Generative artificial intelligence (GenAI) is rapidly expanding across sectors and has now become embedded within educational contexts, where it is actively reshaping teaching and learning practices, especially science education. Unlike traditional digital tools, GenAI systems, particularly large language models (LLMs), can generate tailored explanations, simulations and engage students in argumentation by generating counterexamples or hypotheses in natural language. These potentials offer promising opportunities for personalized learning and effective pedagogical support. However, GenAI tools raise significant cognitive, epistemological, and ethical concerns. Science education is more sensitive to these issues because learning scientific concepts and theories requires engagement in epistemic practices such as modeling, argumentation, and evidence-based reasoning. The integration of GenAI therefore requires a critical examination of its didactic potential and implications for the teaching and learning of science. This study adopts a critical analytical approach based on a review of recent peer-reviewed articles indexed in Scopus, Web of Science, and ERIC. Articles published between 2023 and 2026 were selected using keywords related to generative artificial intelligence, science education, and epistemic practices. The analysis is grounded in key theoretical frameworks in science education research, including didactic transposition, epistemic practices, pedagogical approaches, and cognitive regulation. The analysis indicates that GenAI can support science education through personalized learning pathways, intelligent tutoring systems and feedback, automated creation of learning resources, and assistance in instructional design. At the same time, the analysis underscores substantial challenges, including risks of cognitive offloading, epistemic dependency, algorithmic bias, and the risk of confusing generated content with genuine knowledge construction. The findings indicate that the educational value of GenAI depends on its pedagogical orchestration by teachers and its alignment with disciplinary learning objectives. Instead of replacing didactic approaches, GenAI should be regarded as a didactic mediator capable of supporting epistemic practices when integrated into carefully designed science learning environments.