Limitation or failure · Computer science · General science · Social science

Patterns paper frames data leakage as an ML-based science reproducibility crisis

A peer-reviewed survey and case study of data leakage failures in machine-learning-based scientific claims.

Summary

Kapoor and Narayanan survey reproducibility issues in machine-learning-based science and argue that data leakage can produce overoptimistic scientific claims. Their Patterns article reports leakage across 17 fields and 294 papers, proposes an eight-type taxonomy, and shows in a civil-war prediction case study that corrected analyses remove the claimed advantage of complex ML models.

AI role

Machine-learning models were used as evidence for scientific claims, but leakage in training, evaluation, or feature construction made some reported advantages overoptimistic.

Narrative role

This backfills a concrete methodological limitation from the last few years: AI and ML can accelerate publication and modeling while also amplifying hard-to-detect validity failures across scientific domains.

Caveat

The paper focuses on predictive ML used to support scientific claims, not all AI-for-science workflows, and the count reflects surveyed leakage evidence rather than a complete census of affected literature.

Event

Date: 2023-09-08
Timeline role: Notable
Progress relevance: Caution signal
Validation: Peer reviewed

Evidence

Peer-reviewed paper

Observed metrics

fields affected: 17 fields; The Patterns article reports leakage evidence across scientific fields that adopted ML methods.
papers affected: 294 papers; The peer-reviewed version reports prior reviews showing leakage across hundreds of papers.
leakage taxonomy: 8 types; The authors propose a taxonomy and model info sheets to detect leakage before publication.

Source: Patterns ↗