Departmental Research Seminar

Not Worth the Effort: Generative AI and Effort Opacity in Collaborative Work

Generative AI helps individuals produce more, and sometimes better, work, yet much work is collaborative: one person creates work and another reviews it. Because effort is rarely observable, reviewers infer it from the work before deciding how much effort to reciprocate through review. I argue that AI creates effort opacity; it removes traces of effort from work, leading reviewers to reciprocate less. Critically, this argument predicts that reduced reciprocity cannot be explained by AI use alone: it should diminish when effort is legible, even if AI use remains known. Using a custom AI classifier on code from top GitHub projects with over 1 million stars and 2 million downstream dependencies, I find that AI-assisted work receives 25% less detailed feedback and has a 20% higher rejection rate, even after controlling for quality. A pre-registered study with software developers tests the effort opacity mechanism. Holding quality constant, I vary AI disclosure and effort legibility. Making effort legible closes the reciprocity gap between AI-assisted and unassisted work, even though AI use remains disclosed. This restoration operates through perceived effort rather than quality, negative attitudes toward AI, or perceived competence. These findings show that reciprocity depends on perceived effort, not just work’s value, and when technology creates effort opacity, work receives less of the review needed for its development and integration.

Cognitive Enrichment and Human Capital: Experimental Evidence from Developing Countries

Learning disparities persist in low-income countries despite substantial investment in education, suggesting access alone is insufficient for building human capital. Poverty constrains cognitive development, leaving children with deficits in executive functioning, attention, and working memory. Chess offers a promising, low-cost way to strengthen these abilities through structured, play-based practice in strategic thinking and foresight.
We report two randomized controlled trials in Kenya and Malawi. In Kenya, 551 Grade 4–5 students across six primary schools were individually randomized to a seven-week chess program spanning one academic term. Assignment to chess program increased composite cognitive performance by 0.14–0.16 standard deviations, with the largest gains in attention, visuospatial processing, and executive function. We find no detectable effects on academic or non-cognitive outcomes by the end of the term.
The Malawi experiment expanded implementation to 40 public primary schools and more than 2,100 Grade 4 students. Schools were first randomized to intervention or pure-control status. Within the 20 intervention schools, students were then randomized to a chess program, a tutoring program providing additional teacher-led instruction, or a no-program control group. Our primary estimates compare students within the same schools. Assignment to chess increased the primary cognitive index by 0.144 standard deviations and executive function by 0.122 standard deviations, closely replicating the Kenya result under substantially expanded implementation. Chess also improved planning by 0.128 standard deviations, numeracy by 0.112 standard deviations, and overall self-efficacy by 0.131 standard deviations. The largest self-efficacy gain occurred in the academic domain, where children’s confidence in their ability to succeed at school increased by 0.170 standard deviations. Tutoring improved numeracy by a similar magnitude, driven by gains in directly practiced arithmetic, but produced smaller changes in executive function and self-efficacy. Chess therefore generated a numeracy gain comparable to that from tutoring despite providing no direct mathematics instruction, while producing a significantly larger improvement in overall self-efficacy. Across both experiments, we find that foundational cognitive skills remain malleable during primary school and that a low-cost, infrastructure-light intervention can produce replicable gains across settings and scales. The results identify children’s capacity to learn as a distinct margin of human-capital investment and suggest that structured, cognitively demanding activities can complement conventional efforts to improve instruction where opportunities for stimulation are scarce.