A collaborative research team comprising experts from Google, Google DeepMind, and MIT FutureTech has unveiled a report that evaluates the influence of AI on real-world scientific research settings. The report highlights a notable duality of AI in scientific research: On one front, AI is seamlessly woven into the entire scientific research process. General large-scale models and specialized scientific AI systems work in tandem, substantially expediting the digitization of scientific research. This integration saves researchers an average of nearly 7 hours per week, thus boosting research productivity. On the other hand, research bottlenecks are progressively migrating downstream in the workflow. While AI can swiftly generate a multitude of hypotheses awaiting verification, non-digital processes such as experimentation and real-world data collection cannot be accelerated at the same rate, imposing new capacity limitations. Moreover, AI has introduced a 'verification tax' for scrutinizing model outputs and even fostered a bias in research topic selection towards easily verifiable domains, resulting in a 'streetlight effect.' At present, the structural contradictions of AI for Science are becoming increasingly evident, and the next round of competition will hinge not only on model generation capabilities but also on the efficiency of physical experimental validation. To tackle this issue, both industry and academia have initiated the deployment of novel scientific research infrastructures, such as automated cloud laboratories, to bridge the gaps in the AI-driven research closed loop.