Prediction Machines: The Simple Economics of Artificial Intelligence

by Ajay Agrawal

Prediction Machines is not a recipe for success in the AI economy. Instead, we emphasize trade-offs. More data means less privacy. More speed means less accuracy. More autonomy means less control.

the constant drumbeat of technology news that we numbly recite that the only thing immune to change is change itself.

“[E]verything that civilisation has to offer is a product of human intelligence … [S]uccess in creating AI would be the biggest event in human history.”1

the price of something falls, we use more of it.

The rise of the internet was a drop in the cost of distribution, communication, and search.

What will new AI technologies make so cheap? Prediction. Therefore, as economics tells us, not only are we going to start using a lot more prediction, but we are going to see it emerge in surprising new places.

Kathryn Howe, of Integrate.ai, calls the ability to see a problem and reframe it as a prediction problem “AI Insight,” and, today, engineers all over the world are acquiring it.

Prediction facilitates decisions by reducing uncertainty, while judgment assigns value. In economists’ parlance, judgment is the skill used to determine a payoff, utility, reward, or profit. The most significant implication of prediction machines is that they increase the value of judgment.

PREDICTION is the process of filling in missing information. Prediction takes information you have, often called “data,” and uses it to generate information you don’t have.

Machine prediction is artificially generated useful information.

A] billion hours ago, modern homo sapiens emerged. A billion minutes ago, Christianity began. A billion seconds ago, the IBM PC was released. A billion Google searches ago … was this morning.”

of the prediction and the need for accuracy. These tools are called “power calculations” and tell you how many units you need to analyze to generate a useful prediction.

This might not be true from an economic point of view, which is not about how data improves prediction. It is about how data improves the value you get from the prediction. Sometimes prediction and outcome go together, so the decreasing returns to observations in statistics imply decreasing returns in terms of the outcomes you care about. Sometimes, however, they are different.

Prediction machines utilize three types of data: (1) training data for training the AI, (2) input data for predicting, and (3) feedback data for improving the prediction accuracy.

Thus, organizations need to understand the relationship between adding more data, enhancing prediction accuracy, and increasing value creation.

This human-machine division of labor worked because it overcame human weaknesses in speed and attention, and machine weaknesses in interpreting text.

Understanding the division of labor involves determining which aspects of prediction are best performed by humans or machines. This enables us to identify their distinctive roles.

at Harvard Medical School, presented physicians with two treatments for lung cancer: radiation or surgery. The five-year survival rate recommends surgery. Two groups of participants received different ways of presenting information about the short-term survival rate of surgery, which is riskier than radiation. When told that “the one-month survival rate is 90 percent,” 84 percent of physicians chose surgery, but that rate fell to 50 percent when told that “there is a 10 percent mortality in the first month.” Both these phrases said the same thing, but how the researchers framed the information resulted in major changes in the decision. A machine would not have this outcome.

In other words, the interaction effect may be the most important, and as the number of dimensions for such interactions grows, humans’ ability to form accurate predictions diminishes.

There are known knowns; there are things we know we know. We also know there are known unknowns; that is to say we know there are some things we do not know. But there are also unknown unknowns—the ones we don’t know we don’t know. And if one looks throughout the history of our country and other free countries, it is the latter category that tend to be the difficult ones.9

“counterfactual”: what would have happened if you took a different action.

Prediction machines are better than humans at factoring in complex interactions among different indicators, especially in settings with rich data. As the number of dimensions for such interactions grows, the ability of humans to form accurate predictions diminishes, especially relative to machines. However, humans are often better than machines when understanding the data generation process confers a prediction advantage, especially in settings with thin data. We describe a taxonomy of prediction settings (i.e., known knowns, known unknowns, unknown knowns, and unknown unknowns) that is useful for anticipating the appropriate division of labor.

Prediction machines scale. The unit cost per prediction falls as the frequency increases. Human prediction does not scale the same way. However, humans have cognitive models of how the world works and thus can make predictions based on small amounts of data. Thus, we anticipate a rise in human prediction by exception whereby machines generate most predictions because they are predicated on routine, regular data, but when rare events occur the machine recognizes that it is not able to produce a prediction with confidence, and so calls for human assistance. The human provides prediction by exception.

When someone (or something) makes a decision, they take input data from the world that enables a prediction. That prediction is possible because training occurred about relationships between different types of data and which data is most closely associated with a situation. Combining the prediction with judgment on what matters, the decision maker can then choose an action. The action leads to an outcome (which has an associated reward or payoff). The outcome is a consequence of the decision. It is needed to provide a complete picture. The outcome may also provide feedback to help improve the next prediction.

decision making is ubiquitous throughout our economic and social lives. However, a prediction is not a decision—it is only a component of a decision. The other components are judgment, action, outcome, and three types of data (input, training, and feedback).

The value of substitutes to prediction machines, namely human prediction, will decline. However, the value of complements, such as the human skills associated with data collection, judgment, and actions, will become more valuable.

Judgment involves determining the relative payoff associated with each possible outcome of a decision, including those associated with “correct” decisions as well as those associated with mistakes. Judgment requires specifying the objective you’re actually pursuing and is a necessary step in decision making.

In economic terms, the cost of figuring out the payoffs will mostly be time.

more outcomes mean more judgment means more time and effort. Humans experience the cognitive costs of judgment as a slower decision-making process. We all have to decide how much we want to pin down the payoffs against the costs of delaying a decision. Some will choose not to investigate payoffs for scenarios that seem remote or unlikely. The credit card network might find it worthwhile to separate work trips from vacations but not vacations to Rome from the Grand Canyon.

Judgment, whether by deliberation or experimentation, is costly.

Prediction machines are a tool for humans. So long as humans are needed to weigh outcomes and impose judgment, they have a key role to play as prediction machines improve.

uncertainty increases the cost of judging the payoffs for a given decision.

In these cases, it is more efficient for a human to apply judgment after the prediction machine predicts.

For every bomber that returned from bombing raids over Germany, the engineers could see where they had been hit by antiaircraft fire. The bullet holes in the planes were their data. But were these the obvious places to better protect the plane? They asked statistician Abraham Wald to assess the problem. After some thought and some rather thorough mathematics, he told them to protect the places without bullet holes. Was he confused? That seemed counterintuitive. Didn’t he mean to protect the areas of the plane that did have bullet holes? No. He had a model of the process that generated the data. He recognized that some bombers did not come back from the raids and conjectured that these bombers got hit in places that were fatal. In contrast, the bombers that made it home were hit in places that were not fatal. With this insight, the air force engineers increased the armor in the places without bullet holes, and the planes were better protected.6 Wald’s insight about the missing data required an understanding of where the data came from; given that the problem had not arisen before, the engineers did not have prior examples to draw from. For the foreseeable future, such calculations are beyond the abilities of prediction machines.

Such compromises are a key aspect of how humans make decisions. Economics Nobel Prize–winner Herbert Simon called this “satisficing.” While classical economics models superintelligent beings making perfectly rational decisions, Simon recognized and emphasized in his work that humans cannot cope with complexity. Instead, they satisfice, doing the best they can to meet their objectives. Thinking is difficult, so people take shortcuts.

Carmakers in the United States have reached an agreement with the Department of Transportation to make automatic emergency braking standard on vehicles by 2022.3

Often, the distinction between AI and automation is muddy. Automation arises when a machine undertakes an entire task, not just prediction. As of this writing, a human still needs to periodically intervene in driving. When should we expect full automation?

Humans and machines can accumulate data, whether for input, training, or feedback, depending on the data type. A human must ultimately make a judgment, but the human can codify judgment and program it into a machine in advance of a prediction. Or a machine can learn to predict human judgment through feedback. This brings us to the action. When is it better for machines rather than humans to undertake actions? More subtly, when does the fact that a machine is handling the prediction increase the returns to the machine rather than a human also undertaking the action? We must determine the returns to machines performing the other elements (data collection, judgment, actions) to decide whether a task should be or will be fully automated.

Automation occurs when the return to machines handling all functions is greater than the returns to including humans in the process.

cleverly designed to remove the possibility that robots harm any human.8

AI tools can change work flows in two ways. First, they can render tasks obsolete and therefore remove them from work flows. Second, they can add new tasks. This may be different for every business and every work flow.

In deciding how to implement AI, companies will break their work flows down into tasks, estimate the ROI for building or buying an AI to perform each task, rank-order the AIs in terms of ROI, and then start from the top of the list and begin working downward. Sometimes a company can simply drop an AI tool into their work flow and realize an immediate benefit due to increasing the productivity of that task. Often, however, it’s not that easy. Deriving a real benefit from implementing an AI tool requires rethinking, or “reengineering” the entire work flow. As a result, similar to the personal computer revolution, it will take time to see productivity gains from AI in many mainstream businesses.

Today, AI tools predict the intention of speech (Amazon’s Echo), predict command context (Apple’s Siri), predict what you want to buy (Amazon’s recommendations), predict which links will connect you to the information you want to find (Google search), predict when to apply the brakes to avoid danger (Tesla’s Autopilot), and predict the news you will want to read (Facebook’s newsfeed). None of these AI tools are performing an entire work flow. Instead, each delivers a predictive component to make it easier for someone to make a decision. AI empowers.

Every task has a group of decisions at its heart, and those decisions have some predictive element.

For a business school, for example, it is easy to say that they are focused on recruiting the “best” students, but in order to specify the prediction, we need to specify what “best” means—highest salary offer upon graduation? Most likely to assume a CEO role within five years? Most diverse? Most likely to donate back to the school after graduation? Even seemingly straightforward objectives, like profit maximization, are not as simple as they first appear. Should we predict the action to take that will maximize profit this week, this quarter, this year, or this decade? Companies often find themselves having to go back to basics to realign on their objectives and sharpen their mission statement as a first step in their work on their AI strategy.

Research into the grasping problem uses reinforcement learning to train robots to mimic humans. The Vancouver-based startup Kindred—founded by Suzanne Gildert, Geordie Rose, and a team that includes one of us (Ajay)—is using a robot called Kindred Sort, an arm with a mix of automated software and a human controller.2 Automation identifies an object and where it needs to go, while the human—wearing a virtual reality headset—guides the robot arm to pick it up and move it. In its first iteration, the human can sit somewhere away from a warehouse and fill in the missing link in the fulfillment work flow, deciding the approach angle and grip pressure, through teleoperation of the robotic arm. Long term, however, Kindred is using a prediction machine trained on many observations of a human grasping via teleoperation to teach the robot to do that part itself.

AI tools may augment jobs, as in the example of spreadsheets and bookkeepers. AI tools may contract jobs, as in fulfillment centers. AI tools may lead to the reconstitution of jobs, with some tasks added and others taken away, as with radiologists. AI tools may shift the emphasis on the specific skills required for a particular job, as with school bus drivers.

AI tools may shift the relative returns to certain skills and, thus, change the types of people who are best suited to particular jobs. In the case of bookkeepers, the arrival of the spreadsheet diminished the returns to being able to perform many calculations quickly on a calculator. At the same time, it increased the returns to being good at asking the right questions in order to fully take advantage of the technology’s ability to efficiently run scenario analyses.

It could not have accomplished all this without using a prediction machine to resolve that key uncertainty.

Similarly, the costs and risks associated with AI will fall over time, so that many businesses not at the forefront of developing digital tools will adopt it. In doing so, the demand side will drive them: the opportunity to resolve fundamental dilemmas in their business models by reducing uncertainty.