Zoom Talk: Kathleen Creel (Northeastern)

joint talk with Liam Kofi Bright (LSE)

Title:

Don't Use Machine Learning To Evaluate Grants

Abstract:

Funding science is a chancy business. Promising projects come to naught; strong results fail to replicate. To reduce the uncertainty of their bets, grant-making agencies are encouraging the development of tools that aim to predict which papers will replicate -- and, eventually, it is hoped, which grants to fund. The putative benefits of a tool that could predict the success of funding proposals are clear: time saved, public funds better allocated. And the challenges to constructing such a tool seem equally clear: agreeing on machine-evaluable metrics of success for proposals, predicting future scientific success based on the past successes the future is mean to exceed. The current state of automated prediction reflects these challenges. But whether achievement of this task is likely in the short term -- or even possible -- is immaterial for our argument, which concerns the impact on the epistemic diversity of scientific communities impacted by the establishment of an automated bottleneck.

We first outline the reasons to think machine learning-based predictive instruments are likely to homogenize which science is funded. We then rehearse the arguments for epistemic diversity in science, establishing the stakes. We next generate four possible scenarios for the deployment of algorithmic decision-making tools in grant evaluation by varying the predictive success of the average tool and the degree of correlation between the tools. We show that no matter how predictively successful the tool is, using machine learning to filter grants is likely to reduce the diversity of scientific approaches. We conclude by recommending solutions to grant-making agencies.