Leila's Research

Leila Aghabarari

Do expert-based predictions of occupational AI exposure accurately forecast actual AI adoption and task performance? Using the GAIA Index — a harmonized dataset of 923 O*NET-SOC occupations linking three independent pre-shock exposure measures to post-shock behavioral outcomes — this paper provides the first large-scale empirical test of the predictive validity of occupational AI exposure scores. The three predictors are the AI-rater and human-rater automation scores from Eloundou et al. (2023) and the Suitability for Machine Learning (SML) score from Brynjolfsson, Mitchell & Rock (2018). The behavioral outcome is the Anthropic Economic Index (AEI, February 2026), which provides actual task success rates, AI autonomy rates, and work-relatedness shares for 628 of the 923 occupations. All three exposure predictors fail to forecast 2026 task success: bivariate correlations range from −0.13 to −0.04, and OLS coefficients with group fixed effects are statistically indistinguishable from zero. Against AI autonomy rates, exposure predictions are negative and significant (AI Alpha: β = −1.03, p < 0.001), revealing a structural inversion — the occupations rated as most susceptible to full automation exhibit the lowest autonomous Claude task completion. Selection-corrected Heckman estimates confirm these patterns after adjusting for non-random AEI adoption. Group-level analysis across 22 occupation categories documents systematic inversions: Food Preparation workers average 6.3% predicted exposure but 89.7% task success (+83 percentage points), while Computer & Mathematical workers average 52.1% predicted exposure but only 69.0% task success. Mechanism tests support complementarity as the primary driver: among physical occupation groups, the negative Alpha–success relationship strengthens (β = −2.18, p = 0.021), consistent with low-exposure workers migrating to adjacent, AI-tractable tasks outside their scored task bundle. These findings imply that exposure scores measure automation potential but not AI usefulness, and that workforce transition programs targeting high-exposure occupations may systematically misallocate resources.

Leila Aghabarari · Bernardus Van Doornik (Banco Central do Brasil)

We study how AI exposure affects firm credit outcomes and employment using detailed credit registry data from Brazil. We construct a sector-level AI exposure index from three independent sources and crosswalk it to Brazilian firm sectors. Exploiting the November 2022 generative AI shock in a difference-in-differences framework with 16 years of pre-treatment data, we find that firms in high-AI-exposure sectors face tighter credit conditions post-2022, with effects concentrated in micro and small firms and partially offset by government bank lending. We further trace how credit tightening transmits to employment and wages, documenting a financial accelerator channel for AI-induced labor market adjustment.

Leila Aghabarari · Sophia Chen (IMF) · Deniz Igan (IMF & CEPR) · Bernardus Van Doornik (BCB)

We study how fiscally backed credit interventions affect monetary policy transmission using detailed credit registry data from Brazil. Government direct credit weakens pass-through to loan rates and offsets contractions in private credit, especially during tightening, while supporting SMEs. Government-subsidized earmarked credit interacts with lending relationships to shape pass-through, consistent with a relationship-based hold-up channel. These effects are strongest during tightening and largely muted during loosening.

Leila Aghabarari · Andre Guettler · Mahvish Naeem · Bernardus Van Doornik (BCB)

We exploit the 2008/09 financial crisis as a shock to Brazilian banks and analyze credit union lending behavior. Credit unions tightened members' credit access less than other bank types (insurance effect). The labor market impact is positive for very small firms — micro firms with higher pre-crisis CU lending increased employment and paid higher wages during the crisis. Published in Economic Inquiry, Vol. 59, pp. 1215–1233.