Modelagem e machine learning: constrangimentos na autonomia técnica

Abstract

This dissertation analyzes how social and professional factors shape our technical autonomy in the production of Machine Learning systems, focusing on the influence of hierarchical structures, algorithmic cultures, and MLOps practices on our methodological choices and on algorithmic diversity. We adopt a convergent mixed-methods design: we conduct a quantitative analysis of 142 public GitHub repositories (78 MLOps; 64 academic) to measure the diversity of libraries and imports in Python code, and we carry out qualitative analyses of artifacts and documentation to interpret institutional patterns and technical vocabularies. Our findings reveal segregation between academic and corporate ecosystems: the former exhibits greater plurality and public visibility, while the latter shows higher levels of standardization, governance, and compliance requirements. We observe that standardized pipelines and MLOps platforms correlate with lower algorithmic diversity and reduced individual autonomy, whereas multidisciplinary teams, deliberative practices, and documentation mechanisms mitigate these effects. We highlight the role of boundary objects (frameworks and platforms) in consolidating technological hegemonies and in shaping the institutional rewards that guide our technical choices. We conclude that technical autonomy is socially constructed and hierarchically mediated, recommending participatory governance, robust documentation, cognitive diversity, and audits; we also suggest deepening research using internal corporate data and incorporating ethical training into certification programs.