ENGENHARIA DE PROMPT: Uma análise das "alucinações" em Inteligências Artificiais Generativas

Abstract

In recent years, the evolution of Generative Artificial Intelligence (GAI) and its advanced language model techniques has been remarkable. These advancements enable deep learning capabilities, allowing machines to increasingly approximate human cognitive abilities in creation and communication. At the heart of this revolution are structures that analyze vast volumes of data and perform hundreds of millions of mathematical calculations, obscuring a clear understanding of the logic applied in GAIs. Due to this complexity, these tools often produce "hallucinations"—responses that are out of context or contain unrealistic statements. This study aims to analyze "hallucinations" in the leading language models (e.g., ChatGPT 4.0, Gemini 1.5, Copilot, and Perplexity) available for free to the general public. Methodologically, the study is qualitative with an inductive logic approach. Comparative analyses were conducted among four language models (ChatGPT 4.0, Gemini 1.5, Copilot, and Perplexity) using prompts likely to induce "hallucinations". Our analyses highlight the distinct characteristics of the leading GAIs on the market. We believe that the combined use of these GAIs can lead users to more satisfactory results than using only one.