Atlas of the Errômetro: An Open GenAI-Built Resource for Learning, Questioning, and Testing Election Polls Against the 2026 Ballot Box in Higher Education

Source abstract

In Brazil's 2026 presidential first round, the ballot box contradicted most polls: 136 of the 161 national polls showed Lula da Silva ahead, but Flávio Bolsonaro finished first. Public debate swung between fraud claims and the margin of error, and no open resource let students verify either with real data. Following a design and development approach, this study used generative AI (GenAI) under human curation to build the Atlas of the Errômetro, an open educational resource in which each section lets users learn, question, and test how registered polls compared with the official result. Data on 1,297 registrations came from the Superior Electoral Court, results were located through documentary search, and automated programs gathered report status and 130 compliance audits. The 901 included polls were compared with the ballot using the declared margin, the A index, the audit category, and a retrospective revealed-efficiency measure. Only 5.9% of the 896 comparable polls fell within their margin (31.5% in the last three days), against about 92% expected under unbiased random sampling; errors were systematic, with 530 nominally significant deviations toward Lula against 82 toward Flávio; Compliant or Caution polls erred less (7.1 versus 8.1 percentage points) but leaned toward Lula as often; and declared cost showed no detectable association with accuracy, with random-sample equivalence of about 2.3%. Each finding became an inquiry users can repeat, with activities, exam-style items, and games verified by automated tests. This first design cycle shows feasibility; effects on learning remain to be evaluated.