Something is wrong with automation. If we can find diagnostics performed more than 20 years ago and the conclusions are still current…something is wrong.
- Ironies of Automation (1983)
- FAA Report on Automation (1996)
- Organizational Learning from Experience in High-Hazard Industries (2002)
Of course, we could extend the examples to books like Information Processing and Human-Machine Interaction: An Approach to Cognitive Engineering published by Rasmussen in 1986, Safeware written by Leveson in 1995, Normal Accidents by Perrow in 1999, The Human Interface by Raskin in 2000 and many others.
None of these resources is new but all of them can be read by someone with interest in what is happening NOW. Perhaps there is a problem in the basics that is not still properly addressed.
Certainly, once a decision is made, going back is extremely expensive and manufacturers will try to defend their solutions. An example that I have used more than once is the fact that modern planes have processors so old that the manufacturer does not make them anymore. Since the lifetime of a plane is longer than the lifetime of some key parts, they have to stock those parts since they cannot ask the manufacturers to send them.
The obvious solution should be renewal but this should be so expensive that they prefer having brand-new planes with old-fashioned parts to avoid new certification processes. Nothing to oppose to this practice. It’s only a sample of a more general practice: Keeping attached to a design and defend it against any doubt –even if the doubt is reasonable- about its adequacy.
However, this rationale can be applied to products already in the market. What about the new ones? Why the same problems appear once and again instead of being finally solved?
Perhaps, a Human Factors approach could be useful to identify the root problem and help to fix it. Let’s speak about Psychology:
The first psychologist that won a Nobel Prize was Daniel Kahnemann. He was one of the founders of the Behavioral Economics concept showing how we use heuristics that usually works but we can be misguided in some situations by heuristics. To show that, he and many followers designed interesting experiments that make clear that we all share some “software-bugs” that can drive us to commit a mistake. In other words, heuristics should be understood as a quick-and-dirty approach, valid for many situations but useless if not harming in others.
Many engineers and designers would be willing to buy this approach and, of course, their products should be designed in a way that would enforce a formal rational model.
The most qualified opposition to this model comes from Gigerenzer. He explains that heuristics is not a quick-and-dirty approach but the only possible if we have constraints of time or processing possibilities. Furthermore, for Gigerenzer people extracts intelligence from context while the experiments of Kahnemann and others are made in strange situations and designed to misguide the subject of the experiment.
An example, used by Kahnemann and Tversky is this one:
Linda is 31 years old, single, outspoken, and very bright. She majored in philosophy. As a student, she was deeply concerned with issues of discrimination and social justice, and also participated in anti-nuclear demonstrations.
Which is more probable?
- Linda is a bank teller.
- Linda is a bank teller and is active in the feminist movement.
The experiment tries to show the conjunction fallacy, that is, how many people should choose the second alternative while the first one is not only wider but comprises the second one.
The analysis of Gigerenzer is different: Suppose that all the information about Linda is the first sentence Linda is 31 years old. Furthermore, suppose you don’t give information and simply makes the questions…we could expect that the conjunction fallacy should not appear. It appears because the experimenter provides information and, since the subject is given information, he supposes that this is RELEVANT…otherwise, why is the subject fed with this information?
In real life, relevance is a clue. If someone tells us something, we understand that it has a meaning and that this information is not included to deceive us. That’s why Gigerenzer criticizes the Behavioral Economics approach, which can be shared by many designers.
For Gigerenzer, we decide about how good a model is comparing it with an ideal model –the rational one- but if, instead, we decide about which is the best model looking at the results, we could find some surprises. That’s what he did at Simple Heuristics that Make Us Smart, that is, comparing complex decision models with others that, in theory, should get a worse performance and finding that, in many cases, the “bad” model could get better results than the sophisticated one.
Let’s go back to automation design. Perhaps we are making the wrong questions at the beginning. Instead of “What information would you like to have?” getting a Santa Claus letter as an answer, we should ask what are the cues that you use to know that this specific event is happening?
FAA, in its 1996 study, complained about the fact that some major failures as an engine-stop can be masked by a bunch of warnings about different systems failing, making hard to discern that all of them came from a common root, that is, the engine stop. What if we ask “Tell me one fact –exceptionally I would admit two- that should tell you in a clear and fast way that one of the engines is stopped.”
We have a nice example from QF32 case. Pilots started to distrust the system when they got information that was clearly false. It was a single fact but enough to distrust. What if, instead of deciding this way jumping to the conclusion from a single fact, they should have been “rational” trying to assign probabilities in different scenarios? Probably, the plane should not have fuel enough to allow this approach.
Rasmussen suggested one approach –a good one- where the operator should be able to run cognitively the program that the system was performing. The approach is good but something is still missing: How long should it take for the operator to replicate the functional model of the system?
In real life situations, especially if they have to deal with uncertainty –not calculated risk- people use very few indicators easy and fast to obtain. Many of us remember the BMI-092 case. Pilots were using an indicator to know which engine had the problem…unfortunately, they came from a former generation of B737 and they did not know that the one they were flying had air bleeding in both engines instead of only one. The key used to determine the wrong engine should have been correct in an older plane.
Knowing the cues used by pilots, planes could be designed in a human-centered approach instead of creating an environment that does not fit with the ways used by people to perform real tasks in real environments.
When new flight-deck designs appeared, manufacturers and regulators were careful enough to keep the basic-T, even though it could appear in electronic format but that was the way that pilots used to get the basic information. Unfortunately, this has disappeared in many other things and things like position of power levers with autopilot, position of flightsticks/horns and if they have to transmit pressure or not or if the position should be common to both pilots or not…had a very different treatment from a human-centered approach. Instead, the screen-mania seems to be everywhere.
A good design starts with a good question and, perhaps, questions are not yet good enough and that’s why analyses and complains 20 and 30 years old still keep current.