The Simple Economics of Artificial Intelligence

By: Ajay Agrawal, Joshua Gans, Avi Goldfarb

Published: 2018.

Read: 2019.

Summary:

The basic argument is that AI (or machine learning) is mostly about better and cheaper “prediction”: the ability to use information you already have to make better decisions. It is argued that machines are better than humans at prediction because they are better at dealing with increased uncertainty and complexity.

The book is an economist’s view on some of the ways in which cheap machine prediction may affect employees, companies and society at large. Topics covered include how cheap prediction may change the roles played by humans in decision making, the tasks that they are expected to perform in a job and the potential impact of cheap prediction on a company’s operations, its strategy and industry structure.

Worth Reading:

This book was written by three professors with active entrepreneurial pursuits and that is largely how it reads: a smooth McKinsey-style presentation on how AI can be used as a tool to drive your business.

It’s easy to read, well structured, with some interesting real world examples. The drawback of its clean framework is that it often states the obvious and it mostly skates the surface.

It’s a little loose on concepts and definitions. For instance, the book’s definition of prediction (filling in missing, hidden information) is odd in the context of the decision-making process outlined later, where it looks at prediction more usefully (use input data and knowledge of input/output patterns to predict outputs, what happens or should happen next). Important, because the book is pretty much about prediction.

Practical Takeaways:

• Better prediction allows for handling more uncertainty and complexity.

Key Concepts:

Prediction is the process of filling in missing, hidden information: take information you have (“data”) and use it to generate information you don’t have.

Compared to human prediction, machine prediction generates better and cheaper predictions due to their ability to better handle complex, interactive multi-variable problems. Machine prediction requires (i) input data to train the model, (ii) input data used to make predictions, (iii) feedback data (to check on quality of predictions).

Prediction is a key ingredient in the decision making process:

• Collect input data.
• Make a prediction (compare input data to stored trained data).
• Apply judgment on what matters (assessing the relative pay-offs associated with each potential outcome).
• Choose the appropriate action.
• Monitor the outcome for feedback data.

As machine prediction becomes more accurate, the value of humans shifts from making (general) predictions (what is the fastest route for a taxi) to applying (specific) judgment (what is the best route to take now).

Human judgment is valuable in situations with high uncertainty and many potential outcomes to consider (especially when the cost of a potentially bad outcome is high).

For the moment, humans are better at assessing potential pay-offs due to their ability to process more varied input data (they have more “senses” than machines), they have a better understanding of detailed (human) preferences and machines are not good at predicting by analogy (in situations that happen infrequently and where training data is lacking).

Better machine prediction reduces uncertainty and reduces complexity. It also makes more options feasible by changing their potential pay-off or by making options clearer, changing people’s behavior (for instance, you can leave for the airport later without increasing the risk of missing a flight).

AI redesigns jobs. A job is a collection of tasks. As AI takes up certain tasks, the tasks left for humans change (for instance in accounting, bookkeepers become spreadsheet wizards). Automation may eliminate humans from a task, but not necessarily from a job; likely changes requirements for a job.

AI changes businesses. AI may change the boundaries of organizations (better prediction reduces uncertainty and allows for more complex partnerships), may allow companies to outsource capital or labor focused on data, prediction and action, while retaining labor that is focused on judgment, may increase incentives to own data (or purchase the prediction if data is non-essential).

The introduction of machine prediction to customers can mean sacrificing certain other goals and lead to short-term poor performance (as the system needs to learn from experience, the initial user experience is typically worse). This means that incumbents are typically less incentivized to introduce machine prediction than newcomers (classic innovator’s dilemma).

Other AI risk includes potential for discrimination, vulnerability to hackers, manipulation, etc.

The impact of AI on “everything else”: complex and too early to tell:

• The end of jobs and an increase of inequality: perhaps lower wages for non-automated task, but better opportunities for the highly skilled.
• The dominance of a few large companies: benefits of scale and network economies may lead to monopolies.
• The  dominance of certain countries: regulations on data usage in certain regions may impede performance.