data-centric AI
/ DAY-tuh-SEN-trik AY-EYE /
For years, the way to improve an AI system was to tinker with the model — try a fancier architecture, tune more knobs, add more layers — while treating the dataset as a fixed thing you were stuck with. Data-centric AI flips that priority. It holds the model steady and systematically improves the data instead: fixing wrong labels, filling gaps, removing duplicates, adding examples of cases the model fumbles. The slogan is that better data often beats a better model.
Concretely, a data-centric workflow looks for where a model fails, traces those failures back to the data, and fixes the data: correcting mislabeled examples, making labeling guidelines consistent, collecting more of the under-represented situations, and measuring data quality as carefully as model accuracy. The phrase was popularized by Andrew Ng around 2021 as a deliberate counterweight to the field's long obsession with bigger and cleverer models.
Why it matters: in many real projects the dataset is messier and more impactful than the model, yet it gets far less attention. Cleaning labels and patching coverage gaps frequently yields larger, more reliable gains than swapping algorithms — especially when data is limited. The honest caveat is that this is a shift in emphasis, not a silver bullet: both data and models matter, and the right answer is usually to improve whichever is currently the weakest link rather than to declare one of them king.
A defect-detection model is stuck at 76%. Instead of trying new architectures, the team audits the data and finds inspectors disagreed on what counts as a "scratch." They rewrite the labeling guideline, relabel consistently, and accuracy jumps to 90% — with the exact same model. The fix was in the data all along.
Same model, fixed labels: a 14-point gain that no architecture could buy.
Data-centric AI is a rebalancing of effort, not a claim that models don't matter. In practice you improve whichever is the weakest link right now — and surprisingly often, that link is the data.