{"doc_desc":{"idno":"DDI_PER_2025_GAITS_v01_M","producers":[{"name":"Development Data Group","abbr":"DECDG","affiliation":"World Bank Group","role":"Documentation of the survey"}],"version_statement":{"version":"Version 01 (July 2026)"}},"study_desc":{"title_statement":{"idno":"PER_2025_GAITS_v01_M","title":"Not Faster, but Better: Generative AI and Teacher Skill in Peru 2025","sub_title":"Baseline and Endline Surveys","alternate_title":"GAITS 2025"},"authoring_entity":[{"name":"Carla Z. Glave","affiliation":"University of Wisconsin\u2013Madison"},{"name":"Carolina Lopez","affiliation":"Development Research Group, World Bank"},{"name":"Ezequiel Molina","affiliation":"World Bank"},{"name":"Rom\u00e1n Andr\u00e9s Z\u00e1rate","affiliation":"University of Toronto"}],"production_statement":{"copyright":"\u00a9 World Bank","funding_agencies":[{"name":"Human Capital Project (World Bank)","role":"Primary Founder"},{"name":"Microsoft AI Economy Institute"}]},"distribution_statement":{"contact":[{"name":"Carolina Lopez","affiliation":"Development Research Group, World Bank","email":"carolina_lopez@worldbank.org"}]},"series_statement":{"series_info":"This dataset accompanies the paper \"Not Faster, but Better: Generative AI and Teacher Skill in Peru\" (Glave, Lopez, Molina & Zarate). The data come from a randomized controlled trial conducted in 390 public primary schools in Metropolitan Lima, Peru, during the 2025 school year. The study evaluated a program providing sixth-grade teachers with a one-day training on generative AI, a Microsoft Copilot license, and weekly WhatsApp follow-up; the message sender (education authority vs. peer teacher) and content (general vs. teaching-at-the-right-level) were cross-randomized. The dataset is a teacher-level file combining administrative rosters, training attendance, a baseline survey, weekly Copilot usage records, and an endline survey with blind-graded prompt-quality measures for 1,269 teachers. Pre-registered as AEARCTR-0016336; IRB protocol #48010 (University of Toronto)."},"version_statement":{"version_date":"2026-07-03"},"study_info":{"abstract":"Generative AI is being deployed across the public sectors of developing countries on the premise that it makes workers faster. We test this in a randomized trial with 1,269 sixth-grade teachers across 390 public schools in Lima, Peru. A low-cost program\u2014a one-day training, an AI license, and eight months of light-touch follow-up\u2014raised adoption sharply yet saved teachers no time on any of their core tasks, with precisely estimated nulls throughout. The gains lay elsewhere: treated teachers wrote markedly better prompts, including on a rubric designed independently of the training, and were sharper at catching weak AI outputs. These gains were largest for older teachers, who began furthest behind, narrowing rather than widening skill gaps. Taken together, the results indicate that AI training primarily develops the skills needed to generate and evaluate high-quality AI outputs, rather than reducing the time teachers spend on their work.","coll_dates":[{"start":"2025-04-10","end":"2025-05-11","cycle":"Baseline post training"},{"start":"2025-04-01","end":"2025-12-11","cycle":"Copilot usage records"},{"start":"2025-12-01","end":"2025-12-31","cycle":"Endline"}],"nation":[{"name":"Peru","abbreviation":"PER"}],"geog_coverage":"School (instituci\u00f3n educativa); randomization stratified by UGEL (local education district)","analysis_unit":"Teachers","universe":"Sixth-grade primary school teachers in public schools in Metropolitan Lima, Peru, during the 2025 school year. The sample comprises all 390 public schools in Metropolitan Lima that offer sixth grade of primary education (196 treatment, 194 control) and their 1,269 sixth-grade teachers.","data_kind":"Sample survey data [ssd]","notes":"The dataset covers: program take-up (training attendance, Copilot license activation and weekly usage); teachers' use of AI for professional and personal tasks; time allocation across seven core teaching activities; blind-graded quality of AI prompts (training-based and general rubrics); teachers' ability to evaluate AI-generated output (prompt scenarios); beliefs about AI, equity, and willingness to pay; a discrete choice experiment on hypothetical AI-use scenarios; and baseline teacher and school characteristics."},"method":{"data_collection":{"data_collectors":[{"name":"World Bank research team, in coordination with DRELM"}],"sampling_procedure":"The sample is a census of the 390 public schools in Metropolitan Lima offering sixth grade of primary education. Treatment was assigned in two stages. First, eligibility was randomized at the school-network level: 84 of 143 administrative networks were selected, re-running the randomization 1,000 times and keeping the allocation that maximized the minimum p-value across baseline school characteristics. Second, treatment was randomized at the school level within eligible networks, stratified by UGEL (local school district) and by whether baseline school math performance was above or below the median, using the same re-randomization procedure. Two components were cross-randomized within the treatment group: the sender of the follow-up messages (education authority vs. peer teacher) and their content (general support vs. teaching-at-the-right-level).","coll_mode":["Internet [int]"],"research_instrument":"Baseline Teacher Survey (Spanish, Baseline); Endline Teacher Survey (Spanish, Endline)","coll_situation":"Survey data were collected through self-administered online questionnaires. The baseline survey was administered after the March 2025 training sessions through a form on DRELM's teacher platform. The endline survey was administered in Qualtrics, first during in-person sessions and subsequently distributed online to maximize response rates; the survey modality of each respondent (endline_online) is recorded in the dataset and controlled for in all survey-based analyses. In addition to survey data, the dataset incorporates administrative sources: school and teacher rosters from DRELM, attendance records from the March 2025 training sessions, and weekly Microsoft Copilot license usage reports . Open-ended prompt responses from the endline were blind-graded by evaluators using two rubrics; only the graded scores are included in the dataset.","weight":"No weights are used. All estimates are unweighted; regressions include randomization strata fixed effects.","cleaning_operations":"Data from the five sources were merged at the teacher level. De-identification: direct identifiers (names, national IDs, e-mails, phone numbers), Qualtrics metadata (IP, geolocation, timestamps), open-text responses, granular geography, and school-level variables matchable to the public school census were removed; school, district and teacher identifiers were replaced with anonymous codes assigned with fixed random seeds. Missing covariates are flagged and imputed within the analysis code."}},"data_access":{"dataset_use":{"conf_dec":[{"txt":"The data were collected under IRB protocol #48010 approved by the University of Toronto Research Ethics Board. All participants provided informed consent. Data have been de-identified: names and direct identifiers have been removed. Respondents are adult public school teachers.","required":"yes"}],"contact":[{"name":"Carolina Lopez","affiliation":"Development Research Group, World Bank","email":"carolina_lopez@worldbank.org"}],"cit_req":"Glave, Carla Z., Carolina Lopez, Ezequiel Molina, and Rom\u00e1n Andr\u00e9s Z\u00e1rate. \"Not Faster, but Better: Generative AI and Teacher Skill in Peru.\" Working paper.","conditions":"Access to this dataset requires signing a Data Use Agreement (DUA) with the authors. Users may not attempt to re-identify study participants, share the data with third parties, or use the data for purposes other than research. Publications using these data must acknowledge the original study and cite the accompanying paper.","disclaimer":"The user of the data acknowledges that the original collectors of the data, the authorized distributors of the data, and the relevant funding agencies bear no responsibility for use of the data or for interpretations or inferences based upon such uses. The findings, interpretations, and conclusions expressed in any work using this dataset are entirely those of the authors and do not necessarily represent the views of the World Bank, its affiliated organizations, or the governments they represent."}}},"schematype":"survey","tags":[{"tag":"DOI"}]}