Why has AI not made original scientific discoveries? princeton scholar highlights the need for verifiable infrastructure

On September 10, during the 2026 Inclusion · Bund Conference, Princeton University's Director of the Center for AI Innovation, Wang Mengdi, questioned how close AI is to autonomously discovering new science. She argued that while large models excel at finding the most likely answers, they struggle with true scientific discovery, which often lies in the long tail of probability distributions. Wang Mengdi's team conducted a 'Simulated Life' experiment with models like ChatGPT and Claude, revealing that these models overestimate common scenarios and underestimate minority ones. Wang Mengdi explained that fields like mathematics and coding have clear validators, such as compilers and formal proof systems, enabling iterative learning and correction. In contrast, scientific research involves complex experiments and collaboration, making verification difficult. "Ideas are cheap. You need to show me code or an experiment," she said. Her team is building infrastructure to bridge this gap, including an automated experimental platform for quantum material research and an intelligent operating system called LabOS for research labs. Wang Mengdi emphasized that AI's potential to become a scientific discoverer depends on infrastructure that connects hypotheses, validation, traceability, and replication. "Only when the experimental process in the physical world can be systematically recorded, continuously validated, and repeatedly replicated can AI possibly transform the 'likelihood' learned in training into exploration of 'possibility,'" she concluded.