ETSI's 18 Data Quality Metrics: A Critical Safeguard for AI Applications

ETSI Releases Critical Data Quality Framework for AI Systems

The European Telecommunications Standards Institute (ETSI) has published TR 104 180, a technical report that introduces 18 measurable metrics for assessing data quality before deployment in AI applications. This framework gives companies a concrete methodology to evaluate whether their datasets are suitable for training and running language models and other AI systems—a critical capability as organizations race to implement AI without compromising reliability or security.

Why This Matters for AI Builders and Users

The stakes for data quality in AI are higher than ever. Poor data doesn't just degrade model performance—it introduces systematic biases, enables adversarial attacks, and undermines the guardrails that keep AI systems safe and trustworthy. For companies building LLM applications, unvetted data can turn a promising AI initiative into a liability.

The ETSI report addresses this head-on by providing standardized, formula-based metrics that organizations can implement immediately. Rather than relying on intuition or ad-hoc quality checks, teams now have a structured way to measure data fitness before it reaches their models.

The Four Categories of Data Quality Metrics

The 18 metrics are organized into four groups, each targeting a different dimension of data reliability:

• Foundational Quality: Measures whether data is complete, accurate, consistent, and free of duplicates—the basic hygiene that prevents corrupted training sets.
• Contextual Suitability: Evaluates whether data is relevant and appropriate for the specific AI task at hand.
• Governance and Provenance: Assesses data lineage, documentation, and compliance with regulations.
• Security and Trustworthiness: Examines whether data has been protected from tampering and meets ethical standards.

The LLM-Specific Risk Picture

Language models are particularly vulnerable to poor data quality. When LLMs are trained on or prompted with low-quality data, they can:

• Hallucinate or produce factually incorrect outputs that undermine user trust
• Amplify biases, leading to discriminatory outputs that expose companies to legal and reputational risk
• Become susceptible to prompt injection and jailbreak attacks that exploit data inconsistencies
• Bypass safety guardrails when poisoned data creates conflicting training signals

The ETSI framework directly addresses these risks by making data quality measurable and verifiable before it enters your AI pipeline.

What Builders Should Do Now

Organizations developing or deploying LLM applications should take three immediate steps:

• Audit existing datasets: Apply ETSI's 18 metrics to evaluate the data currently powering your AI systems. You may discover quality gaps you didn't know existed.
• Establish data quality gates: Implement these metrics as checkpoints in your data pipeline. Don't let unvetted data reach your models.
• Document and track metrics: Create dashboards that continuously monitor data quality throughout your system's lifecycle. Data quality degrades over time as new data enters the pipeline.

The Bigger Picture

ETSI's TR 104 180 represents a maturation of AI governance. Rather than treating data as a black box or an afterthought, the framework elevates data quality to a first-class concern—one with measurable standards and clear responsibility assignments.

For teams building production AI systems, this is essential reading. The report provides not just guidance, but implementable formulas you can use immediately to measure whether your data can actually be trusted.

The Bottom Line

Garbage in, garbage out—this adage has never been more relevant than in AI. The ETSI's 18-metric framework gives organizations the tools to ensure their AI systems are built on solid data foundations. For LLM builders, treating data quality as a security and reliability issue isn't optional—it's essential to maintaining trust, safety, and performance in production systems.