Inteligência artificial generativa na elaboração de peças processuais do controle externodiretrizes de governança para uma inovação responsável e segura

Resumo

Objective: This dissertation aims to propose governance guidelines for the ethical and responsible use of generative artificial intelligence (GenAI) in the drafting of procedural documents by Brazilian Courts of Accounts, with a focus on mitigating technical, ethical, legal, and organizational risks. Methodology: A qualitative, exploratory, and descriptive approach was adopted, grounded in the theoretical frameworks of adaptive governance, complex adaptive systems (CAS), and digital ethics. Data collection was conducted through a structured questionnaire submitted under the Brazilian Access to Information Law to all 33 Courts of Accounts in the country, resulting in 31 responses (93.9%). The data were systematized according to three analytical dimensions: technological adoption, practical application, and governance and control mechanisms. Results: The findings indicate a growing adoption of Large Language Models (LLMs) within the Courts of Accounts, albeit asymmetrically and with notable gaps in standardized guidelines, traceability mechanisms, and requirements for human oversight. Despite relevant initiatives – such as the INACIA system (TCU) and Chat-TCE (TCE-AM) – there remains a regulatory and institutional vacuum concerning the use of such tools in technical-decisional documents. Based on the triangulation of empirical data and theoretical foundations, the CRISP-LLM Gov framework was proposed. Inspired by the iterative and cyclical logic of the CRISP-DM model, it is structured into six interdependent phases focused on accountability, mandatory human review, traceability, and continuous institutional learning. Limitations: This study has a theoretical nature and was not empirically validated through practical application. The use of the Brazilian Access to Information Law as the data collection method limited the depth of qualitative responses and reflects a specific timeframe up to May 2025. It does not encompass informal practices or organizational cultures in their entirety. Practical implications: The proposed framework offers a structured and adaptive roadmap for building GenAI governance policies in the Courts of Accounts, contributing to safer, auditable, and more transparent public decisions aligned with the public interest. Contributions to society: By addressing the key risks associated with the use of LLMs in procedural documents that directly influence the decision-making process in external control, this dissertation contributes to strengthening institutional governance, the legitimacy of public decisions, and the protection of fundamental right – thus promoting a human-centered and publicly accountable approach to artificial intelligence. Originality: This research is original in combining an unprecedented nationwide empirical study on the use of LLMs in procedural documents with the development of a proprietary theoretical framework – CRISP-LLM Gov – designed to guide the adaptive governance of GenAI within the context of Brazilian Courts of Accounts.

Ficha do documento

Tipo

Dissertação

Ano

2025

Instituição

Fundação Getulio Vargas

Fonte

Repositório da FGV

Idioma

Português

Acesso

Acesso aberto

Identificador

oai:repositorio.fgv.br:10438/37555

Temas

Administração PúblicaTecnologiaGovernança

Palavras-chave

Inteligência artificial generativaGovernança adaptativaTribunais de contasPeças processuaisLarge language modelsGenerative artificial intelligenceAdaptive governanceCourts of accountsProcedural documentsAdministração públicaInteligência artificialGovernança pública