Inteligência Artificial e Informação Qualificada: Uma Abordagem Experimental para Fundamentar o Jornalismo de Prompt

Abstract

This research investigates the potential of Generative Artificial Intelligence (GenAI), particularly Large Language Models (LLMs), as tools capable of producing qualified information. It starts from the hypothesis that LLMs can discern information. To analyze this premise, a dataset comprising 1,200 pairs of entries was created, formed by news from the fact-checking agency Aos Fatos and responses generated by ChatGPT’s GPT-4o model, formulated using zero-shot and few-shot techniques, the latter with chain-of-thought (CoT). The analysis involved the application of semantic similarity measures, clustering methods (K-Means and Hierarchical Clustering), and non-parametric statistical tests (Sign Test), in order to assess the model’s performance regarding informational verification. The results revealed a high degree of correspondence between the pairs. Similarity medians ranged from 0.80 to 0.95, indicating that, in more than 90% of cases, GenAI responses presented good or near-total correspondence with the contents of human fact-checks. Few-shot CoT techniques, especially the one applied with the DeepSeek model, demonstrated greater critical rigor in the evaluation of semantic nuances. These findings corroborate the initial hypothesis, providing a basis for the substantiation of an emerging field in digital journalism: Prompt Journalism, or Prompt Jornalismo, is an area being founded within the algorithmosphere, situated in the mediation between humans and Generative Artificial Intelligence, dedicated to the process of production and consumption of qualified information through the set of discursive, dialogic, and technical practices, denominated Promptaction (“Promptação”).