For student teachers' professional development, the emergence of generative artificial intelligence (AI) represents both opportunity and challenge. This exploratory quasi-experimental study aims to investigate the effects of "Human-Human" and "Human-Machine" collaborative learning approaches on the STEM teaching training performance of student teachers. Twenty-three student teachers were divided into two groups within a single class, each adopting one learning method. The experiment lasted for two months with weekly three-hour sessions. Data were analysed focusing on critical thinking, learning performance, and cognitive load between the groups. The results indicated that student teachers using ChatGPT showed higher critical thinking systematicity, task completion efficiency, and experienced lower cognitive load. Student teachers paired with in-service teachers slightly outperformed those with ChatGPT on the final teaching design proposal. These findings underscore the potential and varying strengths of AI tools like ChatGPT and human teachers. For further research, refined collaborative learning scaffolding are recommended to explore the impact and potential of AI-assisted and in-service teacher-involved collaboration. The study's implications could guide educators, policymakers, and AI developers in optimizing the AI-enhanced collaborative learning strategies and shed light on the new formation of human-machine collaborative intelligence in the scope of education.