Levy flights in steep potential wells: Langevin modeling versus direct response to energy landscapes

Artificial intelligence systems that can match human scientific creativity remain elusive, according to new research that tested state-of-the-art language models on fundamental scientific tasks. optimistic predictions about AI's potential to accelerate scientific and highlight critical gaps in how these systems process and generate knowledge.

Researchers found that current large language models cannot generate genuinely novel scientific hypotheses or design experiments that produce new knowledge. When tested across multiple scientific domains including physics, biology, and chemistry, the AI systems consistently failed to propose ideas or experiments that went beyond what was already present in their training data.

The study evaluated several leading language models using carefully designed benchmarks that measured true scientific reasoning capabilities. The researchers created tasks requiring the AI to propose new experiments, generate testable hypotheses, and identify promising research directions—all core activities in scientific . The models were assessed on whether their outputs represented genuine novelty versus simply recombining existing information.

showed consistent failure across all tested models. In physics, the AI could not propose experiments to test new physical theories. In biology, it failed to generate hypotheses about unknown biological mechanisms. In chemistry, it couldn't design experiments to discover new chemical compounds or reactions. The models performed best when simply retrieving or rephrasing existing scientific knowledge, but consistently fell short when creativity and true innovation were required.

This limitation matters because scientific progress depends on generating new ideas and testing them through experimentation. If AI systems can only work with existing knowledge, they cannot drive the fundamental breakthroughs that advance human understanding. The research suggests that current approaches to AI development may be missing essential components of scientific reasoning.

The study acknowledges that while language models excel at processing and organizing existing information, they lack the capacity for the type of creative thinking that characterizes human scientific . This gap represents a significant for researchers hoping to use AI to accelerate scientific progress across multiple fields.