Hypothesis generation plays a central role in scientific discovery, yet it has remained one of the least formalized and least automated components of the research process. While advances in automation, machine learning, and self-driving laboratories have transformed how hypotheses are tested, their formulation has largely remained a human endeavor. Recent progress in AI, particularly LLMs, challenges this assumption by enabling the large-scale generation of novel, plausible, and actionable hypotheses through computational recombination of existing knowledge. We argue here that coupling AI-driven hypothesis generation with agentic reasoning systems and autonomous laboratory platforms opens a realistic pathway toward end-to-end automated scientific discovery. We discuss what defines a good scientific hypothesis, assess the opportunities and limitations of contemporary AI systems, and outline how hypothesis generation can be integrated into closed-loop experimental workflows. We conclude by identifying key challenges that must be addressed for such a Hypothesis-Driven Autonomous Discovery System (HADS) to truly accelerate material science.
T. Jesper Jacobsson (2026) studied this question.