This project is ongoing. Descriptions, methods and results may be updated after further validation.
Research question
Was the first official institutional adoption of AI followed by measurable aggregate productivity gains in the state courts?
Motivation
Brazilian courts adopted AI tools at different paces since 2018. This variation over time makes it possible to test, with public administrative data, whether adoption shows up in aggregate productivity indicators — instead of assuming gains from use-case reports.
Data
- Annual panel of the 27 state courts (Justiça em Números, CNJ)
- Audited timeline of the first institutional AI adoption in each court
Methodology
- Court and year fixed effects
- Staggered difference-in-differences (estimators robust to treatment heterogeneity)
- Event studies and imputation estimator
Preliminary findings
Work in progress, version of 12 September 2026. Results are preliminary and have not yet been submitted to peer review. Do not cite as a final result. Check the latest version and the replication materials indicated on this page.
- At the current stage, no robust, statistically detectable average aggregate productivity gains were identified after the first institutional adoption of AI.
- This result does not rule out localised effects by tool, task or judicial unit, which aggregate indicators cannot capture.
Limitations
- Annual aggregate indicators are not very sensitive to gains concentrated in specific tasks.
- The official adoption date may differ from the date of effective use by units.
- A null result is not proof that AI does not work; it is the absence of a detectable average effect with these data.
Implications
- Evaluations of AI in the judiciary need task- and unit-level data, not only aggregate indicators.
- Audited adoption timelines are a reusable public input for other studies.
Planned applied products (possible outreach, subject to registration)
- Public, audited timeline of AI adoption by the state courts (open data, CC BY 4.0, in the repository)