Accelerating Science: How AI Predicts and Expands Discoveries

In a groundbreaking study published in Nature Human Behaviour, researchers from the University of Chicago have delved into the potential of artificial intelligence (AI) to predict and accelerate scientific discoveries. Led by Professor James A. Evans, the team developed models that could predict human inferences and the scientists making them, and generate 'alien' hypotheses—scientifically promising ideas that humans might not consider until the distant future, if at all. This innovative approach suggests that AI, when made aware of human activities and expertise, can complement collective human capacity by exploring uncharted territories, thereby accelerating scientific progress and moving beyond the contemporary scientific frontier.

The researchers simulated reasoning processes by constructing random walks across research literature, starting with a property like COVID vaccination and then jumping to related papers, authors, or cited materials. After running millions of these random walks, their model offered a 400% improvement in predictions of future discoveries compared to focusing on research content alone. Moreover, the model could predict with over 40% precision the actual individuals who would make specific discoveries, as it understood the connections between the property, material, and individual's experience or relationships.

Evans describes the model as a "digital double" of the scientific system, allowing simulations of likely scenarios and experimentation of alternative possibilities. This approach highlights scientists' tendency to stick to familiar methods, properties, and people. Evans explains that some aspects of the current scientific system, like graduate education, are not optimized for discovery but are geared towards job market requirements. To optimize the discovery of new, technologically relevant things, each student would need to be treated as an experiment, crossing novel gaps in the landscape of expertise.

In the second part of the study, the researchers asked the AI model to find predictions that are scientifically plausible but least likely to be discovered by people. These 'alien' or complementary inferences have three features: they are rarely discovered by humans; if discovered, it won’t be for many years until scientific systems reorganize; and they are, on average, better than human inferences, as humans tend to exhaust every possibility of an existing theory before exploring new ones. Because these models avoid the typical connections and configurations of human scientific activity, they explore entirely new territories.

Evans argues that viewing AI as an attempt to replicate human capacity does not help accelerate problem-solving. Instead, we are more likely to benefit from a radical augmentation of our collective intelligence. This involves changing the framing of AI from artificial intelligence to radically augmented intelligence, requiring a deeper understanding of individual and collective cognitive capacities. By understanding more about human cognition, we can design systems that compensate for its limitations, leading to collectively greater knowledge.

Key Highlights:

• UChicago researchers developed AI models that can predict human inferences, the scientists making them, and generate scientifically promising 'alien' hypotheses.
• The models offered a 400% improvement in predictions of future discoveries and could predict with over 40% precision the actual individuals who would make specific discoveries.
• The study suggests that we can benefit more from radically augmenting our collective intelligence rather than artificially replicating it. Reference: "UChicago study explores how A.I. can predict discoveries and who will make them," Nature Human Behaviour.

Resource: [1].