The FAIR Data Principles are guidelines for making data and metadata findable, accessible, interoperable, and reusable. Published in 2016, they provide a framework for improving the discovery, exchange, interpretation, and reuse of digital research resources, especially through automated computational processes. FAIR is not a particular file format, software platform, or certification scheme. It describes properties that data, their descriptions, and the services supporting them should have.
The principles are widely used in research data management and increasingly inform the design of repositories, data infrastructures, and organizational data policies. Their central concern is that digital resources should remain understandable and usable beyond the immediate context in which they were created.
Origins and scope
The principles were formally presented by Mark D. Wilkinson and co-authors in the 2016 article “The FAIR Guiding Principles for scientific data management and stewardship,” published in Scientific Data. They responded to a recurring problem: data may be technically available but difficult to discover, interpret, combine, or reuse without substantial help from their creators.
Earlier practices such as repository deposition, descriptive documentation, and standardized formats addressed parts of this problem. FAIR brought these concerns into a common framework with an explicit emphasis on machine actionability: the ability of computational systems to locate resources, retrieve relevant descriptions, interpret relationships, and carry out useful operations with limited human intervention.
Although the principles were formulated around scientific data, their underlying approach also applies to metadata and other digital research objects, including software, workflows, and analytical models. Applying FAIR to these objects requires practices appropriate to their type; a software package, for example, has different reuse requirements from a collection of environmental measurements.
The four groups of principles
The framework contains 15 principles organized under four headings. The notation “(meta)data” in the original formulation indicates that a requirement concerns data, metadata, or both, as appropriate.
Findable
Findability concerns whether a resource can be identified and discovered.
| Principle | Requirement |
|---|---|
| F1 | Data and metadata are assigned globally unique and persistent identifiers. |
| F2 | Data are described with rich metadata. |
| F3 | Metadata explicitly include the identifier of the data they describe. |
| F4 | Data and metadata are registered or indexed in a searchable resource. |
A persistent identifier provides a durable reference to a resource even when its storage location changes. A Digital Object Identifier is one common example. Persistence depends not only on identifier syntax but also on continuing organizational commitments to maintain the identifier and its resolution service.
Rich metadata may describe a dataset’s title, creators, subject, spatial and temporal coverage, methods, formats, and access conditions. Inclusion in a repository index or data catalog allows users and software to discover the resource without already knowing its identifier.
A filename on a public web server is therefore not sufficient by itself: it may lack a durable identity, searchable description, or clear connection between the data and their metadata.
Accessible
Accessibility concerns how an identified resource can be retrieved and how its availability is documented.
| Principle | Requirement |
|---|---|
| A1 | Data and metadata are retrievable by their identifiers using a standardized communications protocol. |
| A1.1 | The protocol is open, free, and universally implementable. |
| A1.2 | The protocol allows authentication and authorization where necessary. |
| A2 | Metadata remain accessible even when the data are no longer available. |
Here, “open” describes the communications protocol, not an obligation to make every dataset unrestricted. Retrieval through HTTPS, for example, can coexist with authentication and access control.
A2 distinguishes the availability of the data from the persistence of their descriptive record. If data must be withdrawn, deleted, or moved into restricted storage, retained metadata can still explain what existed, who created it, and why it is unavailable. Such records support citation and interpretation of earlier research without implying that the underlying files remain downloadable.
Interoperable
Interoperability concerns whether data and metadata can be meaningfully exchanged and combined across systems.
| Principle | Requirement |
|---|---|
| I1 | Data and metadata use a formal, accessible, shared, and broadly applicable language for knowledge representation. |
| I2 | Data and metadata use vocabularies that themselves follow FAIR principles. |
| I3 | Data and metadata include qualified references to other data and metadata. |
Interoperability involves more than selecting a file format that multiple programs can open. Systems must also be able to interpret the meaning of fields, values, units, entities, and relationships.
A domain ontology or controlled vocabulary can give stable meanings to terms that would otherwise be ambiguous. For example, two datasets may both contain a field labelled “temperature,” but meaningful integration also requires information about units, the measured object, and the measurement conditions.
Technologies such as the Resource Description Framework and JSON-LD can support machine-readable semantic descriptions, but the principles do not mandate either technology.
A qualified reference states the nature of a relationship, rather than merely providing an untyped link. Relations such as “is a revision of,” “was derived from,” or “was generated using” convey information that software can use to interpret connected resources.
Reusable
Reusability concerns whether a resource is sufficiently described and governed for subsequent use.
| Principle | Requirement |
|---|---|
| R1 | Data and metadata are richly described with accurate and relevant attributes. |
| R1.1 | Data and metadata are released with a clear and accessible data usage license. |
| R1.2 | Data and metadata are associated with detailed provenance. |
| R1.3 | Data and metadata meet domain-relevant community standards. |
Reuse requires both practical interpretability and clear conditions of use. A dataset may be easy to download yet difficult to reuse if its variables are undocumented or its legal permissions are unclear.
Provenance records how data were collected, processed, transformed, and assembled. It can include instruments, methods, software versions, processing parameters, and relationships to source datasets. Data lineage captures related information about the movement and transformation of data through processing systems.
Community standards provide discipline-specific expectations. Depending on the field, these may cover measurement units, minimum reporting information, terminology, file structures, or experimental descriptions.
FAIR and open data
FAIR and open data overlap, but they are not equivalent.
Open data concerns whether people may access, use, modify, and redistribute data under appropriate open terms. FAIR concerns the properties that make resources discoverable, retrievable, interpretable, and reusable, including by machines.
Consequently:
- Publicly downloadable data may be poorly FAIR if they lack persistent identifiers, documentation, meaningful structure, or explicit reuse terms.
- Access-restricted data can have FAIR properties when their descriptions are discoverable, access procedures are standardized, and authorized users receive sufficient information for reuse.
- Open and FAIR data can satisfy both broad access goals and technical requirements for effective reuse.
Restrictions may arise from privacy, confidentiality, consent, contractual obligations, or intellectual property. FAIR does not remove these restrictions or establish a legal basis for disclosure. Nor does describing a resource as FAIR establish that its collection or use was ethically justified.
Implementation in data systems
FAIR implementation usually involves coordinated work across the data lifecycle rather than a single action at publication.
A typical implementation includes:
- Planning and description. Select relevant community standards, define variables, and document collection methods and anticipated access conditions.
- Identification. Assign stable identifiers to datasets and, where appropriate, their versions and related research objects.
- Structured metadata. Supply both human-readable explanations and machine-readable descriptions.
- Repository services. Provide identifier resolution, searchable indexing, standardized retrieval, and durable metadata records.
- Semantic integration. Use documented vocabularies and explicit relationships to connect resources.
- Reuse documentation. Record licenses, provenance, processing history, and domain-specific information.
An application programming interface can make metadata or data available to automated clients. However, an API alone does not establish FAIRness: its documentation, identifiers, semantics, access conditions, and long-term maintenance also matter.
Similarly, depositing a dataset in a database or repository does not automatically make it FAIR. Some properties depend on the repository infrastructure, while others depend on the information supplied by data creators and curators.
Example
Consider a repository of river-water measurements. Its implementation could include:
- A persistent identifier for each published dataset version.
- Searchable metadata describing locations, sampling dates, measured substances, and collection methods.
- Retrieval through a standardized protocol, with restricted access where justified.
- Variables linked to shared vocabularies and units expressed using the International System of Units where applicable.
- Explicit links between raw observations, corrected values, and the software used for processing.
- A clear license and documentation of instrument calibration and measurement uncertainty.
These features address different FAIR requirements. None, individually, establishes that the measurements are correct or suitable for every later analysis.
Assessment and limitations
FAIRness is commonly assessed using indicators, questionnaires, maturity models, and automated evaluation tools. Such assessments examine properties including identifier persistence, metadata completeness, protocol support, license information, and semantic references.
The original principles do not define a single universal score, fixed implementation, or mandatory certification procedure. Assessments therefore differ in their interpretation, scope, and weighting. A score is meaningful only in relation to the criteria used and the resource being assessed.
Automated checks are particularly useful for detecting machine-readable metadata, resolvable identifiers, and explicit license statements. They are less able to determine whether documentation is scientifically adequate, terminology is used correctly, or provenance captures the information needed for a specific reuse.
Several distinctions limit what a FAIR assessment can establish:
- FAIRness is not data quality. Data quality includes properties such as accuracy, completeness, consistency, and fitness for purpose. A well-described dataset can still contain serious errors.
- FAIRness is not reproducibility. Reproducing a result may also require executable code, computational environments, dependencies, and details not captured in dataset metadata.
- FAIRness is not ethical approval or legal compliance. Technical accessibility and clear documentation do not settle questions about consent, rights, or acceptable use.
- FAIRness is not guaranteed permanence. Long-term availability requires sustained funding, preservation work, and institutional responsibility.
Implementation also entails costs for curation, terminology maintenance, infrastructure, and stewardship. Different research communities have different levels of standardization and support. Meaningful evaluation therefore needs to distinguish mechanically detectable features from the domain-specific information required for actual reuse.