NBER Working Paper (Guillermo Cruces, Diego Fernández Meijide, Sebastian Galiani, Ramiro H. Gálvez, María Lombardi) • Published: February 2026
Does generative artificial intelligence (AI) reinforce or reduce productivity differences across workers? Existing evidence largely studies AI within firms and occupations, where organizational selection compresses educational heterogeneity, leaving unclear whether AI narrows productivity gaps across individuals with different levels of education. A related question is whether the productivity gains from AI reflect productive use of the tool or mere delegation, fading once AI is unavailable, and whether the answer differs across education groups. We address both questions using a randomized online experiment outside firms, in which 1,174 adults aged 25–45 complete an incentivized, workplace style business problem-solving task with or without a generative-AI assistant, followed by a non-AI-assisted follow-up module. AI increases performance for all participants, with substantially larger gains for lower-education individuals. In the control group without AI, higher-education participants outperform lower-education participants by 0.548 standard deviations; with AI, this gap falls to 0.139 standard deviations, closing about three-quarters of the initial gap. Chat-log measures explain both why this gap narrows and why it does not disappear: lower-education participants successfully obtain substantial assistance from AI, while higher-education participants use the tool somewhat more effectively across several margins. The follow-up results show that these gains are not purely driven by delegation; treated participants do not perform worse than controls once AI is removed, and lower-education participants retain part of their gain, although a sizable education gap re-emerges. Consistent with this interpretation, intensive use of the assistant produces strong task performance regardless of how much time and effort participants devote themselves, but follow-up performance is substantially higher only when intensive AI use is combined with sustained effort on the task. We interpret these findings as evidence that generative AI narrows effective productivity differences in task execution while underlying human-capital differences continue to shape unassisted performance and effective use of the tool.
NBER Working Paper (Daron Acemoglu, David Autor, Simon Johnson) • Published: February 2026
This paper defines pro-worker technologies, including Artificial Intelligence, as technologies that make human skills and expertise more valuable by expanding worker capabilities. Our conceptual framework distinguishes among five categories of technological change: labor-augmenting, capital-augmenting, automating, expertise-leveling, and new task-creating. Only the last category is unambiguously pro-worker, generating demand for novel human expertise rather than commodifying it. We illustrate these distinctions through hypothetical and real-world examples spanning aviation maintenance, electrical services, custodial work, education, patent examination, and gig delivery. While AI’s capacity to automate work is substantial, we argue that its potential to serve as a collaborator, by extending human judgment, enabling new tasks, and accelerating skill acquisition, is equally transformative and currently underexploited. We identify market failures, including misaligned firm and developer incentives, path dependence, and a pervasive pro-automation ideology, that may lead to underinvestment in pro-worker AI. We consider nine policy directions that would change incentives, including targeted investments in health care and education, tax code reform, antitrust enforcement, and intellectual property protections for worker expertise.
NBER Working Paper (Mert Demirer, John J. Horton, Nicole Immorlica, Brendan Lucier, Peyman Shahidi) • Published: February 2026
Production is a sequence of steps that can be executed (1) manually, (2) augmented with AI, or (3) fully automated within contiguous AI-executed steps called “chains.” Firms optimally bundle steps into tasks and then jobs, trading off specialization gains against coordination costs. We characterize the optimal assignment of humans and AI to steps and the firm’s resulting job structure, showing that comparative advantage logic can fail with AI chaining. The model implies non-linear productivity gains from AI quality improvements and admits a CES representation at the macro level. Empirical evidence supports the model’s key predictions that (1) AI-executed steps co-occur in chains, (2) dispersion of AI-exposed steps lowers AI execution at the job level, and (3) adjacency to AI-executed steps increases the likelihood that a step is AI-executed.
NBER Working Paper (Anton Korinek, Lee Lockwood) • Published: February 2026
Transformative artificial intelligence (TAI) - machines capable of performing virtually all economically valuable work - may gradually erode the two main tax bases that underpin modern tax systems: labor income and human consumption. We examine optimal taxation across two stages of artificial intelligence (AI)-driven transformation. First, if AI displaces human labor, we find that consumption taxation may serve as a primary revenue instrument, with differential commodity taxation gaining renewed relevance as labor distortions lose their constraining role. In the second stage, as autonomous artificial general intelligence (AGI) systems both produce most economic value and absorb a growing share of resources, taxing human consumption may become an inadequate means of raising revenue. We show that the taxation of autonomous AGI systems can be framed as an optimal harvesting problem and find that the resulting tax rate on AGI depends on the rate at which humans discount the future. Our analysis provides a theoretically grounded approach to balancing efficiency and equity in the Age of AI. We also apply our insights to evaluate specific proposals such as taxes on robots, compute, and tokens, as well as sovereign wealth funds and windfall clauses.
NBER Working Paper (Erik Brynjolfsson, J. Frank Li, Javier Miranda, Robert Seamans, Andrew J. Wang) • Published: March 2026
This paper studies how minimum wage policy affects firms’ adoption of automation technologies. Using both state-level measures of robot exposure and novel plant-level data on industrial robot imports linked to U.S. Census microdata from 1992–2021, we show that increases in minimum wages raise the likelihood of robot adoption in manufacturing. Our preferred identification exploits discontinuities at state borders, comparing otherwise similar firms exposed to different wage floors. Across specifications, a 10 percent increase in the minimum wage increases robot adoption by roughly 8 percent relative to the mean.
NBER Working Paper (Yijie Wang, Hao Gao, Campbell R. Harvey, Yan Liu, Xinyuan Tao) • Published: February 2026
The standard approach to portfolio selection involves two stages: forecast the asset returns and then plug them into an optimizer. We argue that this separation is deeply problematic. The first stage treats cross-sectional prediction errors as equally important across all securities. However, given that final portfolios might differ given distinct risk preferences and investment restrictions, the standard approach fails to recognize that the investor is not just concerned with the average forecast error - but the precision of the forecasts for the specific assets that are most important for their portfolio. Hence, it is crucial to integrate the two stages. We propose a novel implementation utilizing machine learning tools that unifies the expected return generation process and the final optimized portfolio. Our empirical example provides convincing evidence that our end-to-end method outperforms the traditional two-stage approach. In our framework, each investor has their own, endogenously determined, efficient frontier that depends on risk preferences, investor-specific constraints, as well as exposure to market frictions.
NBER Working Paper (Hemanth Asirvatham, Elliott Mokski, Andrei Shleifer) • Published: February 2026
We present the GABRIEL software package, which uses GPT to quantify attributes in qualitative data (e.g. how “pro innovation” a speech is). GPT is evaluated on classification and attribute rating performance against 1000+ human annotated tasks across a range of topics and data. We find that GPT as a measurement tool is accurate across domains and generally indistinguishable from human evaluators. Our evidence indicates that labeling results do not depend on the exact prompting strategy used, and that GPT is not relying on training data contamination or inferring attributes from other attributes. We showcase the possibilities of GABRIEL by quantifying novel and granular trends in Congressional remarks, social media toxicity, and county-level school curricula. We then apply GABRIEL to study the history of tech adoption, using it to assemble a novel dataset of 37,000 technologies. Our analysis documents a tenfold decline of time lags from invention to adoption over the industrial age, from ~50 years to ~5 years today. We quantify the increasing dominance of companies and the U.S. in innovation, alongside characteristics that explain whether a technology will be adopted slowly or speedily.
NBER Working Paper (Daron Acemoglu, Dingwen Kong, Asuman Ozdaglar) • Published: March 2026
We study how generative AI, and in particular agentic AI, shapes human learning incentives and the long-run evolution of society’s information ecosystem. We build a dynamic model of learning and decision-making in which successful decisions require combining shared, community-level general knowledge with individual-level, context-specific knowledge; these two inputs are complements. Learning exhibits economies of scope: costly human effort jointly produces a private signal about their own context and a “thin” public signal that accumulates into the community’s stock of general knowledge, generating a learning externality. Agentic AI delivers context-specific recommendations that substitute for human effort. By contrast, a richer stock of general knowledge complements human effort by raising its marginal return. The model highlights a sharp dynamic tension: while agentic AI can improve contemporaneous decision quality, it can also erode learning incentives that sustain long-run collective knowledge. When human effort is sufficiently elastic and agentic recommendations exceed an accuracy threshold, the economy can tip into a knowledge-collapse steady state in which general knowledge vanishes ultimately, despite high-quality personalized advice. Welfare is generally non-monotone in agentic accuracy, implying an interior, welfare-maximizing level of agentic precision and motivating information-design regulations. In contrast, greater aggregation capacity for general knowledge—meaning more effective sharing and pooling of human-generated general knowledge—unambiguously raises welfare and increases resilience to knowledge collapse.
No research papers available today.
AI Skill Now