Search Website
Optimizing Human-AI Collaborations
Project Title: Beyond the Accuracy-Interpretability Tradeoff: Optimizing Human-AI Collaborations
Project Funding: SSHRC Insight Development Grant, 2025-2027
This project emerged out of the Computational Epistemology Think Tank project.
This project examines whether interpretable machine learning models lead to better outcomes when evaluated in the context of human-model collaboration, rather than in isolation. The study challenges the common assumption that there's an inherent trade-off between model performance and interpretability, suggesting that while opaque models may show superior predictive accuracy alone, interpretable models might enable better human-machine team performance.
This project will aim to address several key questions:
- How should evaluation criteria for human-model pairs differ from standalone model metrics?
- What insights can be drawn from model usage in scientific research?
- How can dynamic methods like simulations enhance evaluation of human-model collaborations?
- Does empirical evidence support the claim that model explanations improve human-model team performance?
- Under what conditions does this claim hold true?
The project emphasizes two broader contexts often overlooked in current research. First, the use of models in research settings, where understanding internal mechanisms can drive scientific insights, contrasting with the typical focus on decision-making domains like medicine and law, and second, the dynamic nature of sociotechnical systems, where benefits of explanations may emerge over time through increased understanding and model improvements, rather than just immediate performance gains.
We take an interdisciplinary approach that combines empirical evidence, case studies, normative inquiry, and simulations to provide comprehensive insights into human-model collaborations. In terms of longer-term outputs, our aim is to provide practical guidance for designing interpretable AI systems that effectively complement human decision-makers and assist with scientific research, while ensuring that interpretability research focuses on features of AI models that serve genuine epistemic goals.
Primary Investigator:
- Alice Huang, Assistant Professor, Department of Philosophy, Faculty of Arts & Humanities; Department of Computer Science, Faculty of Science
Co-Applicant:
- Chris Smeenk, Professor, Department of Philosophy, Faculty of Arts & Humanities