Accelerating science with human-aware artificial intelligence.

Area of Science:

• Computer Science
• Scientific Discovery
• Artificial Intelligence

Background:

• Artificial intelligence (AI) models traditionally focus on research content, neglecting the crucial role of human scientists in discovery.
• Existing AI models have limitations in predicting future scientific breakthroughs, particularly when literature is sparse.

Purpose of the Study:

• To develop and evaluate AI models that incorporate human expertise for enhanced prediction of scientific discoveries.
• To investigate the impact of human-aware AI on accelerating scientific advancement and generating novel hypotheses.

Main Methods:

• Trained unsupervised AI models on simulated inferences reflecting human expert cognition.
• Incorporated the distribution of human expertise into AI model training alongside research content.
• Tuned AI models to identify scientifically promising hypotheses beyond conventional research directions.

Main Results:

• AI prediction of future discoveries improved by up to 400% compared to models using research content alone.
• Models successfully predicted not only discoveries but also the human scientists likely to make them.
• Human-aware AI generated 'alien' hypotheses unlikely to be discovered through conventional means.

Conclusions:

• Integrating human expertise into AI models dramatically enhances the prediction of scientific discoveries.
• Human-aware AI can accelerate scientific progress by probing current research blind spots and generating novel hypotheses.
• This approach moves AI beyond merely analyzing data to actively participating in the scientific discovery process.