Artificial Intelligence

Forty-five years of airline safety analysis

David Learmount Uncategorized , , , , , , , , , , , , , , ,

For more than 50 years Flight International/FlightGlobal has been publishing a review of world airline safety performance annually, but for the last 45 of them I have been the compiler and author. The most recent review – containing analysis and a list of all the fatal and many of the significant non-fatal airline accidents in 2025 – can be found here on FlightGlobal’s website.

My first safety review for Flight International magazine covered the year 1980. A few years later, in 1985, I reported on the worst year in aviation history in terms of the number of passengers and crew killed: 2,230 deaths in 41 fatal accidents. Last year the figures were respectively 420 and 11, despite the fact that the number of airline passengers carried is now about four times what it was in 1980, and the number of flights has increased about three-fold.

So where are we today in terms of airline safety performance? Obviously massively better than it was in the 1980s, but by comparison with the annual figures in the most recent decade, 2025’s numbers were slightly lower than the average for fatal accidents, and rather higher than the average (276) for the number of annual fatalities.

FlightGlobal/Flight International explains: “The principal reason for the relatively high casualty numbers – a total of 420 for the year – was that more than half of them died in a single, catastrophic crash involving an Air India Boeing 787-8 after departing Ahmedabad International airport on 12 June.”

Flight concedes, however, that “2025 was an unremarkable year – statistically – for fatal accidents.” And those most recent ten-year figures have been remaining fairly constant around a historic best-ever level, so any useful safety review needs to ask what the industry is getting right, as well as studying the mistakes.

The article debates in detail whether the action by one of the Air India pilots of closing both the engines’ fuel control switches seconds after take-off from Ahmedabad was an extraordinary mistake or a deliberate act (as does my previous piece in this blog). The Indian Air Accident Investigation Bureau confirms the pilot’s act of shutting off the fuel, and within about a year its report should be able to provide a verdict on why he did it.

Summing up safety in 2025, FlightGlobal said: “Apart from the Air India loss, almost all the ­accidents can be considered to have been “traditional” in nature. That is, they were caused by exposure to ordinary threats such as bad weather, pilots taking avoidable risks, or errors of omission or commission, bird-strikes, turbulence encounters and maintenance shortcomings.”

One of the primary reasons air travel is so much safer than it used to be is that the aircraft and engines are better engineered than they were forty years ago, and smart avionics give the crews more information presented more intuitively. There is a parallel between engineering advances in the avation and automotive industries. Anyone who owns a new car today recognises that its reliability and its technology is much better than cars built in the 1980s.

The theoretical downside of “smarter” aircraft is that they are highly computerised and thus more complex, but in practice that concern is not borne out because the systems are so reliable they rarely go wrong, and they are self-monitoring so the pilots are kept informed of systems health.

The human factors worry is that the sheer reliability means pilots may be lulled into a comfortable, non-critical mindset, so when something does go wrong they are startled and may react inappropriately. Again – theoretically – there is more that can go wrong because aircraft may still be traditional mechanical machines, but they are overlaid with software-driven sensors and computerised flight management systems. On the rare occasions when these go wrong, however, the crew may be confronted with a problem that has never presented before, so there is no checklist to deal with it.

For that reason, today’s crews in their basic training are still confronted with traditional problems like engine failure, but their advanced and continuation training is designed to inculcate a resilient mindset, based on flight priorities, paticularly when something unidentified has gone wrong: 1. Aviate, 2. Navigate, 3. Communicate.

Aviate: is the aircraft flying at the appropriate speed, height and attitude? Navigate: what is the aircraft’s position, is it heading in the direction it should be, and what is the fuel state? Communicate: report your situation to ATC, then work with other crew to deternine the best course of action to deal with the problem, and tell the cabin crew chief what is going on. Finally, while dealing with the problem, revisit your priorities over and over again: are you aviating right, are you navigating right, are you communicating what people need to know?

On this theme of training pilots to interface with the aircraft’s systems, the air transport industry – like all others – has to prepare to make good use of the next level of information technology: artificial intelligence (AI). In a thoughtful paper entitled Artificial Intelligence in Aviation, IFALPA (International Federation of Airline Pilot Associations) warns the industry to be ready to use AI with care to support the piloting task. It advises: “The role of AI in the operation of a flight should always be to support the humans in the system”, adding: “For this to be effective, whatever the intended capability of an AI system, it should only present options to a pilot, never a fixed outcome. There should also be transparency to the pilot as to how these options have been selected, and the level of confidence associated with them.”

2025 may have played out with a relatively small number of low-tech airline accidents, but we have to be ready for something different, and hopefully even better.

Boeing 737: the beginning of the end.

David Learmount Uncategorized , , , , , , , , , , , , , , , , , , , , , , , , , , , , , ,

It hasn’t surprised anyone in the industry to hear rumours – via the pages of the Wall Street Journal – that Boeing is working on the design of a new narrowbody jet, because it’s what everyone – including Boeing – knows the manufacturer should have done instead of launching the 737 Max.

The confidence in that statement is completely un-influenced by hindsight.

Now the Max has been purged of the ghastly design mistake that was MCAS (manoeuvring characteristics augmentation system), and Boeing has radically overhauled its corporate safety culture under a new leader – former Rockwell Collins engineer CEO Kelly Ortberg – the 737 series can once again trade on the lazy confidence that comes from the fact that – with all its faults and its antique technology – it’s a known quantity.

As a result the 737 is selling well, but nothing like as well as its competitor the Airbus A320 series.

The first 737-100 entered service in 1968, initially to fly the routes that the larger 727 series trijet was too big for. Its basic control technology was – and still is – just-post-war, except that in the latest versions the power-assisted cranks and pulleys are overlaid with electronic flight instrument systems and flight management computers to the extent that the pilots could almost believe the aircraft is fly-by-wire. They know, however, that they themselves are the flight envelope protection.

The industry needs a new-technology narrowbody competitor to the A320, and if Boeing doesn’t supply it, perhaps a development of China’s Comac C919, Russia’s Sukhoi Superjet 100, or a new product from Brazil’s Embraer will fill the gap.

Boeing’s first fly-by-wire airliner was its highly successful 777 widebody, which entered service in 1995 with virtually no birthing pains.

In 2011 it launched the 787 Dreamliner series, also highly commercially successful, but suffering from multiple early problems, some of them still being worked on.

Right now Boeing is struggling to re-launch the 777 as the 777X. The fact that it had a planned 2019 in-service date, but now its launch customer – Lufthansa – will not receive it until 2026, suggests how difficult a task bringing an entirely new narrowbody (the 797?) to service readiness may yet be.

The challenge is always to deliver a safe, trouble-free product, but the staggeringly advanced, fully-integrated electronic technology by which the aircraft and all its systems will be managed and controlled, plus the fact that there must be fallback systems that the pilots can access easily if it all goes wrong, mean its service entry will not be quick.

Look beyond this to the fact that the new systems will inevitably employ artificial intelligence, which makes passengers – and even engineers – nervous when it comes to managing safety-critical systems, and the size of the challenge becomes clear.

So the venerable 737 series will be with us for many years yet.

AI suggests why the Philadelphia medevac crash happened

David Learmount Uncategorized , , , , , , , , , , ,

Can artificial intelligence (AI) provide the factors behind aviation accidents? Maybe we should find out, because people can suddenly believe they are experts as a result of using it.

A reader contacted me on February 1st with the answer that an AI app provided when he asked it what caused the 31 January fatal Learjet air ambulance crash in Philadelphia.

Before discussing the AI’s verdict, here is what we actually know about the short flight: the Learjet 55 took off in the evening dusk (18:06:10 local time) from Philadelphia Northeast Airport runway 24, bound for Springfield, Missouri, with a stretchered patient and five other people on board. The temperature and dew point were both 9deg, the cloud-base was at 400-500ft, light wind and reasonable visibility.

When airborne the aircraft was told to turn right onto heading 290deg, and the pilot received an instruction from Tower to change frequency to Philadelphia Departure Control. He read the new frequency back correctly, and bid the Tower controller a good day.

According to ADS-B data, the Learjet had climbed rapidly to a maximum 1,650ft by about a minute after take-off (18:06:56) , then the aircraft entered a steep, uninterrupted descent to impact with the ground. The impact point was in a suburban area about 2.5nm from the airport, close to the runway extended centreline. The pilot never did make contact with Departure Control, and broadcast messages addressed to the flight by Departure did not receive a response.

According to initial reports by Philadelphia police, no-one on board survived the Learjet crash, one person on the ground was killed and 19 were injured, That is my summary of the basic known facts of what happened.

Meanwhile my reader who asked AI to provide him with an explanation for the crash told me he had, in his question, given the AI app (which he didn’t name) all the facts known at that point.

The executive summary part of the AI answer said this: “Preliminary data from ADS-B tracking, witness reports, and aviation system analysis suggest that Learjet 55 XA-UCI suffered a catastrophic runaway trim event (nose-down), leading to an unrecoverable dive and high-speed impact.” It also supplied what I would describe as cogent arguments to back this verdict up, but no actual evidence for the alleged runaway trim or the electrical fault that it proposed was the reason for it. The whole proposal, however, was delivered in a decidedly confident style.

I decided to take a different approach to test AI on the same subject. Given what we know happened, I asked Chat GPT whether the pilot suffering spatial disorientation as a result of somatogravic illusion could be the explanation for the Learjet accident? ChatGPT’s response first explained what somatogravic illusion is, then responded that, yes, it could indeed be a plausible explanation, but advised me to wait for the National Transportation Safety Board’s report.

Somatogravic illusion is an illusion generated by the delicate human inner ear balance organs when they are subjected to acceleration, linear or rotational. For example, passengers seated in the cabin of an aircraft beginning its acceleration along the runway for take-off can feel that the whole aircraft has tilted slightly nose-up, especially if they are looking straight ahead. But a glance out of the cabin window during the take-off run will prove that no such upward tilt has taken place.

Pilots experience the same somatogravic effect during take-off that passengers do, but since they are looking ahead out the cockpit windscreen – and providing the external visibility is good – their powerful visual sense will overcome the misleading feedback from their balance sensors.

If, however, the acceleration continues after take-off and the crew lose sight of the outside world because of darkness or entering cloud, the misleading feedback from their balance sensors returns. And the natural reaction to believing the aircraft’s nose is higher than it should be is to push forward on the control column, pushing the nose down. The physical feeling that a nose-down push is demanded can entirely overcome the intellectual information presented by the pilots’ flight instruments, because the latter is artificial, unlike powerful instinctive feelings or sight of a natural external horizon.

The Learjet series has a reputation for sporty performance. Its take-off acceleration and rate of climb when airborne are impressive. And the point in this short flight where it all appears to have gone wrong happens to occur at the moment when the pilot is likely to have taken his eyes off the flight instruments for a moment to change the radio frequency. The latter may be just coincidence, however.

There is no data here that could be regarded as evidence about the reasons for the Philadelphia crash, but I do know that the runaway trim explanation is plausible, and so is the pilot spacial disorientation theory.

There could be other reasons, however, and I know well after 45 years in this business that listing “what if” explanations is a waste of time because there are too many. The truth will out, via actual evidence. These days it does not take long, because investigators now strive to provide periodic interim factual reports which signpost the emerging truth.

But full understanding – and thus the ability confidently to act to prevent repetition – only comes with the full facts.