In this paper, my coauthors and I show how machine learning can help researchers discover meaningful patterns in complex data without taking over the human work of explanation. The core idea is simple: instead of using AI only to test existing theories, we use it earlier in the process to detect robust, repeatable patterns that people might otherwise miss, especially in large datasets. We then separate that step from the human task of making sense of those patterns and turning them into theory. Algorithms can scan for patterns with speed, consistency, and protection against overfitting, while researchers still provide interpretation and causal explanation. In other words, I find that AI is most valuable not when it replaces theorizing, but when it helps humans see more clearly what is happening in the data and build stronger explanations from it.

In this research, I extend my earlier work with Shrestha et al by moving from pattern detection to pattern explanation. I focus on a form of reasoning called abduction: the human ability to look at an unexpected pattern. This kind of reasoning is essential in a world where many interacting causes shape innovation, but it is also vulnerable to bias and oversimplified “just-so” stories. The key idea here is that humans and AI can play complementary roles: machine learning can help detect robust patterns in large datasets, while generative AI can help people generate and compare multiple possible explanations. But AI should not replace human judgment; we still need to decide which questions matter, evaluate which explanations are credible, and test them carefully. In that sense, this research shows how intelligent machines can strengthen human cognition by expanding our ability to reason, while keeping humans responsible for critical thinking.

This research shows that predictive AI not only help people perform a task in the moment but also help them learn to do the task better on their own. Studying medical novices diagnosing lung cancer from CT scans, we find that AI improves performance not only when it is available during practice, but also after AI is taken away, which suggests a learning effect. Most importantly, the best results come when AI is woven into both training and practice: in that case, novices’ accuracy rises to the level of expert doctors. We also show that how AI is deployed matters. Some forms of AI deployment improve individual accuracy but make people’s mistakes more similar, which weakens the value of second opinions in team decision-making. Other forms improve both individual learning and diversity of judgments. The key message: with the right deployment strategy, AI can act as a coach that builds human expertise and strengthens collective decision-making, rather than as a shortcut that degrades human capability.