Exploring the Impact of Artificial Intelligence: Prediction versus Judgment

Source: Ajay Agrawal, Joshua S. Gans, Avi Goldfarb; Information Economics and Policy (NBER w24626, 2018) · 2019

Based on recent developments in the field of artificial intelligence (AI), we examine what type of human labor will be a substitute versus a complement to emerging technologies. We argue that these recent developments reduce the costs of providing a particular set of tasks – prediction tasks. Prediction about uncertain states of the world is an input into decision-making. We show that prediction allows riskier decisions to be taken and this is its impact on observed productivity although it could also increase the variance of outcomes as well. We consider the role of human judgment in decision-making as prediction technology improves. Judgment is exercised when the objective function for a particular set of decisions cannot be described (i.e., coded). However, we demonstrate that better prediction impacts the returns to different types of judgment in opposite ways. Hence, not all human judgment will be a complement to AI. Finally, we show that humans will delegate some decisions to machines even when the decision would be superior with human input.

Introduction. Artificial intelligence (AI) has advanced markedly in the past decade. With advances in machine learning – particularly deep learning and reinforcement learning – AI has conquered image recognition, language translation, and games such as Go (Brynjolfsson and McAfee, 2014).

This has raised the usual questions with regard to the impact of such new general purpose technologies on human productivity (Cockburn, Henderson, and Stern 2018; Brynjolfsson, Rock, and Syverson 2018). Will AI substitute or complement humans in the workforce? (Autor, 2015; Markov, 2015; Acemoglu and Restrepo, 2016). In this paper, we build a simple model that takes a careful approach to precisely what new advances in AI have generated in a technological sense and applies this to a microeconomic model of task production. In so doing, we are able to provide some insight on the complements/substitutes question as well as where the dividing line between human and machine performance of cognitive tasks might be. Our approach is, naturally, a first step in exploring the impact of AI but we believe sets the stage for more substantive investigations.

At the core of our approach is noting that recent developments in AI are all advances in prediction – in its statistical sense. Prediction is when you use information you do have to produce information you do not have. For instance, using past weather data to predict the weather tomorrow. Or using past classification of images with labels to predict the labels that apply to an image you are currently looking at. This is all machine learning does. It does not establish causal relationships and it must be used with care in the face of model uncertainty and limited data (Ng, 2016; Agrawal, Gans and Goldfarb, 2018a). But in an economic sense, if we were to model the impact of AI, the starting point would be a dramatic fall in the cost of providing quality predictions.

With this insight, we embed these changes in a standard model of decision making under uncertainty. As might be expected, having better predictions leads to better and more nuanced decisions – in particular, signal-contingent decisions that vary depending upon predictions received. But we note, however, that better predictions also change the returns to understanding what the payoff functions are from different (state, action) pairs. After all, if a state is so rare that you never take an action based on it, then there is no real return to understanding the payoff in that state. We term this process of understanding ‘judgment’. At the moment, it is uniquely human as no machine can form those payoffs. This allows us to consider how better prediction interacts with the process of human judgment.

In this paper, our approach is to delve into the weeds of what is happening currently in the field of artificial intelligence (AI) to examine precisely what type of human labour will be a substitute versus a complement to emerging technologies. In Section 2, we show that prediction allows riskier decisions to be taken and this is its impact on observed productivity although it could increase the variance of outcomes as well. In Section 3, we make our stand on the issue of “what computers cannot do” and consider the role of human judgment in decision-making. Judgment is exercised when the objective function for a set of decisions cannot be described (i.e., coded).

However, we demonstrate that better prediction impacts the returns to different types of judgment in opposite ways. Hence, not all human judgment will be a complement to AI. Section 4 then considers the design of prediction technology when prediction may be unreliable. Section 5 then examines span of attention issues for human judgment, demonstrating that humans may give the machines real authority even when human input would lead to better decisions. Finally, in the conclusion, we conjecture what will happen when AI learns to predict the judgment of humans.

Method. We now turn to model the impact of a reduction in the cost of prediction on decision-making.

This model uses many of the same elements as our prior work (Agrawal, Gans, and Goldfarb 2018b,c). We assume there are two actions that might be taken: a safe action and a risky action.

The safe action generates an expected payoff of S while the risky action’s payoff depends on the state of the world. If the state of the world is good then the payoff is R while if it is bad, the payoff is r. We assume that R > S > r. These expected payoffs represent the decision-maker’s utility from each action.

Which action should be taken depends on the prediction of how likely the good rather than the bad state will arise. To keep things simple, we will suppose that the probability of each state is 1⁄2. Prediction is of value because it makes taking the risky action less risky. To capture this, we assume that:

Then in the absence of a prediction, the decision-maker will take the safe action. Better prediction means that the decision-maker is more likely to face a probability that is closer to 1 or 0. Thus, better prediction increases the likelihood the decision-maker receives ‘good news’ and takes the risky action.

To make this concrete, suppose that the prediction technology is such that with probability, e, the decision-maker learns the true state. Thus, if the prediction technology is available then the decision-maker’s expected payoff is:

Thus, the better is the prediction technology (e), the higher the expected payoff. The returns to creating a better prediction technology depend on (R – S); the difference between the upside payoff and the safe payoff.

Prediction and Human Judgment Having specified the baseline role of prediction in decision-making, we now turn to consider the application of judgment. The need for judgment arises because the decision-maker cannot describe the utility function perfectly in advance. Specifically, having received a prediction regarding the likely state of the world, the decision-maker engages in thought that allows them to assess the payoff from each action. In this mode, the payoffs specified previously are simply prior beliefs and having engaged in thought, the decision-maker can update those beliefs.

To model this, we assume that there are hidden attributes that, if known, can change the assessment of return to the risky action in the good state. Specifically, we assume that with probability �/2 (� < 1), a hidden opportunity that boosts R by D arises. Similarly, we assume with the same probability �/2 a hidden cost that reduces R by D arises. We assume that hidden opportunities and hidden costs are mutually exclusive events (i.e., they cannot both arise). Thus, the expected payoff from the risky action in the good state remains R. We focus attention on hidden attributes that are consequential. Therefore, we assume that:

Thus, if the decision-maker is uncertain regarding the state, identifying an opportunity will cause them to choose the risky action. If they are certain about the state, then identifying a cost will cause them to choose the safe action.

Judgment is the ability to recognize hidden attributes when they arise. Humans might be able to exercise judgment in identifying hidden opportunities or hidden costs and we treat their abilities in this regard as distinct. Thus, we assume that with probability, �:, the decision-maker gets ‘good’ news and can discover a hidden opportunity (if it arises) while with probability, �;, the decisionmaker gets ‘bad’ news can discover a hidden cost (if it arises).1 To begin, if there are no hidden attributes, the decision-maker will choose the safe action. If such attributes are discovered, the exercise of the resulting judgment will only change the decision if a hidden opportunity is uncovered. Therefore, the expected payoff becomes:

: �1� 0 0 0 0 0 0 0 0 The application of judgment for hidden costs does not impact on the expected payoff precisely because the discovery of such costs will not change the decision made. Notice that this opens up the possibility that a risky action will be taken in what turns out to be the bad state.

Conclusion. In this paper, we explore the consequences of recent improvements in machine learning technology that have advanced the broader field of artificial intelligence. In particular, we argue that these advances in the ability of machines to conduct mental tasks are driven by improvements in machine prediction. Therefore, we examine sources of comparative advantage in the presence of improved machine prediction. Specifically, we identify judgment as an activity, distinct from prediction, that can improve decision-making outcomes. Whereas prediction is information regarding the expected state of the world that can be easily described, judgment relies on factors that are indescribable. These are things often classed as intuition, transference, and the drawing of analogies for unfamiliar situations. To be sure, judgment is not a passive activity. Instead, it requires deliberative cognitive application.

With this distinction formalized, we demonstrate several outcomes. The first is that when it comes to whether more prediction enhances the value of judgment or not, the type of judgment matters. This is because prediction tends to favor choosing actions that are riskier when undertaken without information. So, while, in the absence of prediction, judgment might push decisions towards riskier actions if it identifies hidden opportunities (that is, factors that make the riskier action have greater up-side potential), when there is prediction, the reverse is true. In such situations, the role of judgment is to move away from the riskier action to safer ones. And this will happen if that judgment is directed at identifying hidden costs associated with the riskier action.

Thus, we argue that as the cost of prediction falls and prediction is applied to more decisions, then the type of judgment that is valuable will move from judgment regarding good news to judgment regarding bad news.

We then turn to consider trade-offs in the design of prediction technology. In this regard, we argue that there is a potential trade-off between more predictions being generated and made available to decision-makers and the reliability of those predictions.

Limitations. Our analysis, however, remains just the beginning and there are numerous directions in which our exploration of prediction and judgment can proceed. For instance, thusfar, we have defined judgment in the negative: an intervention that cannot be described. This is necessary because if it can be described then it is conceivable that it could form part of the prediction that machines can undertake. In reality, this is useful only in a static context. One of the hallmarks of machine learning is that prediction can improve over time as the machines are able to observe the outcomes associated with an observed state and a chosen action. Thus, for instance, when a human intervenes because they assess there are hidden costs, this intervention becomes data for the machine. Given enough such interventions, the machine can potentially form an inference regarding when those costs arise. Hence, the machine would be less likely to recommend the risky action if similar circumstances arise in the future. From our perspective, at the margin, this new data would then allow prediction to replace judgment.

Machines could even be designed so that the process of having prediction replace judgment is deliberate. To get machines to learn judgment, the machines need examples of judgment; so they can observe it in more environments and learn to mimic it.