In April this year, I wrote an article to summarize some thoughts I have had about AI's influence to scientific research. I think we are in a period of fundamental paradigm shift, which I name the "agentification of scientific research". This article was posted as arXiv:2604.14718
A key point I would like to emphasize is that the AI revolution should not be understood only as automation. Its deeper meaning is that human know-how, not just explicit knowledge, is becoming easier to carry, reproduce, and share. DNA once made biological information inheritable, and human language made cultural information transferable. Large language models now represent another transition in the dynamics of information on planet earth, because they begin to encode practical patterns of reasoning, judgment, explanation, and workflow that previously had to be learned through close apprenticeship.
For science, this matters because much of research is tacit. Papers record final results, methods, and selected evidence, but they rarely preserve all the failed attempts, intermediate decisions, debugging habits, and contextual choices through which a result was actually produced. If AI agents can work inside real research workflows, use the tools scientists use, and learn from the process itself, they may gradually move from assistants for routine tasks to collaborators that help carry scientific experience across people, fields, and projects.
I call this gradual transformation the agentification of scientific research. It begins with tool use and automation, but its more important consequences may appear in collaboration and publishing. In the long run, a scientific result may be published not only as a static paper, but also with an interactive research agent that can explain the work, reproduce parts of it, answer questions at different levels, and help others build on it. Such agentic publishing would try to preserve both a stable archival record and a more flexible interface to the know-how behind the record.
Scientific research also requires new capability of AI systems. In particular, original discovery depends on diversity of ideas: different tastes, intuitions, and choices about what is worth pursuing. If AI systems simply reproduce the dominant patterns of their training data, they will remain useful but not deeply creative. The future of AI for Science therefore depends not only on stronger models, but on building open systems where human researchers and AI agents can learn from one another through real scientific work.