What is data-centric AI?
Data-centric AI is the practice of improving a model by improving its data. Where a model-centric mindset holds the data fixed and iterates on architectures and hyperparameters, a data-centric mindset holds the model fixed and iterates on the data: fixing labels, filling coverage gaps, removing noise and duplicates, and adding examples where the model is weak.
The premise is that on many real problems, better data yields larger and more reliable gains than a better model.
Key takeaways
- Data-centric AI treats data quality as the main lever for performance.
- It systematically fixes labels, coverage, and noise instead of only tuning models.
- On real problems, better data often beats a better model.
How it works
The loop is diagnose, then improve the data. You evaluate the model, slice results to find weak segments, run error analysis to find failure patterns and label mistakes, then act by relabeling, curating, augmenting, or collecting targeted data. Retraining on the improved data closes the loop, and the cycle repeats.
Why it matters
Modern architectures are widely available and often good enough, which shifts the competitive edge to data. A disciplined data-centric process turns curation, annotation quality, and evaluation into a repeatable engine for improvement, which is especially valuable when labeling is the most expensive part of the pipeline.
Frequently asked questions
How is data-centric AI different from model-centric AI?
Model-centric work improves the architecture and training while holding data fixed. Data-centric work improves the data while holding the model fixed.
Does data-centric AI mean ignoring the model?
No. It means prioritizing data quality as the primary lever, while still using a solid model. The two are complementary.
Related terms