With the rise of artificial intelligence, AI driven scientific discoveries have become a new paradigm in scientific research. However, whether the research path that relies on data and algorithms can truly constitute “scientific discoveries” still faces fundamental philosophical and methodological questions. Scientific epistemology requires that knowledge claims must undergo falsifiability testing. However, current AI scientific discoveries often appear as “black boxes” due to the lack of interpretability, with outputs mostly statistical predictions rather than logically complete rational claims, resulting in the failure of falsifiability mechanisms: there are no clear theoretical targets to criticize, and scientific debates have become superficial debates on input-output. This reveals a deeper crisis: the traditional scientific cognition based on human intuitive reasoning is difficult to adapt to the complexity patterns revealed by AI. Therefore, resolving this crisis cannot only pursue the transparency of AI, but also requires promoting the expansion of the cognitive boundaries of the scientific community - developing new conceptual systems and formal languages that represent complexity, so as to understand the complex world revealed by AI while maintaining the fundamental mechanisms of scientific theory criticism and evolution.