What Research Is Suitable for AI for Science—and What Is Not

AdAI for Science is rapidly automating research activities ranging from literature search and hypothesis generation to experimentation, analysis, and paper writing. Yet treating “Can AI autonomously conduct scientific research?” as a single question conflates research settings with different epistemic structures. This paper distinguishes Type I research (derivable or statistically continuous discovery), in which candidate hypotheses can be generated and evaluated relatively continuously from an existing search space and evaluation criteria, from Type II research (transformative or distribution-distant discovery), in which initial evidence is weak and potentially valuable hypotheses lie far from current knowledge distributions or evaluation axes. In Type I settings, rapid generation, evaluation, rejection, and re-exploration are major strengths of AI. In Type II settings, however, a valuable hypothesis may disappear under low initial evaluation before it has been sufficiently developed.I therefore define Hypothesis Persistence as a function distinct from hypothesis generation, and Premature Hypothesis Abandonment as the loss of a potentially valuable hypothesis before it becomes adequately evaluable. I further propose Human-Anchored Hypothesis Persistence + AI Peripheral Exploration: a research configuration in which a researcher serves as an exploration anchor by preserving the semantic identity of a hypothesis core and keeping it reconnectable to new concepts, evidence, theories, technologies, or cross-disciplinary links, while AI performs large-scale exploration around that anchor. This is not a claim of general human superiority. It is a design hypothesis that AI Autonomy Suitability varies across research problems and research stages. Recent work including AutoResearchEval, Anthropic’s Automated Alignment Researchers, and HypoForge supports the importance of decomposing research processes and distinguishing differences in evaluability, feedback, and supervision across stages.