ISO/IEC 5259-2: The Data Quality Standard for Reliable AI

Enterprise trust takes years to build and only a moment to lose.

ISO/IEC 5259-2, the focus of this article, establishes that benchmark as an international standard. Pebblous assesses and improves data quality based on ISO/IEC 5259-2 for enterprise organizations across telecommunications, manufacturing, mobility, and other industries, including LG Uplus, one of Korea's three major telecommunications companies.

In this article, we will walk through what ISO/IEC 5259-2 is, how it applies to real-world data, and how to interpret measurement results.

What Makes ISO/IEC 5259-2 Different?

At Pebblous, we place particular emphasis on ISO/IEC 5259-2 among the many data quality standards. Why does this standard matter so much?

Put simply, it's the data quality standard best suited to where AI is headed.

To explain why, let's compare ISO/IEC 25012, ISO/IEC 5259, and ISO/IEC 5259-2.

ISO/IEC 25012

Introduced around 2015, this standard was designed to manage the quality of structured data stored in databases. To date, many government agencies and enterprises have used this standard as a basis for data quality management.

• Consider a dataset. The file names are valid, the image resolution is uniform across files, and there are no obvious missing records. The labeling is also well executed. Under ISO/IEC 25012, this dataset would likely be considered high quality.

• But what if 90% of the samples come from only one specific condition, or if all images were captured from the same angle and environment? The data format may be flawless, but AI performance will suffer.

ISO/IEC 5259

This is why ISO/IEC 5259 was developed. ISO/IEC 5259 builds on ISO/IEC 25012 and extends it for AI. It also defines additional data quality characteristics essential for AI and machine learning, such as diversity, representativeness, and similarity.

Pebblous leverages cutting-edge research on the ISO/IEC 5259 series to drive real-world impact. To dig deeper into ISO/IEC 5259 measurement criteria, check out our detailed guide to ISO/IEC 5259-2 and our ISO/IEC 5259 series information hub.

ISO/IEC 5259-2

As described above, ISO/IEC 5259 is an international standards series for managing the quality of AI training data. Within this series, ISO/IEC 5259-2 focuses on how data quality should be measured and evaluated in practice.

ISO/IEC 5259-2 Measurement Criteria

From this point on, we'll focus specifically on ISO/IEC 5259-2. Its measurement items are broadly divided into four groups.

1. Intrinsic Data Quality Characteristics

These characteristics assess the fundamental properties of the data itself. They evaluate whether the content of the data is valid based on five criteria: accuracy, completeness, consistency, credibility, and currentness.

2. Intrinsic and System-Dependent Data Quality Characteristics 

Data quality can change depending on the system that stores and processes the data. In other words, this category does not evaluate the properties of data in isolation. It evaluates the quality created by the combination of the data and the system that stores and processes it. It examines how well data is managed within a system, including accessibility, compliance, and traceability.

3. System-Dependent Data Quality Characteristics 

These items evaluate the performance of the IT systems and infrastructure that hold the data. Even high-quality data is difficult to use in practice if the system is unstable. This category assesses whether data is available when needed, portable to other environments, and recoverable in the event of failure.

4. Additional Data Quality Characteristics for Analytics and ML 

These ML-specific data quality characteristics consist of 26 items across 10 subcategories, including auditability, balance, diversity, validity, identifiability, suitability, representativeness, similarity, and timeliness.

For detailed measurement criteria for each item, including QM codes, see the Pebblous blog.

How Does Pebblous Apply ISO/IEC 5259-2 to Real Assessments?

Pebblous applies ISO/IEC 5259-2 to Data Clinic, our data quality management solution, to assess real datasets. We organize the quality characteristics of ISO/IEC 5259-2 by assessment level and analyze data quality accordingly.

Level I : Foundational Quality Assessment 

This level checks the basic quality condition of a dataset. Think of it as a baseline health check for your dataset.

Level II·III : AI-Centric Quality Assessment 

However, AI data quality cannot be determined by missing values or formatting errors alone. To build data that AI can learn from effectively, Data Clinic uses DataLens. It projects data into an embedding space and analyzes quality in the way AI perceives data.

Think of an 'embedding space' as a massive digital map where similar data points are clustered together like neighboring cities. By analyzing this map, AI doesn't just read data; it visualizes the entire landscape, allowing us to spot 'deserts' where data is missing, or 'congested zones' where redundant data is piled up. 

Why ISO/IEC 5259-2 Is Difficult to Apply in Practice, Even When You Understand It

Even when organizations understand ISO/IEC 5259-2, improving data quality to the desired level is often difficult in practice. Pebblous has already addressed many of these challenges. Based on that experience, we will explain why many teams struggle and how those challenges can be solved.

1. Data Quality Cannot Be Solved by Standards Alone. 

Standards are important, but standards are not everything. (quote)

• In real-world operations, a single standard can't answer every question that comes up. Conditions vary widely by industry, and so does the complexity of each problem. Standards alone can't keep pace with these realities.

• Organizations must also comply with AI regulations and governance requirements. Yet many ethical criteria remain insufficiently defined.

In this context, Pebblous’ data quality management solution, Data Clinic, does not rely solely on ISO/IEC 5259-2.

The agent's training scope consists of the following five areas. These bodies of knowledge work together to improve data quality.

• Data quality standards, including ISO 5259 (28.84%)

• Regulations and governance frameworks, such as the EU AI Act, GDPR, NIST AI RMF (U.S. AI Risk Management Framework), and the Executive Order on AI (45.09%)

• Domain applications in robotics, manufacturing, and public safety (14.06%)

• Foundations of data science based on recent academic conference papers, including ICML (9.89%)

• Pebblous internal technical documents (2.11%)

2. Data Is Viewed Only from a Human Perspective, Not an AI Perspective. 

Artificial intelligence may appear highly capable, but its operating principles are fundamentally different from those of the human brain.

Humans understand data intuitively, but AI does not understand data in the same way. AI learns patterns in data. As a result, data that appears normal to humans can become biased training data for AI. (call out)

That is why organizations need to shift perspective. Data should be examined not from the human point of view, but from the AI point of view.

Two Solutions Developed by Pebblous 

After extensive research, Pebblous identified two solutions.

1. Neural Network-Based DataLens 

Rather than treating AI data as a simple table, this technology analyzes how AI actually perceives the data, how data points relate to one another, and what the statistical distribution looks like.

Data Clinic uses two types of DataLens: general-purpose and customized. Among them, customized DataLens is developed to reflect the characteristics of each domain.

2. Data Imaging

Data Imaging is a technology that makes the data seen by AI visible to humans. Data is high-dimensional, often containing thousands of features. Data Imaging visualizes it in 2D or 3D space, allowing teams to see at a glance where data is concentrated, which areas are empty, and where outliers exist. Pebblous holds a U.S. patent for this approach, which combines ISO 5259 quality criteria with Data Imaging to assess data quality, and the technology is proprietary to Pebblous.

A Real Data Quality Assessment Process Using Data Clinic 

Let's look at how Data Clinic, which combines all the technologies described above, assesses data quality. We'll use a diagnostic report on a recyclable waste dataset as an example.

Below are two issues Data Clinic identified by projecting images into an embedding space and analyzing them.

Class Imbalance 

According to ISO/IEC 5259-2, this dataset showed a clear class imbalance issue. When comparing the video_tape class with the filament class, the counts were 646 images and 33 images, respectively. In class analysis, a ratio exceeding 10:1 between the largest and smallest classes is generally considered a severe imbalance. In this case, the ratio was 19.6:1.

We also analyzed the degree of imbalance using L2 (general DataLens) and L3 (domain-specific DataLens). L2 density values ranged from 0.06 to 0.18, while L3 values ranged from 0.5 to 1.5. The units differ because L2 is based on general-purpose embeddings with 1,280 dimensions, whereas L3 uses optimized embeddings with 32 dimensions through fine-tuning for the recycling domain. As a result, L3 captures recycling-related features more sensitively.

Looking first at the L2 chart, we can see that video_tape samples are clustered with similar samples in the embedding space.

Through the L3 domain-specific lens, this pattern becomes even clearer. When analyzed from the recycling-domain perspective, video_tape shows a much larger density gap compared with other classes. This clearly reveals that video_tape accounts for 37.8% of the dataset and that similar patterns repeat within that class.

You can find the full explanation of the diagnostic report at the link below.

Where Pebblous Is Headed Next 

Pebblous is not satisfied with today's data quality standards. We anticipate the quality standards of the future.

When Pebblous was founded in 2021 and first began developing Data Clinic, its data quality management solution, ISO 5259 did not yet exist. Because there was no adequate quality standard tailored to AI training data, Pebblous developed its own criteria.

1. Under review for designation as a KOLAS-accredited body

KOLAS is a Korean government accreditation scheme that evaluates and officially accredits the technical competence of testing, inspection, and calibration bodies under Korea's Framework Act on National Standards. It can be understood as Korea's counterpart to ANAB in the United States.

Pebblous is currently undergoing review to be recognized by KOLAS as a trusted organization authorized to issue certificates for the ISO/IEC 5259-2 standard. We have successfully completed the on-site assessment and expect to obtain the accreditation body designation in August. 

2. Implementing automated assessment for all items in the AADS project

In the AADS (Agentic AI Data Scientist) project currently being carried out by the Ministry of Science and ICT, we are developing technology that automatically assesses data quality based on the 5259 standard. We successfully automated part of the process in Phase 1, and in Phase 2, all 5259 items will be assessed automatically. Neuro-symbolic technology and AI Agent technology are at the core of this work.

Are you curious whether your AI data currently meets the criteria above?

Data Clinic assesses your dataset against ISO/IEC 5259-2 and delivers a report that pinpoints exactly where the issues are. Using Pebblous' proprietary technology, we help you build a data quality foundation that stays resilient as AI regulations and standards continue to evolve.