In the early hours of June 1st 2009, Air France Flight 447 crashed into the Atlantic Ocean. Till the black boxes of AF447 were recovered in April 2011, the exact circumstances of the crash remained a mystery. The most widely accepted explanation for the disaster attributes a large part of the blame to human error when faced with a partial but not fatal systems failure. Yet a small but vocal faction blames the disaster and others like it on the increasingly automated nature of modern passenger airplanes.
This debate bears an uncanny resemblance to a similar debate as to the causes of the financial crisis – many commentators blame the persistently irrational nature of human judgement for the recurrence of financial crises. Others such as Amar Bhide blame the unwise deference to imperfect financial models over human judgement. In my opinion, both perspectives miss the true dynamic. These disasters are not driven by human error or systems error alone but by fatal flaws in the interaction between human intelligence and complex, near fully-automated systems.
In a recent article drawing upon the black box transcripts, Jeff Wise attributes the crash primarily to a “simple but persistent mistake on the part of one of the pilots”. According to Wise, the co-pilot reacted to the persistent stall warning by “pulling back on the stick, the exact opposite of what he must do to recover from the stall”.
But there are many hints that the story is nowhere near as simple. As Peter Garrison notes :
every pilot knows that to recover from a stall you must get the nose down. But because a fully developed stall in a large transport is considered highly unlikely, and because in IFR air traffic vertical separation, and therefore control of altitude, is important, transport pilots have not been trained to put the nose down when they hear the stall warning — which heralds, after all, not a fully developed stall, but merely an approaching one. Instead, they have been trained to increase power and to “fly out of the stall” without losing altitude. Perhaps that is what the pilot flying AF447 intended. But the airplane was already too deeply stalled, and at too high an altitude, to recover with power alone.
The patterns of the AF447 disaster are not unique. As Chris Sorensen observes, over 50 commercial aircrafts have crashed in “loss-of-control” accidents in the last five years, a trend for which there is no shortage of explanations:
Some argue that the sheer complexity of modern flight systems, though designed to improve safety and reliability, can overwhelm even the most experienced pilots when something actually goes wrong. Others say an increasing reliance on automated flight may be dulling pilots’ sense of flying a plane, leaving them ill-equipped to take over in an emergency. Still others question whether pilot-training programs have lagged behind the industry’s rapid technological advances.
But simply invoking terms such as “automation addiction” or blaming disasters on irrational behaviour during times of intense stress does not get at the crux of the issue.
People Make Poor Monitors for Computers
Airplane automation systems are not the first to discover the truth in the comment made by David Jenkins that “computers make great monitors for people, but people make poor monitors for computers.” As James Reason observes in his seminal book ‘Human Error’:
We have thus traced a progression from where the human is the prime mover and the computer the slave to one in which the roles are very largely reversed. For most of the time, the operator’s task is reduced to that of monitoring the system to ensure that it continues to function within normal limits. The advantages of such a system are obvious; the operator’s workload is substantially reduced, and the [system] performs tasks that the human can specify but cannot actually do. However, the main reason for the human operator’s continued presence is to use his still unique powers of knowledge-based reasoning to cope with system emergencies. And this is a task peculiarly ill-suited to the particular strengths and weaknesses of human cognition…..
most operator errors arise from a mismatch between the properties of the system as a whole and the characteristics of human information processing. System designers have unwittingly created a work situation in which many of the normally adaptive characteristics of human cognition (its natural heuristics and biases) are transformed into dangerous liabilities.
As Jeff Wise notes, it is impossible to stall an Airbus in most conditions. AF447 however went into a state known as ‘alternate law’ which most pilots have never experienced where the airplane could be stalled:
“You can’t stall the airplane in normal law,” says Godfrey Camilleri, a flight instructor who teaches Airbus 330 systems to US Airways pilots….But once the computer lost its airspeed data, it disconnected the autopilot and switched from normal law to “alternate law,” a regime with far fewer restrictions on what a pilot can do. “Once you’re in alternate law, you can stall the airplane,” Camilleri says….It’s quite possible that Bonin had never flown an airplane in alternate law, or understood its lack of restrictions. According to Camilleri, not one of US Airway’s 17 Airbus 330s has ever been in alternate law. Therefore, Bonin may have assumed that the stall warning was spurious because he didn’t realize that the plane could remove its own restrictions against stalling and, indeed, had done so.
This inability of the human operator to fill in the gaps in a near-fully automated system was identified by Lisanne Bainbridge as one of the ironies of automation which James Reason summarised:
the same designer who seeks to eliminate human beings still leaves the operator “to do the tasks which the designer cannot think how to automate” (Bainbridge,1987, p.272). In an automated plant, operators are required to monitor that the automatic system is functioning properly. But it is well known that even highly motivated operators cannot maintain effective vigilance for anything more than quite short periods; thus, they are demonstrably ill-suited to carry out this residual task of monitoring for rare, abnormal events. In order to aid them, designers need to provide automatic alarm signals. But who decides when these automatic alarms have failed or been switched off?
As Robert Charette notes, the same is true for airplane automation:
operators are increasingly left out of the loop, at least until something unexpected happens. Then the operators need to get involved quickly and flawlessly, says Raja Parasuraman, professor of psychology at George Mason University in Fairfax, Va., who has been studying the issue of increasingly reliable automation and how that affects human performance, and therefore overall system performance. ”There will always be a set of circumstances that was not expected, that the automation either was not designed to handle or other things that just cannot be predicted,” explains Parasuraman. So as system reliability approaches—but doesn’t quite reach—100 percent, ”the more difficult it is to detect the error and recover from it,” he says…..In many ways, operators are being asked to be omniscient systems administrators who are able to jump into the middle of a situation that a complex automated system can’t or wasn’t designed to handle, quickly diagnose the problem, and then find a satisfactory and safe solution.
Stored Routines Are Not Effective in Rare Situations
As James Reason puts it:
the main reason why humans are retained in systems that are primarily controlled by intelligent computers is to handle ‘non-design’ emergencies. In short, operators are there because system designers cannot foresee all possible scenarios of failure and hence are not able to provide automatic safety devices for every contingency. In addition to their cosmetic value, human beings owe their inclusion in hazardous systems to their unique, knowledge-based ability to carry out ‘on-line’ problem solving in novel situations. Ironically, and notwithstanding the Apollo 13 astronauts and others demonstrating inspired improvisation, they are not especially good at it; at least not in the conditions that usually prevail during systems emergencies. One reason for this is that stressed human beings are strongly disposed to employ the effortless, parallel, preprogrammed operations of highly specialised, low-level processors and their associated heuristics. These stored routines are shaped by personal history and reflect the recurring patterns of past experience……
Why do we have operators in complex systems? To cope with emergencies. What will they actually use to deal with these problems? Stored routines based on previous interactions with a specific environment. What, for the most part, is their experience within the control room? Monitoring and occasionally tweaking the plant while it performs within safe operating limits. So how can they perform adequately when they are called upon to reenter the control loop? The evidence is that this task has become so alien and the system so complex that, on a significant number of occasions, they perform badly.
Wise again identifies this problem in the case of AF447:
While Bonin’s behavior is irrational, it is not inexplicable. Intense psychological stress tends to shut down the part of the brain responsible for innovative, creative thought. Instead, we tend to revert to the familiar and the well-rehearsed. Though pilots are required to practice hand-flying their aircraft during all phases of flight as part of recurrent training, in their daily routine they do most of their hand-flying at low altitude—while taking off, landing, and maneuvering. It’s not surprising, then, that amid the frightening disorientation of the thunderstorm, Bonin reverted to flying the plane as if it had been close to the ground, even though this response was totally ill-suited to the situation.
Deskilling From Automation
As James Reason observes:
Manual control is a highly skilled activity, and skills need to be practised continuously in order to maintain them. Yet an automatic control system that fails only rarely denies operators the opportunity for practising these basic control skills. One of the consequences of automation, therefore, is that operators become de-skilled in precisely those activities that justify their marginalised existence. But when manual takeover is necessary something has usually gone wrong; this means that operators need to be more rather than less skilled in order to cope with these atypical conditions. Duncan (1987, p. 266) makes the same point: “The more reliable the plant, the less opportunity there will be for the operator to practise direct intervention, and the more difficult will be the demands of the remaining tasks requiring operator intervention.”
Opacity and Too Much Information of Uncertain Reliability
Wise captures this problem and its interaction with a human who has very little experience in managing the crisis scenario:
Over the decades, airliners have been built with increasingly automated flight-control functions. These have the potential to remove a great deal of uncertainty and danger from aviation. But they also remove important information from the attention of the flight crew. While the airplane’s avionics track crucial parameters such as location, speed, and heading, the human beings can pay attention to something else. But when trouble suddenly springs up and the computer decides that it can no longer cope—on a dark night, perhaps, in turbulence, far from land—the humans might find themselves with a very incomplete notion of what’s going on. They’ll wonder: What instruments are reliable, and which can’t be trusted? What’s the most pressing threat? What’s going on? Unfortunately, the vast majority of pilots will have little experience in finding the answers.
A similar scenario occurred in the case of the Qantas-owned A380 which took off from Singapore in November 2010:
Shortly after takeoff from Singapore, one of the hulking A380’s four engines exploded and sent pieces of the engine cowling raining down on an Indonesian island. The blast also damaged several of the A380’s key systems, causing the unsuspecting flight crew to be bombarded with no less than 54 different warnings and error messages—so many that co-pilot Matt Hicks later said that, at one point, he held his thumb over a button that muted the cascade of audible alarms, which threatened to distract Capt. Richard De Crespigny and the rest of the feverishly working flight crew. Luckily for passengers, Qantas Flight 32 had an extra two pilots in the cockpit as part of a training exercise, all of whom pitched in to complete the nearly 60 checklists required to troubleshoot the various systems. The wounded plane limped back to Singapore Changi Airport, where it made an emergency landing.
Again James Reason captures the essence of the problem:
One of the consequences of the developments outlined above is that complex, tightly-coupled and highly defended systems have become increasingly opaque to the people who manage, maintain and operate them. This opacity has two aspects: not knowing what is happening and not understanding what the system can do. As we have seen, automation has wrought a fundamental change in the roles people play within certain high-risk technologies. Instead of having ‘hands on’ contact with the process, people have been promoted “to higher-level supervisory tasks and to long-term maintenance and planning tasks” (Rasmussen, 1988). In all cases, these are far removed from the immediate processing. What direct information they have is filtered through the computer-based interface. And, as many accidents have demonstrated, they often cannot find what they need to know while, at the same time, being deluged with information they do not want nor know how to interpret.
Absence of Intuitive Feedback
Among others, Hubert and Stuart Dreyfus have shown that human expertise relies on an intuitive and tacit understanding of the situation rather than a rule-bound and algorithmic understanding. The development of intuitive expertise depends upon the availability of clear and intuitive feedback which complex, automated systems are often unable to provide.
In AF447, when the co-pilot did push forward on the stick (the “correct” response), the behaviour of the stall warning was exactly the opposite of what he would have intuitively expected:
At one point the pilot briefly pushed the stick forward. Then, in a grotesque miscue unforeseen by the designers of the fly-by-wire software, the stall warning, which had been silenced, as designed, by very low indicated airspeed, came to life. The pilot, probably inferring that whatever he had just done must have been wrong, returned the stick to its climb position and kept it there for the remainder of the flight.
Absence of feedback prevents effective learning but the wrong feedback can have catastrophic consequences.
The Fallacy of Defence in Depth
In complex automated systems, the redundancies and safeguards built into the system also contribute to its opacity. By protecting system performance against single faults, redundancies allow the latent buildup of multiple faults. Jens Rasmussen called this ‘the fallacy of defence in depth’ which James Reason elaborates upon:
the system very often does not respond actively to single faults. Consequently, many errors and faults made by the staff and maintenance personnel do not directly reveal themselves by functional response from the system. Humans can operate with an extremely high level of reliability in a dynamic environment when slips and mistakes have immediately visible effects and can be corrected……Violation of safety preconditions during work on the system will probably not result in an immediate functional response, and latent effects of erroneous acts can therefore be left in the system. When such errors are allowed to be present in a system over a longer period of time, the probability of coincidence of the multiple faults necessary for release of an accident is drastically increased. Analyses of major accidents typically show that the basic safety of the system has eroded due to latent errors.
This is exactly what occurred on Malaysia Airlines Flight 124 in August 2005:
The fault-tolerant ADIRU was designed to operate with a failed accelerometer (it has six). The redundant design of the ADIRU also meant that it wasn’t mandatory to replace the unit when an accelerometer failed. However, when the second accelerometer failed, a latent software anomaly allowed inputs from the first faulty accelerometer to be used, resulting in the erroneous feed of acceleration information into the flight control systems. The anomaly, which lay hidden for a decade, wasn’t found in testing because the ADIRU’s designers had never considered that such an event might occur.
Again, defence-in-depth systems are uniquely unsuited to human expertise as Gary Klein notes:
In a massively defended system, if an accident sneaks through all the defenses, the operators will find it far more difficult to diagnose and correct it. That is because they must deal with all of the defenses, along with the accident itself…..A unit designed to reduce small errors helped to create a large one.
Two Approaches to Airplane Automation: Airbus and Boeing
Although both Airbus and Boeing have adopted the fly-by-wire technology, there are fundamental differences in their respective approaches. Whereas Boeing’s system enforces soft limits that can be overridden at the discretion of the pilot, Airbus’ fly-by-wire system has built-in hard limits that cannot be overridden completely at the pilot’s discretion.
As Simon Calder notes, pilots have raised concerns in the past about Airbus‘ systems being “overly sophisticated” as opposed to Boeing’s “rudimentary but robust” system. But this does not imply that the Airbus approach is inferior. It is instructive to analyse Airbus’ response to pilot demands for a manual override switch that allows the pilot to take complete control:
If we have a button, then the pilot has to be trained on how to use the button, and there are no supporting data on which to base procedures or training…..The hard control limits in the Airbus design provide a consistent “feel” for the aircraft, from the 120-passenger A319 to the 350-passenger A340. That consistency itself builds proficiency and confidence……You don’t need engineering test pilot skills to fly this airplane.
David Evans captures the essence of this philosophy as aimed at minimising the “potential for human error, to keep average pilots within the limits of their average training and skills”.
It is easy to criticise Airbus‘ approach but the hard constraints clearly demand less from the pilot. In the hands of an expert pilot, Boeing’s system may outperform. But if the pilot is a novice, Airbus’ system almost certainly delivers superior results. Moreover, as I discussed earlier in the post, the transition to an almost fully automated system by itself reduces the probability that the human operator can achieve intuitive expertise. In other words, the transition to near-autonomous systems creates a pool of human operators that appear to frequently commit “irrational” errors and is therefore almost impossible to reverse.
* * *
People Make Poor Monitors for Some Financial Models
In earlier post, I analysed Amar Bhide’s argument that a significant causal agent in the financial crisis was the replacement of discretion with models in many areas of finance – for example, banks’ mortgage lending decisions. In his excellent book, ‘A Call for Judgement’, he expands on this argument and amongst other technologies, lays some of the blame for this over-mechanisation of finance on the ubiquitous Black-Scholes-Merton (BSM) formula. Although I agree with much of his book, this thesis is too simplistic.
There is no doubt that BSM has many limitations – amongst the most severe being the assumption of continuous asset price movements, a known and flat volatility surface, and an asset price distribution free of fat tails. But the systemic impact of all these limitations is grossly overstated:
- BSM and similar models have never been used as “valuation” methods on a large scale in derivatives markets but as a tool which tries to back out an implied volatility and generate useful hedge ratios by taking market prices for options as a given. In other words, volatility plays the role of the “wrong number in the wrong formula to get the right price”.
- When “simple” BSM-like models are used to price more exotic derivatives, they have a modest role to play. As Emanuel Derman puts it, practitioners use models as “interpolating formulas that take you from known prices of liquid securities to the unknown values of illiquid securities”.
Nevertheless, this does not imply that financial modelling choices have no role to play in determining system resilience. But the role was more subtle and had to do less with the imperfections of the models themselves as with the imperfections of how complex models used to price complex products could be used by human traders.
Since the discovery of the volatility smile, traders have known that the interpolation process to price exotic options requires something more than a simple BSM model. One would assume that traders would want to use a model that was accurate and comprehensive as possible. But this has rarely been the case. Supposedly inferior local volatility models still flourish and even in some of the most complex domains of exotic derivatives, models are still chosen based on their intuitive similarities to a BSM-like approach where the free parameters can be thought of as volatility or correlation e.g. The Libor Market Model.
The choice of intuitive understanding over model accuracy is not unwarranted. As all market practitioners know, there is no such thing as a perfect derivatives pricing model. Paul Wilmott hit the nail on the head when he observed that *“the many improvements on Black-Scholes are rarely improvements, the best that can be said for many of them is that they are just better at hiding their faults. Black-Scholes also has its faults, but at least you can see them”.
However, as markets have evolved, maintaining this balance between intuitive understanding and accuracy has become increasingly difficult:
- Intuitive yet imperfect models require experienced and expert traders. Scaling up trading volumes of exotic derivatives however requires that pricing and trading systems be pushed out to novice traders as well as non-specialists such as salespeople.
- With the increased complexity of derivative products, preserving an intuitive yet sufficiently accurate model becomes an almost impossible task.
- Product complexity combined with the inevitable discretion available to traders when they use simpler models presents significant control challenges and an increased potential for fraud.
In this manner, the same paradoxical evolution that have been observed in nuclear plants and airplane automation is now being experienced in finance. The need to scale up and accommodate complex products necessitates the introduction of complex, unintuitive models in combination with which human intuitive expertise is unable to add any value. In such a system, a novice is often as good as a more experienced operator. The ability of these models to tackle most scenarios on ‘auto-pilot’ results in a deskilled and novice-heavy human component in the system which is ill-equipped to tackle the inevitable occasion when the model fails. The failure is inevitably taken as evidence of human failure upon which the system is made even more automated and more safeguards and redundancies are built into the system. This exacerbates the problem of absence of feedback when small errors occur. The buildup of latent errors again increases and failures become even more catastrophic.
* * *
My focus on airplane automation and financial models is simply illustrative. There are ample signs of this incompatibility between human monitors and near-fully automated systems in other domains as well. For example, Andrew Hill observes:
In developed economies, Lynda Gratton writes in her new book The Shift, “when the tasks are more complex and require innovation or problem solving, substitution [by machines or computers] has not taken place”. This creates a paradox: far from making manufacturers easier to manage, automation can make managers’ jobs more complicated. As companies assign more tasks to machines, they need people who are better at overseeing the more sophisticated workforce and doing the jobs that machines cannot….
The insight that greater process efficiency adds to the pressure on managers is not new. Even Frederick Winslow Taylor – these days more often caricatured as a dinosaur for his time-and-motion studies – pointed out in his century-old The Principles of Scientific Management that imposing a more mechanistic regime on workers would oblige managers to take on “other types of duties which involve new and heavy burdens”…..
There is no doubt Foxconn and its peers will be able to automate their labour-intensive processes. They are already doing so. The big question is how easily they will find and develop managers able to oversee the highly skilled workforce that will march with their robot armies.
This process of integrating human intelligence with artificial intelligence is simply a continuation of the process through which human beings went from being tool-users to minders and managers of automated systems. The current transition is important in that for the first time, many of these algorithmic and automated systems can essentially run themselves with human beings performing the role of supervisors who only need to intervene in extraordinary circumstances. Although it seems logical that the same process of increased productivity that has occurred during the modern ‘Control Revolution’ will continue during the creation of the “vast,automatic and invisible” ‘second economy’, the incompability of human cognition with near-fully automated systems suggests that it may only do so by taking on an increased risk of rare but catastrophic failure.