Human-AI Teams Make Better Decisions Than Superhuman Machines Alone

MPINB researcher Felipe Yáñez and team publish paper in Patterns

How can humans and artificial intelligence make better decisions together? A new study led by Felipe Yáñez, researcher in the In Silico Brain Sciences group, shows that combining human and machine judgments can outperform either partner alone, even in tasks where AI is already superior.

The work, now published in Patterns (Cell Press), introduces a simple and efficient method for integrating decisions from both humans and machines, such as large language models (LLMs). Instead of relying on complex or computationally expensive models, the team developed a confidence-weighted logistic regression approach that captures and integrates uncertainty from both human and machine judgments. The model is adaptable and scalable, easily interpretable, and capable of incorporating any number of teammates.

Why collaboration works

The study builds on previous research showing that LLMs can surpass human experts in predicting results of neuroscience experiments. This raised an important question: Can humans still contribute meaningfully when machines outperform them?

The answer, the new paper shows, is yes, when two conditions are met:

1. Confidence calibration: When both humans and AI reliably link confidence to accuracy, their judgments can be meaningfully combined.
2. Diversity of errors: Humans and machines tend to make different mistakes. This diversity becomes an advantage when their judgments are integrated.

Tested in two domains

The team tested their method in:

• Object recognition: When human decisions were added to purely machine-based teams, overall accuracy increased, even when machines performed better on average.
• Neuroscience forecasting (BrainBench): In knowledge-intensive tasks where LLMs outperformed human experts, adding a human teammate still improved performance, but only when confidence ratings were included.

These results demonstrate that humans can strengthen AI systems even in domains dominated by machine performance, highlighting the value of human intuition and complementary judgment.

A general approach for human-machine teaming

Because the method works with any number of teammates and only requires that each can express confidence in their decisions, it can be applied far beyond the two tasks studied here. The authors hope that this approach will support more effective and transparent collaboration between humans and AI systems, foreshadowing a future in which machines actively contribute to human discovery and innovation.

Publication

Yáñez, F., Luo, X., Valerio Minero, O., & Love, B. C.

Confidence-weighted integration of human and machine judgments for superior decision-making.
Patterns, Cell Press (2025).

https://doi.org/10.1016/j.patter.2025.101423

About the researcher

Felipe Yáñez is part of the In Silico Brain Sciences group. His research focuses on computational neuroscience, prediction, and human-AI collaboration.