Will AI Ever Win a Nobel Prize? (www.scientificamerican.com)

In 2016 Hiroaki Kitano launched the “Nobel Turing Challenge”: build an AI scientist that, working highly or fully autonomously, can generate hypotheses, design and run experiments, and analyze results well enough to make a Nobel‑worthy discovery — originally framed as achievable by 2050, though some researchers predict it could come much sooner. Recent milestones show AI accelerating science (AlphaFold’s protein-structure work and 2024 Nobel recognition for neural‑network pioneers), and projects like Coscientist (LLM-driven robotic chemistry), Sakana AI’s automated ML work, and Agents4Science (AI-authored papers and reviews) demonstrate growing end-to-end capabilities. Small wins include AI finding overlooked biological signals in published datasets and dramatically speeding up computational chemistry tasks that once took humans months. But significant technical and social hurdles remain. Benchmarks paint a mixed picture: an Allen Institute study of 57 agents found ~70% success on discrete science tasks but only ~1% success at taking a project from idea through experiment to analysis. Core gaps include hallucinations, lack of real‑world embodied experience, weak meta‑reasoning (the ability to evaluate and revise one’s own reasoning), and difficulty learning underlying scientific principles rather than surface patterns. Closing those gaps likely requires new research directions, larger investments, robotic embodiment, and rigorous human oversight — plus legal, ethical and reproducibility safeguards. For the AI/ML community the Nobel Turing Challenge crystallizes both a bold roadmap for autonomous scientific agents and a test bed for critical advances in autonomy, reasoning, and trustworthy deployment.