Machine Learning Reconfigures Science From Inference to Prediction

According to the arXiv preprint by Malena Mendez Isla, Vincent Lariviere, and Diego Kozlowski, the authors map machine-learning research using 4.9 million publications and 255 ML techniques. The paper reports a core-periphery structure in which physical sciences form the methodological core and health sciences represent a major adoption area. It also reports that predictive techniques are concentrated in computer science, while inferential approaches remain distributed across applied fields.

The preprint's central practitioner issue is the trade-off between predictive performance and traceability. A shift from inference-oriented methods to predictive architectures can improve benchmark performance while making causal interpretation, uncertainty explanation, and replication harder. The post-2022 dependence on external commercial models adds another layer because training data, model versions, and internal processes may be inaccessible to downstream researchers.

For AI/DS teams, this argues for stronger experiment records: model versioning, dataset provenance, disclosed prompts or APIs where possible, and evaluation methods that separate predictive fit from explanatory claims. In regulated or clinical settings, the same shift can create review friction when a strong predictor cannot support an interpretable claim.

Track whether journals and conferences start requiring clearer disclosure for external model use, model-version pinning, or interpretability checks when predictive systems replace inferential methods. Those rules will shape whether the reported shift becomes a reproducibility problem or a managed methodological transition.