I recently finished “Range” [link], a fun and breezy read in the 'airport business-ish pop-psych’ genre exemplified by Malcom Gladwell. It’s pretty good, albeit suffers from the classic ‘book that could have been an article’ problem. The key idea is pretty simple and succinct, which is that trying a lot of different things makes you more productive than hyper-specializers. It’s an idea I find inherently appealing, and as a result I’m very suspicious of it. But it also references a few times a framing device I’m new to, but is incredibly useful when thinking about how to apply machine learning/”artificial intelligence” to real-world problems, specifically “wicked problems”.
What is a ‘wicked problem’?
A wicked problem is one that defies easy characterization [link] - literally the first classification rule is “there is no definitive characterization of a wicked problem’ (the second rule is that you don’t talk about Fight Club). More helpfully, it is one that can not easily be solved because the functional requirements cannot easily be described. Often there are elements of feedback, where the underlying phenomenon can change in response to things you do to solve them. It’s the opposite of a “tame problem”, which is one that can be easily understood, and where there can be a closed-form solution. For those who are interested, it’s well worth diving into the 1973 paper [link] which lays out the seminal definition.
Most problems worth solving in the social, economic, or policy sphere are wicked problems. It’s often the case that precisely describing the problem is a key part of what makes it wicked, or in the authors’ words, “Goal-finding is turning out to be an extraordinarily obstinate task”
. Disagreement over what is an appropriate goal to a wicked problem is why it’s wicked, even before you get to the part where the problem you seek to solve starts to fight back against you. These are problems like California’s homelessness problem, where two very intelligent people can easily disagree over whether the problem at hand is one of law and order on the streets or one of insufficient affordable housing.
The authors lay out a number of criteria of a wicked problem, but I particularly love #8 - every wicked problem can be described as a symptom of another problem. In the California homelessness example, the essential issue in addressing the problem lies in whether the problem to be solved is the presence of homelessness (a problem for those with homes) or lack of homes (a problem for those without). As the authors point out, “The level at which a problem is settled depends upon the self-confidence of the analyst and cannot be decided on logical grounds”.
The question of deciding which problem/symptom is to be solved is so often the essence of a wicked problem.
Think about that and then ask yourself how how an algorithm will solve homelessness.
AI/Machine Learning Cannot Beat Wicked Problems
“AI and Machine Learning” mean many things to many people, but fundamentally it’s about more efficiently encoding complicated datasets. The idea is that you feed a ton of data into a computer and something else comes back out. This “something else” can take many forms - it might be dimension reduction, which lets you know which seemingly separate attributes actually have a lot in common. It might be classification, which is a prediction of what “type” an object is. These can be hideously complex, but classification can be quite simple - “if it walks like a duck and quacks like a duck, it’s a duck” is a classification algorithm.
There are a lot of interesting things about machine learning, but what these tools are at root is systematize information about tame problems. That is to say, coming up with a single number or prediction that is a condensation of all the information that has been revealed in the past. Chatbots for handling customer service interactions are tremendously impressive technology, but ultimately it’s a tool for the handling of completely routinized interactions with clear(ish) success and failure modes. Building a system to route a driverless car around an object is a tame problem, building it to handle traffic is much less tame - and trying to figure out how to handle a city full of kinda-buggy semi-driverless cars is a very, very wicked problem.
“Machine learning” is ultimately meant for machine data. That is to say, data that is produced in a somewhat predictable fashion from inputs that will remain constant over time. There’s no easy machine learning solution for problems where the goal isn’t constant over time, or where the subjects of your model fight back when you try to study and predict them. In a business context, this might be a question like identifying high performers. Once it’s clear how people might be stacked and ranked, those people will start devoting effort to gaming those rankings.
Whats the upshot here?
Most interesting problems in life are essentially wicked. Tinder may be able to predict quite well what your chances of a swipe are, but it can’t tell you how to find love or what to value in a partner. If you can’t help but swipe right on people who make you miserable, it’ll keep feeding you misery all day long. It can’t tell you when you’re swiping for all the wrong reasons.
A wicked world punishes those who treat it like a tame world. Naively accepting the outputs of a machine learning system will work great just until the moment it doesn’t. This might be a shift in the context that renders your model out of date, or changing circumstances that change what you actually need or value. I’d treat predictions about how machine learning will change the world with a lot of skepticism - while there’s amazing stuff that it can do, it’s best employed with the watchful eye of a human trainer to make sure it’s not leading us astray.