Generative AI and the Productivity Divide: Human-AI Complementarities in Education and Knowledge Work
Citation. Idan, L., & Anand, B. (2026). Generative AI and the productivity divide: Human-AI complementarities in education and knowledge work (Working paper). https://doi.org/10.2139/ssrn.6624038
Summary. In a randomized Texas A&M self-study experiment, LLM access raised post-intervention LLM-knowledge scores relative to traditional resources, but gains varied by AI Interaction Competence (AIC): eliciting, filtering, verifying, and applying model outputs. The authors report that a conceptual-roadmap scaffold modestly improved novice outcomes and reduced score dispersion, while additional required study time did not. The paper frames access, interaction skill, and workflow structure as complementary inputs rather than treating GenAI as a uniform productivity intervention.
Key claims
- Among completers, the LLM group averaged .56 on the post-intervention exam versus .48 in the no-LLM group; the controlled treatment coefficient was positive at p < .05.
- GPA and pre-intervention topic knowledge did not significantly moderate the LLM treatment, whereas the Treatment × AIC interaction was positive at p < .05.
- For novices, the conceptual-roadmap condition averaged .45 (SD = .14) versus .38 (SD = .22) for unguided LLM use; the reported difference and AIC interaction were weakly significant (p < .10), so the equalizing interpretation is suggestive rather than definitive.
- Requiring novices to study four rather than three hours per day did not produce a significant performance gain over unguided LLM use.
Evidence & limitations
- Randomized, short-horizon self-study experiment with 179 Texas A&M students, primarily engineering majors; it is a proxy for early-career knowledge work rather than workplace productivity evidence.
- The outcome is an LLM-knowledge exam after three days, not accepted work quality, organizational performance, durable transfer, or longitudinal capability development.
- AIC is central to the paper’s account, but its measurement and discriminant validity relative to AI literacy, prior knowledge, confidence, and general learning skill need independent validation.
- The Drive binary exactly matches the public arXiv v1 PDF (SHA-256 de02659b2a57a291164ed400154e64552e3cec06f5a2deaaf27f7c44a118a7d8); Crossref identifies the matching SSRN working-paper DOI as canonical provenance.