Inquiring into AI Models: GPT-4, Sabiá-3 and Deepseek-R1 on Historical News ArFcles on Brazil’s Military Dictatorship

Abstract

This study analyzes how large language models (LLMs)—ChatGPT (OpenAI), Sabiá-3 (MaritacaAI), and DeepSeek (China)—interpret Brazilian historical texts from the military dictatorship (1964–1985). Using a curated dataset of sixteen OCR-processed newspaper articles pre-classified by stance (supportive, critical, or neutral), the study selected three representative texts for detailed analysis. It evaluates model responses to a political sentiment prompt through cosine similarity, lexical frequency, and class-based TF-IDF (c-TF-IDF). Results show semantic convergence among the models but notable divergence in discursive strategies. ChatGPT tends to neutralize ideological framing, DeepSeek focuses on rhetorical deconstruction, while Sabiá-3 favors literal interpretation and often misses implicit criticism or historical nuance. These patterns are linked to model architecture, linguistic design, and especially the quality and curation of training data. Despite being a Brazilian model, Sabiá-3 demonstrates interpretive shortcomings, suggesting that domestic training data alone is insufficient without critical historiographic input. The findings highlight concerns about algorithmic bias, digital sovereignty, and the cultural politics embedded in AI. The research underscores the importance of historically informed and critically curated corpora in training LLMs for underrepresented languages. Future work will broaden the dataset and explore how prompt design and model versioning affect interpretative outcomes.