ANÁLISE COMPARATIVA DE SITES DE DETECÇÃO DE TEXTOS GERADOS POR INTELIGÊNCIA ARTIFICIAL

Abstract

This paper presents a comparative analysis of the performance of three tools that detect texts generated by Artificial Intelligence, namely ZeroGPT, EditPad, and QuillBot, focusing on the Brazilian Portuguese language. The motivation for the research comes from the rapid popularization of models such as ChatGPT and the consequent difficulty in distinguishing human writing from synthetic writing, a challenge that directly impacts the educational and scientific environment. To conduct the study, we selected the most relevant detectors with support for the chosen language and built a database with 404 texts: half composed of essays that received a perfect score of 1000 on the ENEM (2018–2024) and the other half generated by AI, following the same themes and rules as the ENEM. The evaluation was automated via Python with the Selenium framework, allowing the extraction of precise metrics such as sensitivity, specificity, and linguistic patterns (use of verbs and adjectives). The results indicated QuillBot as the most effective tool, achieving 95.3% accuracy and the highest sensitivity in identifying AI. ZeroGPT also proved reliable, with 93.81% accuracy. On the other hand, EditPad performed unsatisfactorily, correctly identifying only 50% of cases and failing to distinguish between texts, classifying almost everything as human. Linguistically, it was noted that AI tends to overuse adjectives and use fewer verbs, although the repetition of words is similar to that of humans. It is concluded that, although competent detectors exist, they are not infallible and should serve only as support for human analysis.