Publications
We produce independent audits and public metrics to assess transparency, risk and bias of AI models in the Brazilian context. Here are the technical reports resulting from this work.
Err or Withhold? Apparent Fairness Through Omission When LLMs Decide Under Stigma
Rodrigo Cavalcanti • Julho, 2026
This work evaluates how seven artificial intelligence models each respond to 7,626 scored everyday situations (from a 7,653-prompt instrument with 27 neutral controls) involving people marked by social stigma. The analysis distinguishes prudence, systematic refusal and the reproduction of bias. The results show that apparently safe models may only be avoiding decisions, and that performance varies according to the type of stigma and the context presented.
The living and continuous version with more than 15 evaluated models is available in the Avalia section.