Taking the Easy Way Out - AI, Self-Control, and Human Capital Formation

Sep 10, 2026

AI use by junior researchers may interfere with building human capital, and this paper studies that tradeoff in a dynamic choice model. A short-term project yields immediate payoff, while a long-term project yields a larger payoff later but builds skill through learning by doing. AI raises success when human capital is low yet blocks human-capital accumulation, and self-control costs can make welfare fall and even trigger switching to short-term work; under some conditions AI encourages human-capital formation.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages33

Reading Time1 h 23 min

AccessFree to download

Introduction

Modeling delegation to AI

Self-control and temptation benchmark

Four effects of AI and self-control

Blowing off

Discouragement

Reversal

Encouragement

Taking the Easy Way Out: AI, Self-Control, and Human Capital Formation ∗ Drew Fudenberg† David K. Levine‡ August 19, 2026 Abstract Some researchers worry that AI use by junior researchers inhibits acquisition of human capital. We study a dynamic model with a short-term project with a direct payoff now and a long-term project with a larger direct payoff next period. AI increases success with low human capital, but prevents learning by doing. Increasing available options ordinarily cannot decrease welfare, but if direct payoffs are tempting and self-control is costly, AI reduces welfare even if it is not adopted, and can induce the researcher to switch to the short-term project. More surprisingly, AI may increase the acquisition of human capital. ∗We thank Daron Acemoglu, Joshua Gans, Andrew Koh, and Stephanie Wang for helpful comments and suggestions, and NSF grant SES-2417162 and the Leverhulme Trust for financial support. †Department of Economics, MIT ‡Department of Economics, RHUL and Washington University in St, Louis 1 Introduction An AI agent may give a junior researcher a better chance of producing a useful output. But as Hogg [2026] and Karamanis [2026] emphasize, the value of long-term research work lies not only in the output it generates, but also in the skills and judgment acquired by doing it. We share their concern that by replacing hard research work, AI can weaken the process through which researchers acquire judgment and skill: A researcher who uses AI to generate results and text can produce a passable paper while giving up the learning by doing that would make future work better. This paper studies the decision of whether to delegate to an AI agent in a simple dynamic model where the researcher can either delegate the task to an AI agent with minimal human input or not use it at all. Each period, the researcher chooses between a short-term project and a long-term one. The short-term project provides a direct payoff now, while the long-term project provides a direct payoff only later. In low-human-capital states, working without AI can generate learning by doing and lead to the high-humancapital state, whereas using AI reduces human-capital acquisition. In a standard model without a self-control problem, the availability of AI cannot lower the agent’s average present value: They will not use it if doing so decreases their average present value, and if they do not use AI they are no worse off than if it were not available. The welfare effect of AI is no longer straightforward when the agent is tempted to forgo future direct payoffs for current direct payoffs and so has a cost of self control. To model this, we assume that the cost of resisting temptation depends on the best direct payoff available in the current period, and that this cost may be incurred even if the AI is not used. By making a high-direct-payoff option available, AI raises the temptation benchmark. As a result, the availability of AI can worsen welfare and distort project choice. The issue is not specific to academic research. Many valuable skills are acquired while performing tasks that are initially slow, frustrating, or only imperfectly productive. AI changes the private return to those tasks by making it possible to obtain acceptable output without going through the costly process of skill formation. Thus the central question is not only whether AI raises current productivity, but also whether it changes the incentives to acquire the human capital on which future productivity depends. This makes AI different from a pure productivity shock: it changes both what can be done today and what the worker has reason to learn for tomorrow. The interaction of AI and self-control gives rise to four distinct effects. • Blowing off: the researcher uses AI to carry out the long-term project in the lowhuman-capital state and forgoes the accumulation of human capital. • Discouragement: when AI becomes available, the research

FAQs

How does the paper model AI delegation decisions?

The researcher chooses each period between a short-term project and a long-term project, either delegating to an AI agent with minimal human input or not using AI at all.

Why can AI reduce welfare even without being adopted?

Because self-control costs depend on the best current-period direct payoff available, AI can raise the temptation benchmark and increase the cost of resisting even when the researcher does not use AI.

What are the four distinct effects of AI on behavior and human-capital acquisition?

Blowing off prevents accumulation by using AI on the long-term task; discouragement shifts to the short-term task in low human capital; reversal switches to short-term even when AI is not used; encouragement leads to human-capital acquisition that would not occur without AI.

Sep 10, 2026

Uploaded by 8796093062539