The Economics of Bicycles for the Mind

Source: Ajay K. Agrawal, Joshua S. Gans and Avi Goldfarb; NBER Working Paper 34034 · 2025-07

Steve Jobs described computers as “bicycles for the mind,” a tool that allowed people to dramatically leverage their capabilities. This paper presents a formal model of cognitive tools and technologies that enhance mental capabilities. We consider agents engaged in iterative task improvement, where cognitive tools are assumed to be substitutes for implementation skills and may or may not be complements to judgment, depending on their type. The ability to recognise opportunities to start or improve a process, which we term opportunity judgment, is shown to always complement cognitive tools. The ability to know which action to take in a given state, which we term payoff judgment, is not necessarily a complement to cognitive tools. Using these concepts, we can synthesise the empirical literature on the impact of computers and artificial intelligence (AI) on productivity and inequality. Specifically, while both computers and AI appear to increase productivity, computers have also contributed to increased inequality. Empirical work on the impact of AI on inequality has shown both increases and decreases, depending on the context. We also apply the model to understand how cognitive tools might influence incentives to automate processes and allocate decision-making authority within teams.

Introduction. A bicycle is a tool used by people to travel more efficiently. Indeed, a 1973 article in Scientific American calculated that it might be the lowest cost per kilogram of any means of transportation, both natural and artificial. Steve Jobs was influenced by this notion when he characterised computers as “bicycles for the mind.”1 The idea is that computers are tools that make minds more efficient in much the same way as a bicycle makes legs more efficient. This paper is about computers and artificial intelligence (AI) as cognitive tools. A growing empirical literature explores how computers and AI affect work. We build a model that enables interpretation of many of the results in this literature, both the historical evidence on computers and the emerging work on AI. It provides a unifying framework for interpreting the following empirical results:2 • Computers increased productivity of adopting workers and firms, and at the macroeconomic level (e.g. Bresnahan et al. (2002); Jorgenson and Stiroh (2000)).

• AI adoption increases productivity of adopting workers and firms (e.g. McElheran et al. (2025); Cui et al. (2025)).

• Computers increased inequality, within adopting industries and economy-wide (e.g. Dranove et al. (2014); Autor et al. (2006)).

• AI often reduces inequality within adopting firms (e.g. Brynjolfsson et al. (2025); Kanazawa et al. (2022)).

• The task-based approach to anticipating AI’s impact on the economy suggests highincome occupations will be most impacted (e.g. Brynjolfsson and Mitchell (2017); Felten et al. (2021); Eloundou et al. (2024)).

• Computer and AI adoption may not lead to full automation (e.g. Agrawal et al. (2018b); Scharre (2016)).

• For both computers and AI, team composition changes (e.g. Teodoridis (2018); Law and Shen (2025))

This framework examines how agents engage in iterative task improvement, where in each period, agents have the opportunity to enhance the quality of their work. The model has three key components: the agent’s level of effort in implementing improvements (which affects both costs and probability of success), their ability to know how those improvements translate into action, and their ability to locate new improvement opportunities. Each successful improvement increases quality by a fixed amount, but the process eventually ends when the agent fails to identify a new opportunity for enhancement. The model then examines how the introduction of a tool, defined as something that improves the ratio of success probability to cost for any given level of effort, affects this process. When a cognitive tool is adopted, agents reduce their level of direct effort since the tool makes it easier to achieve good outcomes. However, because the tool increases the probability of success relative to costs, the net benefit from improvements rises. Our first motivating example is the computer. A knowledge worker is tasked with designing a product. Absent a computer, the worker performs calculations and creates visualisations by hand. With a computer (say, used for Computer-Aided Design or CAD), the worker can now iterate more quickly. In addition to the implementation skill of performing calculations and creating visualisations, the knowledge worker also identifies opportunities to improve the outcome, which we label ‘opportunity judgment’. In this case, the introduction of the cognitive tool amplifies the value of opportunity judgment by increasing the net benefit of acting on opportunities. Our second motivating example is AI prediction in the service of decision making (Agrawal et al., 2019), for example, as in a physician providing a diagnosis. Even when a diagnosis is made, there is uncertainty about whether a treatment is available to address that diagnosis. This uncertainty creates a need for a different kind of skill, which we label ‘payoff judgment’. This refers to the decision maker’s ability to translate the predicted knowledge of the state into an optimal decision. Unlike opportunity judgment, payoff judgment increases accuracy per unit of effort but decreases effort. Thus, payoff judgment is only a complement to a cognitive tool when the tool does not decrease effort too much. We then build a general model that combines the insights from these motivating examples. The general model nests both computers and AI predictions, as well as generative AI, which provides aspects of both types of judgment. This general model yields several insights. First, cognitive tools lead to more output with less effort. They are productivity-enhancing. Second, as with the motivating examples, cognitive tools are substitutes for implementation skill, are complements to opportunity judgment, and are generally complements to payoff judgment (as long as the reduction in effort is not so large as to reduce output).

Related work. This section has documented that the impact of a cognitive tool on inequality depends on the distribution of human skills with respect to implementation and opportunity judgment. As a cognitive tool gets better, inequality may decline as the tool replaces the value created by highly skilled humans at implementation. Eventually, if the tool improves enough, inequality may begin to increase again as the humans with high judgment benefit from the complementarity between judgment and the tool. This result helps interpret two distinct literatures: the empirical literature on whether computers and AI have increased or decreased labour market outcomes and the literature that examines current labour markets to anticipate how AI might impact jobs and inequality in the future.

The empirical literature suggests that computers increased labour market inequality by automating routine tasks, because non-routine cognitive skills were rare relative to non-routine manual skills (Autor et al., 2006, 2008). Goldin and Katz (2008) argue that education did not keep up with increasing demand for non-routine skills, and therefore labour market inequality increased. The automation of human routine actions is analogous to implementation effect in our model, and may underlie the evidence that factory robots in the United States led to increased inequality through the automation of human actions (Acemoglu and Restrepo, 2018, 2020). The increased value of non-routine cognitive tasks includes the ability to recognise opportunities to increase the quality of work. Such opportunity judgment may underlie evidence that computer adoption was primarily valuable to those with access to skilled expertise (Dranove et al., 2014; Forman et al., 2012). Overall, the empirical literature on computers and inequality, when interpreted through the lens of Proposition 3, suggests that the variance of judgment skills is sufficiently large relative to the variance of implementation skills that inequality rose as a consequence of the cognitive tool of computers. In contrast, the empirical literature on AI suggests that AI has generally reduced productivity differences. Much of this literature examines what happens to different workers within the same workplace after the adoption of AI. For example, Brynjolfsson et al. (2025) show that newer and less productive call centre employees benefit most from AI recommendations, and Cui et al. (2025) show that less experienced developers benefit most from an AI-based coding assistant. Others with similar workplace results include Hui et al. (2024), Wiles et al. (2023), Dell’Acqua et al. (2023), Peng et al. (2023), and Kanazawa et al. (2022).

Method. Before providing a general model of cognitive tools, it is instructive to build up from some motivating contexts that capture the two kinds of judgment separately. The first examines computers, which are cognitive tools that follow instructions directly. The second examines the role of AI prediction in a decision-making context.

Conclusion. This paper provides a unifying economic framework for understanding how cognitive tools, specifically computers and artificial intelligence, interact with human capabilities in iterative task improvement. By decomposing human skills into implementation ability, opportunity judgment, and payoff judgment, we demonstrate that cognitive tools consistently substitute for implementation skills while complementing opportunity judgment. The relationship with payoff judgment proves more nuanced, depending on whether tools reduce human effort sufficiently to offset their productivity benefits. These insights help reconcile seemingly contradictory empirical findings: computers increased both productivity and inequality by amplifying the value of non-routine cognitive skills, while AI often reduces inequality within adopting firms by disproportionately helping workers with weaker implementation skills. Our model predicts that this equalising effect may reverse as AI tools become more sophisticated, eventually favouring workers with superior judgment capabilities and creating a U-shaped pattern in wage inequality. Our approach complements Autor and Thompson (2025) and responds to the call in Agarwal et al. (2024) for economists to specify the ways in which AI is interdependent with human decision-making skills. Our framework highlights how human judgment complements AI’s shortcomings (Loaiza and Rigobon, 2024). This complementarity explains why full automation remains elusive despite rapid technological progress: The benefit of responsive human judgment rises as AI improves.