Professor Ding Wenxuan Publishes Generative AI Study in Nature

Source:emlyon business schoolDate:2025-04-29

Ding Wenxuan
Professor of Artificial Intelligence and Business Analytics
DBA Advisor at emlyon business school


Professor Ding Wenxuan holds a Ph.D. in Cognitive Science and Information Technology from Carnegie Mellon University (USA).
She serves as an expert in artificial intelligence for both the French Academy of Sciences and the European Science Foundation.
Her research focuses on frontier basic AI theory and interdisciplinary innovation in AI applications, including AI for scientific discovery, smart healthcare, risk analysis and decision-making.
Her academic work has been published in top-tier international journals and premier conferences in the field of computer science.


Professor Wenxuan Ding from emlyon business school recently published a groundbreaking research paper in the internationally renowned journal Nature.

(Original title: Generative AI lacks the human creativity to achieve scientific discovery from scratch: https://www.nature.com/articles/s41598-025-93794-9).

She also presented her findings at the AAAI 2025 Spring Symposium held at Stanford University in March 2025, sparking widespread attention across both academia and industry.

This study provides an in-depth analysis of the capabilities and limitations of generative AI for scientific discovery, offering forward-looking insights on the future of AI for Science and AI-empowered Scientific Discovery.

Scientific discovery involves some of the most creative and complex forms of human reasoning, a capacity traditionally considered unique to the human brain. With the advent of generative AI, new possibilities have emerged for creativity, prompting scientists to question whether generative AI could achieve disruptive scientific breakthroughs comparable to those made by humans. Opinions, however, have remained mixed.

Breaking away from conventional perspectives, Professor Ding’s research investigates whether and how generative AI can independently formulate original scientific hypotheses, design goal-driven experiments, and make disruptive discoveries, as well as what scientific discoveries generative AI can make.

In the study, generative AI was assigned the role of a "scientist" and tasked with achieving a Nobel Prize-level discovery in molecular genetics using a semi-automated computational molecular genetics lab.

The findings reveal that, although current generative AI models (such as ChatGPT) excel at summarizing, synthesizing, and suggesting based on existing knowledge, they lack the human-like abilities of intuition, curiosity, and sudden insight necessary for true "zero-to-one" scientific exploration.

Generative AI tends to favor only incremental discoveries based on existing datasets, rather than producing breakthroughs from scratch. The study also discovered that current generative AI systems exhibit overconfidence in their discovery processes, mistakenly believing that they have made a completely successful discovery — a cognitive bias that poses challenges to the rigor of scientific research.

As such, current generative AI is mainly good at discovery tasks where the domain knowledge is already structured in a machine-executable form or where human-established knowledge bases are accessible.

Beyond delineating the boundaries of generative AI in scientific discovery, the study also proposes strategies to address its current limitations, including potential ethical and bias-related challenges.

It emphasizes that combining AI’s computational power with human intuition, curiosity, and insight will be critical to achieving transformative advances in scientific exploration.