We usually think of prediction as forecasting the future. While that’s true, it’s more than that. Prediction is filling in gaps in information, whether in the past, present or future. High fidelity predictions are immensely valuable, offering the potential to unlock new opportunities that we can’t imagine today.
Prediction is at the heart of making decisions under uncertainty. Our businesses and personal lives are riddled with such decisions. Prediction tools increase productivity—operating machines, handling documents, communicating with customers. Uncertainty constrains strategy, which means that better prediction creates opportunities for new business structures and strategies.Agrawal, Gans, Goldfarb
All decisions implicitly involve a view of the future, which is getting progressively difficult to do with any level of certainty. Meanwhile machines are entering the prediction game. In the book Prediction Machines, Agrawal, Gans and Goldfarb hypothesize that the real value of AI and Big Data are to reduce the cost of prediction, allowing many more things to be forecast, in finer detail, closer and closer to real-time.
When we think of prediction as “filling in gaps in information” we can reframe the role of ML and data. AI drops the cost of prediction, making prediction cheap in much the same way as the PC made arithmetic cheap, which made all sorts of problems become arithmetic problems—music and photos are canonical examples of digital transformation made possible by inexpensive arithmetic.
With AI, all sorts of problems become prediction problems, such as language, vision and even human behavior. This means we need to imagine all the places to fill in missing information, which is something that is really difficult and we aren’t very good at. If you’d given someone a calculator in 1995, would they have imagined all the uses for low cost arithmetic?
As prediction gets cheaper and more problems are prediction problems, what happens? Consider what happens when coffee gets cheap. What goes up in value? Cheaper coffee means more people drinking coffee which means more demand for cream and sugar. What’s the “cream and sugar” of AI? The answer is action and judgement. The more AI in a workflow, the more valuable action and judgement become. This is because there are more actions to take and judgements required to handle situations. Predictions are not a guarantee, they are a probability which places humans into the position of assigning values to new things.
One way to think about assigning new values is to imagine an airline that switches to using a facial recognition algorithm to streamline check in. We know that facial recognition is less accurate for women and people of color. In the event that the algorithm fails, the role of humans is not just to act as a “human in the loop” when the system fails, it is also to assign a value to the cost of failure. This is required to so that key design tradeoffs can be made about fairness versus accuracy. These costs include both employee and customer satisfaction (positive and negative) and the cost to the brand in the event the company being perceived as sexist and/or racist.
Judgment is also more valuable because it is paired with more options for action. Judgment is the process of determining what the reward is to a particular action in a particular environment, that is, figuring out what’s going to happen if you put your plan into action. It’s becoming much harder to determine if an action will yield a reliably predictable result and what the range of outcomes could be; it’s much harder to plan.
AI can support decision making because it provides for many more ways to fill in missing information. Experimentation, exploration and learning become more valuable strategies. Action becomes more valuable which shifts decision making to a bias for action. Judgment becomes more valuable which means that action must include data gathering on both the environment and the action itself.
This article is one in a series of hacks, tips and tricks for making better decisions.