Data quality and augmentation

Lens adapted to this topic: Benefits, methods, and conditions for reliable use

The mainstream technical view is that improving data is often as important as improving model architecture. It supports a lifecycle combining quality management—such as profiling, cleansing, labeling, monitoring, and maintenance—with augmentation methods tailored to the task. Augmentation can reduce labeling burdens or improve coverage, but should be designed around label preservation, domain relevance, and empirical evaluation rather than treated as universally beneficial.

• Data-centric AI emphasizes training-data development, inference-data development, and data maintenance as central objectives alongside model design.
• Augmentation methods can create additional training data through rules, transformations, external knowledge, interpolation, conditional generation, and policy learning.
• Label-preserving transformations anchored to real examples are presented as a distinct and comparatively controlled form of augmentation.
• Automated data-quality systems can support profiling, entity matching, cleansing, monitoring, and anomaly investigation, helping scale processes across heterogeneous data.
• Synthetic data can be useful when it is filtered, verified, or curated rather than used as an unexamined replacement for real data.

Lens adapted to this topic: Risks, limits, and alternatives to synthetic expansion

A serious dissenting view argues that augmentation is not automatically a quality improvement. Synthetic examples inherit assumptions from the generator, may miss rare cases and the complexity of the real world, and can create misleading confidence if evaluated against the data that produced them. This view favors withheld-real-data testing, selective use of verified synthetic data, and more interaction or real-world experience instead of simply increasing dataset volume.

• A 2026 analysis argues that many synthetic-minority methods fail to demonstrate meaningful improvement when tested against withheld real data, and that validity depends on the underlying distribution.
• Richard Sutton’s critique holds that simulations are necessarily limited representations of the physical world and cannot straightforwardly substitute for experience or other people’s minds.
• Recursive training on unfiltered model output can erase distributional tails and reduce variance; filtering and curation are presented as important distinctions.
• Some analyses argue that the next valuable data may be interaction traces from agents acting in environments and receiving feedback, rather than more passive synthetic text.
• A broader critique questions whether ever-increasing data volume should be the default objective, pointing to rules, symbolic representations, constraints, and other approaches as alternatives.