Reduced, Reused and Recycled: The Life of a Dataset in Machine Learning Research

Benchmark datasets show increasing concentration: The research finds that machine learning communities are relying on fewer and fewer datasets over time (2015-2020), meaning researchers are repeatedly using the same benchmarks rather than developing diverse evaluation approaches. This concentration could limit the breadth of problems being tested and create blind spots in AI system evaluation.

Elite institutions control dataset creation: Datasets are increasingly concentrated among researchers at a small number of elite institutions. This institutional concentration raises equity concerns, as it means researchers at less-resourced institutions may have less influence over which problems are prioritized and how progress is measured in the field.

Datasets migrate across research areas: The paper documents significant adoption of datasets from one task area into others—meaning datasets designed for one purpose are being repurposed for different applications. While reuse can be efficient, this practice risks applying datasets in contexts they weren't designed for, potentially introducing hidden biases or limitations.

Implications for scientific evaluation and AI ethics: The authors explicitly note their findings have consequences for how we evaluate AI systems and approach responsible AI. Heavy reliance on a narrow set of benchmarks may mask real-world performance problems and create a false sense of progress that doesn't translate to robust, ethical AI systems in practice.

Access and participation barriers: The concentration patterns suggest structural inequities in who gets to shape AI research priorities, which could disadvantage researchers and communities outside elite institutions and limit whose values and concerns are represented in benchmark design.