O uso da inteligência artificial (ia) para elaboração de minutas de sentença em processos dos Juizados Especiais Federais do Maranhão: estudo comparativo de qualidade, sob a perspectiva de Robert Alexy, em distintos cenários de utilização

Abstract

This dissertation empirically investigates the argumentative quality of sentences drafted by generative artificial intelligence compared with human judgments in the Federal Small Claims Courts of the Maranhão Judicial Section (JEFs/SJMA). The experimental design compares, case by case, the original human judgment with three synthetic sentences generated by Gemini 2.5 Pro across three scenarios: Free (open prompt), Directed (outcome indicated), and Controlled (structured report and normatively anchored prompt). The stratified sample comprises 300 cases covering banking civil liability, continued benefit for the elderly (BPC-I), and continued benefit for persons with disabilities (BPC-PCD). The 1,200 sentences were submitted to triple independent coding by a blind human evaluator, Claude Opus 4.5, and GPT-4.5, under an Alexy-derived six-dimension rubric (D1–D6), totaling 3,600 codings. Friedman, Wilcoxon, and Bonferroni tests demonstrate that AI-generated sentences present greater textual conformity with the D1–D6 rubric standards than human sentences, with very high statistical significance and very large effect size (Friedman χ²(3) > 308; p < 10⁻⁶⁶), with an identical hierarchy across all three independent evaluators: free > directed > controlled > human. The correlation between the human evaluator and Claude Opus 4.5 (r = 0.936) provides robust evidence of instrument calibration. The largest differences concentrate in D3 (empirical external justification) and D5 (argumentative saturation), the dimensions most frequently weakened in mass litigation under Article 489, §1, of the Brazilian Civil Procedure Code. In the banking stratum (n=18), the human sentence outperforms the Directed and Controlled scenarios across all evaluators; the Free scenario, however, surpasses the human sentence in two of three evaluators, a finding that prevents generalizing human advantage in principiological matters to all AI modes. The automated verification workflow confirms 75.5% of suspicious marks in human sentences as real material inaccuracies (209/277), a pattern consistent with reused boilerplate judgments. The dissertation concludes that the integration of generative AI into mass litigation adjudication is empirically sustainable, provided it operates through protocols differentiated by argumentative complexity, namely controlled mode for rule-based and evidentiary matters and open or auxiliary mode for principiological matters, with qualified human review as a necessary and non-waivable condition. Keywords: