Data Quality in AI: Why Better Data Beats Better Models

Science Glossary · Explore this series ›

Tormod Guldvog

Editor & Duowriter

March 21, 2026

Key Takeaways

• Model performance is bounded by training data quality.
• MIT found 3.4% label errors in major ML benchmark datasets.
• ISO/IEC 5259 (2024) standardizes AI data quality measurement.

✓ Fact-checked · See verification details

AI assistant: Research, Drafting, QA

Data quality in AI is the measure of how accurate, complete, consistent, and representative a dataset is for training, testing, and validating machine learning models.

Why it matters

The performance of any machine learning model is bounded by the quality of the data it learns from. Andrew Ng, co-founder of Google Brain and Stanford's AI Lab, formalized this principle in 2021 when he launched the data-centric AI movement at NeurIPS. His argument was direct: for most practical applications, improving the data produces better results than improving the model.

The consequences of poor data quality are measurable. MIT researchers found an average of 3.4% label errors across ten commonly used machine learning benchmark datasets, including over 2,900 mislabeled images in ImageNet's validation set alone. Ng identified inconsistent labeling in 20 to 30 percent of computer vision datasets. These errors do not simply reduce accuracy. They introduce systematic biases that compound through training.

In healthcare AI, the effects can be stark. Diagnostic models trained on imbalanced datasets have missed 44% of liver disease cases in women, compared to 23% in men. Facial recognition systems have shown error rates up to 35% for dark-skinned women while dropping below 1% for white men, as documented in MIT Media Lab's Gender Shades project (Buolamwini and Gebru, 2018). In each case, the algorithm performed exactly as its training data instructed.

How it works

Data quality in AI is measured across several dimensions, now codified in the ISO/IEC 5259 standard series (published 2024). The standard defines five core characteristics: accuracy (whether values correctly represent reality), completeness (whether all necessary elements are present), consistency (whether data remains uniform across sources), timeliness (whether data reflects current conditions), and representativeness (whether the dataset reflects the real-world population the model will serve).

Traditional software engineering adopted the phrase "garbage in, garbage out" decades ago. In machine learning, the principle operates with greater force. A 15% inaccuracy rate in training data can undermine model performance entirely. Unlike traditional software bugs, data quality problems often remain invisible until a model is deployed and begins producing biased or unreliable outputs.

Addressing data quality involves several stages: profiling (understanding what the data contains), cleaning (correcting errors and removing duplicates), augmentation (expanding datasets to improve representativeness), and monitoring (tracking quality over time as distributions shift). An estimated 91% of deployed ML models experience temporal degradation, where accuracy drops within months as real-world conditions diverge from training data.

Key context

The concept of data quality predates AI by decades. The phrase "garbage in, garbage out" appeared in computing literature as early as the late 1950s. But the field gained specific urgency for machine learning in the 2010s, as deep learning models began consuming datasets of unprecedented scale. When OpenAI trained GPT-3 on 570 GB of internet text in 2020, questions about what that text contained, and what it excluded, became unavoidable.

In 2021, Andrew Ng's data-centric AI workshops at NeurIPS marked a formal turning point. The workshops argued that the ML community's emphasis on model architecture had overshadowed the more fundamental question of data quality. As of 2024, ISO/IEC 5259 provides the first international standard framework for measuring and managing data quality specifically for AI and machine learning applications.

FAQ

Sources

• Primary: Data-Centric AI Competition, NeurIPS 2021 (Ng, A. et al., 2021)
• Primary: Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks (Northcutt, C.G., Athalye, A., Mueller, J., 2021)
• Standard: ISO/IEC 5259-1:2024, Artificial Intelligence: Data Quality for Analytics and Machine Learning
• Primary: Gender Shades: Intersectional Accuracy Disparities (Buolamwini, J. and Gebru, T., 2018)
• Survey: Data-Centric Artificial Intelligence: A Survey (Zha, D. et al., 2023)

Related Reading

Will AI Replace Artists? Perhaps - Because Clients Might Demand It

AI Model Reads DNA's Hidden Switches, One Letter at a Time

AI Materials Discovery: 5 Things to Know

AI Slop Is Flooding Scientific Publishing - and It Shows

Tormod Guldvog

Editor & Duowriter

Tormod Guldvog is the editor of Science Reader, covering the space where human and artificial intelligence meet. With a background in technology and science communication, he writes about how AI is changing scientific discovery — and what that means for us. More about Science Reader

Cite this article

Tormod Guldvog. (2026, March 21). Data Quality in AI: Why Better Data Beats Better Models. Science Reader. https://sciencereader.com/glossary/data-quality-in-ai/