AI has far-reaching potential not just in society, but also in science and research. How can its recent success be captured and described and, then, explained? This paper focuses on objects and fields of research; it is not limited to AI methods and techniques, as much contemporary philosophy of AI does. My question is whether the object systems of various scientific disciplines in which AI methods are successfully used do not have fundamental similarities? The answer and main thesis is that the common characteristic of these objects lies in their dynamical complexity. The latter is generated by underlying instabilities. AI can be seen as central to complex systems research and complexity science. Four case studies from climate research, neuroscience, astrophysics and biomedicine support the thesis. It will be shown why established traditional (classical-modern) sciences, i.e. research before AI, had comparatively limited access to dynamical complex objects. While limitations and blind spots cannot be fully overcome by AI-assisted sciences, these novel sciences, which can be called late-modern, enable new approaches and display impressive success in some respects (e.g., predictions and classifications). The paper concludes that the turn towards a late-modern regime of science is ambivalent.