Is metadata always stored in a separate file?

No. It can be embedded or external, as long as the relationship to the resource is maintained.

Can metadata be wrong?

Yes. It needs validation and maintenance just like other information.

Are there different kinds?

Yes. Common groupings include descriptive, structural, and administrative metadata, with technical, rights, and preservation concerns often discussed within broader schemes.

AI Glossary / Data governance and quality / Intermediate

Metadata

Definition

Metadata is information that describes, explains, organizes, or helps manage other information resources. It can identify a dataset's creator and purpose, define its fields and units, record a file's format, or describe relationships between versions. Metadata makes information easier to discover, interpret, govern, and preserve. The familiar phrase data about data is a useful starting point, but concrete examples are needed to show what a particular resource requires.

What it means

A table of numbers is difficult to use responsibly if no one knows what the columns mean or how the values were collected. Metadata supplies that context. It can be embedded in the resource or maintained separately, provided the relationship remains reliable. The distinction between data and metadata is contextual: a timestamp may describe a file in one system while becoming a measured variable in another analysis. Metadata itself also needs quality control.

How it works

Identify the information users need to find, understand, manage, and reuse a resource. Select suitable metadata fields and conventions, then capture them as close as practical to creation and subsequent changes. Link metadata to the correct resource and version. Validate required fields, use consistent meanings, and update the description when the underlying resource changes. Combine automated technical extraction with human knowledge about purpose, methods, limitations, and rights. Make metadata accessible to the intended users while protecting sensitive details it may contain.

Three practical examples

Dataset interpretation: field descriptions specify that a temperature is in degrees Celsius and explain how missing observations are represented.
Digital collections: descriptive records help people locate a document, while structural information records the order and relationships of its parts.
AI governance: a model package records the version, intended use, training-data references, evaluation context, and known limitations needed for an informed review.

Think of it this way

Think of a label and instruction sheet attached to a box of scientific samples. The contents matter, but the labels explain what they are, where they came from, and how to handle them. The analogy is limited because metadata can be machine-readable, distributed across systems, and sensitive in its own right, and a wrong label can be as misleading as missing one.

Worked example

In an original example, a dataset contains a column named distance with values 2, 5, and 8. Without metadata, a user cannot tell whether these represent meters, kilometers, or another quantity. The metadata specifies kilometers, measurement from a defined origin, and a missing-value code of -1. A pipeline that treats -1 as a real distance would calculate an incorrect average if such a value appears. The team therefore preserves the unit and missing-value definition alongside the dataset and tests the transformation against them. A descriptive field name alone would not carry all this meaning, and converting units requires updating the relevant description.

Illustrative case study

Fictional example: Oakline AI publishes two model files with similar names. Reviewers cannot tell which preprocessing configuration belongs to each. The team creates version-specific metadata linking the artifact, configuration, dataset references, and evaluation report. It checks those links during release. This invented case illustrates that useful metadata is not merely a long description; it must accurately identify the relationships needed to interpret and operate the resource.

Common misconceptions and mistakes

Do not assume automatically extracted file properties explain business meaning or data limitations. Do not treat metadata as inherently non-sensitive: creator identities, locations, and access details can reveal information. Avoid copying a prior dataset's description without checking what changed. Also, a metadata standard provides shared structure and meaning but cannot guarantee that the supplied statements are accurate. Stale or inconsistent metadata can produce confident errors, especially when automated systems use it to select or interpret data.

Business use

Good metadata reduces avoidable clarification work and supports reliable discovery, integration, and governance. For AI projects, it helps teams connect data preparation, model versions, evaluation results, and deployment decisions. Decide which metadata is essential for the intended task and assign responsibility for maintaining it. Check whether users can actually answer important questions from the description. A compact, accurate record of units, provenance, limitations, and ownership can be more valuable than many automatically populated fields with unclear meaning.

Check your understanding

A dataset is converted from kilometers to meters, but its unit description remains kilometers. Is preserving the metadata unchanged the right result?

Answer: No. The metadata must reflect the transformed resource, while provenance should retain the relationship to the original representation and conversion.

Related terms

Schema
Data catalog
Digital continuity
data dictionary
provenance
lineage
versioning

Learn next

take a small non-sensitive dataset and ask someone unfamiliar with it to interpret three fields. Record the questions they cannot answer. Add only the metadata needed to resolve those uncertainties, then check how the description will stay linked to the correct version when the data changes.

Sources

Sources: NISO, Understanding Metadata primer, https://www.niso.org/publications/understanding-metadata-2017 , publication overview consulted; NISO, Metadata Demystified, https://www.niso.org/sites/default/files/2017-08/Metadata_Demystified.pdf , relevant definition and function excerpts consulted; NISO, A Framework of Guidance for Building Good Digital Collections, https://www.niso.org/sites/default/files/2017-08/framework3.pdf , descriptive, administrative, and structural metadata excerpts consulted. Taxonomies are presented as common groupings, not one mandatory universal scheme. Dataset and AI examples are original; the full standards corpus was not reviewed.

Previous: Metaclass (knowledge representation) Next: Metadata content standard