What You Don’t Know
We are hardwired to not like holes in our memory. So much so that our brains have evolved to quite effectively fill in data that is missing, to literally invent memories to smooth out the edges of our experiences. This is one of the reasons that eyewitness accounts are so unreliable. It is also the cause of many, if not most, arguments and disagreements. People literally experience the same event but recall it in vastly different ways. Our brains construct diverging narratives, and we often reach false conclusions. This happens more than we realize (and more than we would like to admit) and it can cause us to ignore missing information.
One of the most illustrative examples of ignoring missing information comes from World War II. Air combat was still a relatively new capability and the Allies were losing an alarming number of planes. To mitigate the losses, they started studying the bullet holes of returning airplanes to determine where to reinforce them. The initial plan was to reinforce the planes where most of the bullet holes were. Luckily, Abraham Wald, a statistician disagreed. He argued that if the planes could safely return with those bullets holes, then reinforcing those areas would provide little benefit. He suggested that the areas NOT showing bullet holes were the ones that should be reinforced because the planes hit in those areas were not able to return safely. He was correct. The original approach had only considered the surviving planes, not the ones that had crashed.
A more modern example comes in the form of how we diagnose heart attacks. If I asked you to tell me the primary symptoms of someone having a heart attack, what would you say? You might say chest pain or tingling in your arms and neck, or shortness of breath. You would be partially correct; but, historically, we have primarily studied heart attacks in men. As a result, most people (including many medical professionals) are familiar with heart attack symptoms of men but not of women, which can be vastly different. As a result, more women die of heart attacks than men do. In this case, the information learned from studying men is not wrong, but it is missing significant data from half the population.
From a technology perspective, each of us interacts with artificial intelligence and machine learning on a daily basis, but these systems are only as good as the information we use to train them. Although we think of computers as rational and unbiased, a recent Forbes article describes several types of biases that affect algorithm effectiveness and impartiality. For instance, we often use datasets that over-represent certain demographics when training facial recognition algorithms. In many instances, this causes racial bias or age bias or regional bias in how our systems operate. This can even be a problem in transcription algorithms. My oldest daughter recently worked as in intern transcribing audio from recordings of public events. When I asked her why they did not use software to do the transcription, she said that it is not reliable in an uncontrolled environment, especially with people who have strong accents. The algorithms have just not seen (or heard) enough variations to be accurate in uncontrolled settings.
Although it is still challenging to address the unknowns, I routinely employ several simple, but effective, actions to mitigate the impacts of missing information and bad assumptions.
The first step is to acknowledge that we are usually missing something or that there are gaps in the data. Just being aware that we might have incomplete information is often enough to address the most damaging impacts or at least to prevent us from being surprised by the impacts.
Second, I strive to interrogate implied and explicit assumptions. We often create assumptions based beliefs that we do not question or validate. We might assume that the other team will deliver the infrastructure on time. We assume that the client has sufficient understanding of their requirements. We assume that employees have a good environment at home to work remotely. Perhaps even worse, we think that if our assumptions are not true, someone will say something. Most of the problem-projects that I have seen are the result of assumptions that did not hold up. Interrogating assumptions up front helps to validate them and to mitigate the impact if they end up being wrong.
Finally, I also endeavor to incorporate a wide and diverse set of people in activities and decisions, especially people who are not experienced with the domain or the problem I am trying to solve. We all have our biases and we are often blind to them. By including people with different experiences and different perspectives, we stand a greater chance of identifying gaps in our understanding or filling in the holes in our data. Speaking of data, we should also ensure our data samples are sufficiently large and diverse. An “anomaly” we decide to exclude may only be an anomaly in some instances. For another set of users or clients, it could be the norm.
Knowing what we don’t know is always challenging, but with some forethought and preparation, we can mitigate the impacts of what we don’t know. If we rigorously interrogate our assumptions and incorporate wide a diverse perspectives and datasets, we can use what we don’t know to our advantage to create more robust and inclusive results and solutions.