus. The victims' families appealed. On 21 May 2026 the Paris Court of Appeal reversed that outcome, convicting both companies of corporate manslaughter (homicides involontaires) and imposing the maximum corporate fine of 225,000 euros on each. Both companies announced they would appeal to the Court of Cassation. As of mid-2026 the case remains, in legal terms, open.

Timeline

31 May 2009, 22:29 UTC
Departure
Flight 447 leaves Rio de Janeiro Galeão for Paris Charles de Gaulle as a routine overnight crossing, an Airbus A330-203 registered F-GZCP with 216 passengers and 12 crew.
1 June 2009, ~01:35 UTC
Last contact
The aircraft makes its final routine radio contact and passes beyond radar coverage over the Atlantic, approaching the Intertropical Convergence Zone and its towering storms.
~02:02 UTC
Captain leaves the cockpit
Captain Marc Dubois goes to rest, leaving the two first officers in control; Pierre-Cédric Bonin is the pilot flying.
02:10:05 UTC
Autopilot disconnects
Ice crystals block the pitot tubes; airspeed readings become invalid and the autopilot and autothrust disengage. The aircraft is handed to the crew at altitude in turbulence.
02:10:05–02:10:50 UTC
Nose-up and a climb
Bonin makes nose-up and roll inputs; the A330 zoom-climbs from 35,000 feet, losing speed. The stall warning sounds.
~02:11:10 UTC
Apex
The aircraft reaches about 38,000 feet, its highest point, with airspeed decaying badly.
~02:11:40 UTC
Stalled
The angle of attack passes 40 degrees; the aircraft is fully stalled and begins a high rate of descent while remaining nose-high.
02:11:40–02:14:28 UTC
The descent
The jet falls at roughly 10,000 feet per minute for about three and a half minutes, stalled throughout; the crew never identifies the stall or applies nose-down recovery. The captain returns but cannot resolve it in time.
02:14:28 UTC
Impact
The aircraft strikes the ocean belly-first, wings near level, at high vertical speed. All 228 aboard are killed.
1–2 May 2011
Recorders found
After four search campaigns, the flight data recorder and cockpit voice recorder are recovered from about 3,900 metres depth and read out successfully.
5 July 2012
BEA final report
The BEA publishes its conclusions, centring on the crew's response, manual-flying training, and warning ergonomics.
March 2023
Acquittals
A Paris court acquits Air France and Airbus of involuntary manslaughter; families appeal.
21 May 2026
Convictions on appeal
The Paris Court of Appeal reverses the acquittals, convicting both companies and fining each 225,000 euros; both announce appeals to the Court of Cassation.

The Aircraft and the Crossing

The A330-203 is a twin-engine, fly-by-wire wide-body, and F-GZCP was a sound example of it: delivered in 2005, it had no relevant defect on the night of the crossing. The flight was the ordinary kind of long-haul work that thousands of crews fly without incident — a night sector across the South Atlantic with a three-pilot complement so that the captain could take a rest break in cruise. Nothing about the load, the fuel, the weather briefing, or the aircraft's condition marked the flight as unusual.

The one known weakness was external and small. The aircraft's pitot probes were an earlier Thales model that operators and manufacturers already knew could be susceptible to icing at high altitude in certain conditions. A replacement programme was under way across the A330/A340 fleet, but the modification had not yet reached F-GZCP, and the regulator had not made it mandatory. On the route the flight took, through the convective weather of the Intertropical Convergence Zone, the probes encountered ice crystals capable of obstructing them. The result was not catastrophic in itself: a loss of valid airspeed data for under a minute, of a kind the type's procedures were designed to ride out.

By design, when the airspeed inputs disagreed, the flight-control system dropped the autopilot and autothrust and reverted to a less-protected control law. The aircraft remained fully flyable. The required response was simple and trained: hold attitude and thrust, fly the aircraft straight and level, and wait for the airspeed information to return — which it did. The transient was a problem the crew had only to not make worse.

The Stall Nobody Named

When the autopilot disconnected at 02:10:05, Bonin took manual control and almost immediately pulled the nose up. The reason has never been fully explained; the BEA noted that startle, the turbulence, and confusing indications all played a part. Whatever the trigger, the input was the wrong one. The A330 climbed steeply away from its cruise altitude, trading speed for height it did not need, and the airspeed — once the probes cleared and valid readings returned — fell toward the stall.

The stall warning activated and then sounded almost continuously. Yet a perverse feature of the situation worked against recognition: when the angle of attack became extreme and the airspeed reading dropped below a threshold, the stall warning logic stopped, then resumed when a nose-down input briefly raised the indicated speed. A pilot pushing the nose down — the correct action — could thus be met with the stall warning starting up again, an apparent punishment for doing the right thing. The crew never assembled the cues into the single correct diagnosis: the aircraft is stalled.

For three and a half minutes the A330 descended at roughly 10,000 feet per minute, nose pitched up around 15 degrees, wings broadly level, engines spooled up — the classic, counterintuitive signature of a deep stall in which the wing is no longer flying. The pilots spoke of not understanding what was happening; the recorders captured confusion over whether the aircraft was climbing or descending and which displays to trust. Captain Dubois returned to the cockpit partway down but, arriving into an unexplained emergency with the ground rushing up, could not reconstruct the sequence in the seconds available. The aircraft struck the water belly-first, intact and stalled, at 02:14:28.

The Search and the Verdict

Recovering the truth from the seabed was itself a multi-year investigation. The first floating debris and bodies were found within days, but the main wreckage and the recorders lay on rough terrain nearly 4,000 metres down. Four search campaigns over almost two years finally located the debris field in 2011; the flight data recorder and cockpit voice recorder were raised on 1 and 2 May and, remarkably, yielded full data. Only then could the BEA reconstruct the cockpit second by second.

The BEA's final report of 5 July 2012 traced the accident to the crew's handling of a recoverable situation. The temporary loss of airspeed information caused the autopilot to disengage; the pilot flying made inappropriate nose-up inputs that destabilized the flight path and led to a climb and loss of speed; and the crew failed to recognize the resulting stall and to apply the recovery. Behind these immediate actions the report identified systemic factors: pilots were not adequately trained or practised in manual flying at high altitude or in handling unreliable-airspeed events; crew coordination broke down under startle; and the design and logic of the indications — including the intermittent stall warning — did not help the crew build an accurate picture. The report issued recommendations on training, on the certification of warnings and human-machine interaction, and on flight-recorder and underwater-locator improvements to speed future recoveries.

Crucially, the BEA's role was to determine cause and prevent recurrence, not to assign blame; under Annex 13 its findings carried no legal liability. The courts handled that separately and reached a different emphasis. Where the technical report led with crew response, the 2026 appeal court found Air France and Airbus criminally at fault for organizational failings — inadequate pilot training for sensor failures and shortcomings around the known pitot vulnerability — framing the crash as a foreseeable risk that had not been addressed in time.

The Five Factors

01
A small sensor fault handed off a whole aircraft
Iced pitot tubes caused only a brief loss of airspeed data, but that was enough to disconnect the automation and transfer control of a high-altitude jet to the crew without warning. Systems that degrade gracefully on the engineering side can still hand humans an abrupt, high-workload problem; the handoff itself is a hazard to be designed for.
02
Manual flying at altitude was an untrained edge
The crew rarely hand-flew at cruise altitude, where the margins between overspeed and stall are narrow and aircraft response is unfamiliar. A skill that is automated away in normal operation atrophies precisely where it is most needed in an emergency. Recurrent training must cover the rare manual case, not just the routine automated one.
03
The stall was never recognized
Despite a near-continuous stall warning and a sustained high rate of descent, the crew did not identify an aerodynamic stall or apply the standard nose-down recovery. Recognition, not just recovery technique, is the first and decisive step; training and instrumentation must make the stalled state unambiguous.
04
Warnings that fought the correct action
The stall-warning logic could fall silent at extreme angles of attack and restart when a nose-down input raised the indicated speed, so doing the right thing appeared to reawaken the alarm. An alert that contradicts the recovery it should prompt undermines the crew at the worst moment. Warning systems must remain coherent across the whole flight envelope, including its corners.
05
Startle and crew coordination
A sudden, ambiguous failure at night produced confusion over basic facts — climbing or descending, who was flying, which display to trust. The captain, resting at the trigger moment, could not re-enter the loop in time. Crew-resource-management training must rehearse the startle scenario and clear role-handover under stress, not only calm-cockpit decision-making.

Aftermath

AF447 became the central modern case study in the limits of cockpit automation and the erosion of manual flying skill. Its most concrete legacy lay in training and procedures: regulators and operators reworked stall-recovery training to emphasize reducing angle of attack over preserving altitude, expanded practice in unreliable-airspeed and high-altitude upset scenarios, and pressed the pitot-probe replacement and certification questions the report had raised. The accident also drove improvements in flight-recorder requirements and underwater locator beacons, intended to prevent another two-year hunt for the truth.

The human and legal reckoning was slower and more contested. The deep-sea search that recovered the recorders also recovered most of the victims, allowing identification and burial that the families had long awaited. The criminal process stretched across seventeen years: an investigation, an initial acquittal of Air France and Airbus in March 2023, and the reversal on 21 May 2026 when the Paris Court of Appeal convicted both companies of corporate manslaughter and imposed the maximum 225,000-euro fine on each — a sum widely criticized as token against 228 deaths. With both companies pursuing a further appeal to the Court of Cassation, the legal arc of the case had still not fully closed.

Lessons

  1. Treat the moment automation disengages as a designed-for emergency, not a neutral handover; the abrupt transfer of control to the crew is itself a hazard that training and interface design must address.
  2. Recognize the stall before trying to recover it. The first failure of AF447 was diagnostic, so make the stalled state unmistakable through training and instrumentation, and teach nose-down recovery over altitude preservation.
  3. Keep warning systems coherent across the entire flight envelope; an alert that goes quiet, or that restarts when the pilot finally does the right thing, can defeat the very recovery it exists to prompt.
  4. Preserve and practise manual flying at high altitude. A skill automated out of daily use will not be there in the rare crisis that demands it.
  5. Address known component vulnerabilities and the human procedures around them before they combine; the pitot susceptibility was understood and the fix already in progress, but neither the hardware change nor the training had reached this aircraft and crew in time.

References