Advancements in AI Research and Social Responsibility : From Technology to Control and Governance
Drawing upon recent academic publications and reports, this paper systematically examines current developments in artificial intelligence (AI) through a five-stage framework encompassing technology, applications, expectations and reality, risks and safety, and control and governance. Based on this analysis, the paper identifies key challenges and proposes future directions for AI research and development. Following the emergence of ChatGPT, large language models (LLMs) have rapidly proliferated. Their core component, the transformer, has expanded beyond natural language processing into diverse domains such as computer vision (Vision Transformer, ViT) and multivariate time-series analysis (Time Series Transformer, TST). Furthermore, LLMs are evolving into agent-based systems and are being applied to the automation of scientific research, as exemplified by Google’s Co-Scientist, AlphaEvolve, and AlphaGenome. These developments have significantly heightened expectations regarding the transformative potential of AI across society. However, a gap remains between these expectations and the current state of technology, as structural issues such as data bias and hallucination continue to pose substantial risks. In particular, hallucination is interpreted as a phenomenon arising from the model’s tendency to maximize expected evaluation outcomes, and it is identified as a critical challenge for ensuring AI safety. Accordingly, the safe deployment of AI requires effective monitoring of model behavior and improved interpretability of chain-of-thought (CoT) reasoning processes, red-teaming activities at both macro- and micro-levels, and the establishment of international governance frameworks, including those in the healthcare domain such as guidelines from the World Health Organization (WHO). In conclusion, while AI is driving profound changes not only in science and technology but also across society as a whole, addressing technical challenges— such as mitigating hallucination, preventing catastrophic forgetting in continual learning, and improving data efficiency—must be accompanied by the development of control and governance systems aligned with human values. In particular, international governance initiatives are needed to reduce disparities between countries and address polarization at the global level.