AI Research Is More Documented—but Is It More Reproducible?

A new study in Communications AI & Computing, published October 22, analyzes documentation practices in 56,800 papers from five major AI conferences (2014–2024). The authors estimate that reproducibility rose from 28% to 64% over that period—but the estimate is inferred from documentation and prior findings, not from independently rerunning the studies. (nature.com)

That distinction matters for decisions: better records make a result easier to inspect and reuse, but they do not show on their own that another team can reproduce it. The study’s evidence is useful for tracking research practice; it is not a direct pass rate for replicating results.

Read the study. When should a well-documented result be trusted for a new decision—and what additional check would you want first?