O Uso de Processamento de Linguagem Natural e Inteligência Artificial para Análise da Determinação da Significância dos Impactos em Estudos de Impacto Ambiental

Environmental Impact Assessment (EIA) is a fundamental instrument for integrating
environmental considerations into decision-making. However, its effectiveness faces persistent
challenges, particularly regarding the Determination of Impact Significance, widely recognized
as the core of the EIA process, yet frequently lacking systematization and transparency in
Environmental Impact Statements (EIS). This research investigated the potential of Artificial
Intelligence (AI), through Natural Language Processing (NLP) and Large Language Models
(LLMs), to support the documentary analysis of how the Determination of Significance is
addressed in Brazilian EIS. The methodology involved a systematic literature review that
resulted in a set of six key assessment questions for evaluating significance practices in EIS;
the empirical analysis of 35 Brazilian EIS using this instrument; the development of an
experimental system based on Retrieval-Augmented Generation (RAG) with two commercial
LLMs (GPT-4o and Claude 3.5 Sonnet); and a comparative evaluation between model-
generated responses and human reference analysis. The empirical analysis revealed significant
gaps in Brazilian practice, with only 22 of the 35 EIS providing a definition of what they
consider significant and 16 cases where it was not possible to verify the coherence between the
adopted definition and the judgments made. The system demonstrated compatibility with
human analysis for objective questions, with Claude 3.5 Sonnet showing better performance in
identifying formal definitions and conceptual structures. Limitations were identified, including
GPT-4o's tendency to infer beyond the explicit text and Claude's generation of inconclusive
summaries in some cases. It is concluded that NLP and LLMs can support the documentary
analysis of the Determination of Significance in EIS, with potential for initial screening and
information retrieval, without replacing specialized human analysis, which remains
indispensable when documents lack transparency in their judgment criteria.