Overview

Lesson 3.1: FAIR Principles and AI Research Data

This lesson teaches you how to apply FAIR principles (Findability, Accessibility, Interoperability, Reusability) to research data in projects where AI has played a significant role. You'll learn to create data management plans that account for AI, select and document repositories appropriately, implement metadata standards that support discovery of AI-assisted data, and address unique reproducibility challenges that arise when AI is involved in data creation or processing.

Title

Lesson 3.1: FAIR Principles and AI Research Data

Purpose

This lesson teaches you how to apply FAIR principles (Findability, Accessibility, Interoperability, Reusability) to research data in projects where AI has played a significant role. You'll learn to create data management plans that account for AI, select and document repositories appropriately, implement metadata standards that support discovery of AI-assisted data, and address unique reproducibility challenges that arise when AI is involved in data creation or processing.

FAIR Principles: Why They Matter More Than Ever for AI Research

FAIR principles were formalized in a landmark 2016 paper in Scientific Data (Wilkinson et al.) as a set of aspirational standards for research data: data should be Findable, Accessible, Interoperable, and Reusable. In the decade since, FAIR has moved from aspirational to expected, many funders now require FAIR-compliant data management plans, and the European Commission and major research councils have embedded FAIR into open science mandates.

AI reshapes why FAIR matters and what FAIR requires. On one hand, AI makes FAIR more important. Machine learning thrives on well-documented, consistently formatted, richly described datasets. An AI tool attempting to discover and integrate data across repositories is far more sensitive to FAIR failures than a human researcher who can phone a colleague or interpret ambiguous metadata with domain knowledge. When data is not findable by machines, it is excluded from the large-scale AI-driven analyses that are becoming standard in many fields. When data is not interoperable, AI integration across datasets is fragile or impossible. FAIR compliance is increasingly a prerequisite for your data to be useful in the AI era.

On the other hand, AI makes FAIR more challenging. AI introduces new types of data artifacts, prompts, interaction logs, model configurations, fine-tuned weights, for which no established FAIR conventions yet exist. AI-generated data may have complex provenance that is difficult to describe with standard metadata schemas. And the question of what constitutes the 'data' in an AI-assisted study is often genuinely unclear: is the primary data the inputs, the outputs, or both?

At the advanced level, your task is not just to apply FAIR to your own research data. It is to extend FAIR thinking proactively to the AI components of your research, to help establish norms in your field for how AI-related artifacts should be described and shared, and to design your data management plans with sufficient depth to satisfy not only today's FAIR requirements but the more detailed requirements that are clearly coming.

Findable: Making AI-Assisted Research Data Discoverable

Findability requires that data have persistent, unique identifiers; that data be registered in searchable resources; and that metadata be rich enough to enable discovery even without knowing where to look.

For AI-assisted research, findability raises specific questions. What data objects require identifiers? Traditionally, you assign a DOI to your dataset. In AI research, you may need separate identifiers for: (1) the primary dataset used as input, (2) the AI-generated outputs (annotations, summaries, classifications, synthetic data), (3) the prompts or prompt templates used, and (4) the model configuration or fine-tuned weights. Each of these is potentially a distinct data object that other researchers might want to find, cite, and reuse independently.

Metadata richness for AI-assisted datasets requires fields that standard metadata schemas may not yet address. Consider what a researcher discovering your AI-generated annotation dataset needs to know: What model generated the annotations? What version? What prompts were used? What was the model's performance on a validation set? What are the known failure modes? How were AI outputs validated or corrected by human reviewers? Standard Dublin Core or DataCite metadata schemas include fields for creator, title, description, and license, but not for AI model version or prompt template. You will need to supplement standard metadata with custom fields or use emerging specialized schemas (such as those being developed by communities like Schema.org and bioschemas.org for AI/ML data).

Registry and discoverability also require thought. Depositing your dataset in a general-purpose repository like Zenodo ensures findability by search engines and AI harvesters. But for AI-assisted datasets, you may also want to register in domain-specific repositories that are developing AI-specific discoverability features (such as Hugging Face Datasets for ML-adjacent research, or domain repositories that have added model card-style documentation). The choice of repository affects not only discoverability but also the metadata standards expected and the community that will discover your data.

Accessible: Ensuring Appropriate and Persistent Access

Accessibility under FAIR does not mean open access to all data. It means that clear, documented, machine-readable conditions for access exist, and that those conditions are persistent. Data can be accessible under FAIR principles even if access is restricted, as long as the restriction conditions are clearly specified and enforced through a transparent mechanism.

For AI research, accessibility raises two important questions. First: What are the access conditions for AI-generated artifacts? If you generated annotations on a restricted clinical dataset using AI, the annotations themselves may or may not be as sensitive as the original data. This requires careful analysis. Some annotations (e.g., disease classification labels) may carry nearly as much diagnostic information as the underlying data; others (e.g., structural syntax annotations on anonymized text) may be safely shared even when the source data cannot be. Developing clear, documented access conditions for each type of AI-generated output is part of your data management responsibility.

Second: What about the AI model itself? If you fine-tuned a model on your data and the resulting model is a significant research artifact (as is often the case in computational biology, NLP, and related fields), the model's accessibility conditions must be documented. A model card, now a standard documentation format for AI models, addresses this by specifying what the model does, how it was trained, what data it was trained on (at a high level, respecting any data access restrictions), how to access it, and what its limitations are. Sharing model cards alongside or instead of the full model weights is an emerging practice that balances accessibility with practical and ethical constraints.

Persistence of access is also critical for AI artifacts. Unlike a dataset that can be archived in a repository indefinitely, AI models may depend on specific computational environments, libraries, or hardware. A model that requires a specific GPU architecture and a specific version of a deep learning framework may become inaccessible within years. Documenting the computational requirements and providing containerized environments (Docker images, reproducible conda environments) alongside model weights significantly extends practical accessibility.

Interoperable: Enabling Integration Across AI Research Datasets

Interoperability requires that data use standard formats, standard vocabularies, and standard ontologies that other systems and datasets can work with: enabling cross-dataset integration, meta-analysis, and automated use.

AI research creates interoperability challenges at multiple levels. At the data format level, AI outputs may be generated in vendor-specific formats, proprietary schemas, or ad-hoc structures that other researchers cannot easily parse. AI-generated annotations in a commercial annotation tool's native format, for example, are less interoperable than the same annotations exported to a standard format like JSON-LD with vocabulary drawn from a standard ontology. When choosing AI tools and output formats, prioritize those that support standard, open formats for export.

At the vocabulary level, AI-generated data often uses the terminology of the specific model or vendor, which may differ from the standard vocabulary of your research domain. A large language model generating disease classifications may use different terminology than the standard ICD-11 codes your domain uses. Mapping AI outputs to standard ontologies, and documenting that mapping, is essential for interoperability. Tools like BioPortal (for biomedical ontologies) or the Gene Ontology provide controlled vocabularies that can anchor AI-generated classifications to established domain standards.

At the model level, interoperability is emerging as a priority through initiatives like the ONNX (Open Neural Network Exchange) format, which enables AI models trained in one framework to be deployed in another. For research purposes, sharing models in interoperable formats rather than framework-specific formats (PyTorch .pt vs. Hugging Face transformers vs. ONNX) significantly expands the community that can access and reuse your models.

Perhaps most importantly for cross-study integration, the metadata describing AI processes needs to be interoperable. Emerging standards like the Model Card format, Data Sheets for Datasets, and the Croissant metadata format for ML datasets are working toward interoperable description of AI training data and model characteristics. Adopting these standards, even before they are universally required, positions your research at the leading edge of FAIR practice for AI and contributes to their establishment as community norms.

Reusable: Documenting AI Research Data for Future Use

Reusability requires that data be described with sufficient richness that others can understand and appropriately reuse it, including understanding its provenance, its limitations, its license, and the context of its creation. For AI research data, reusability documentation is substantially more complex than for traditionally-generated data.

Provenance documentation for AI-assisted data must describe the chain of decisions that produced it. This is the 'data lineage' question: Where did the input data come from? How was it preprocessed? What AI model processed it, in what version, with what parameters, with what prompts? How were outputs validated? What human review or correction was applied? What are the known biases or limitations of the AI system that generated the data? Each of these questions represents a dimension of provenance that affects how other researchers can appropriately use your data. Omitting provenance documentation is the FAIR equivalent of reporting your experimental result without your methods section, the result exists but cannot be evaluated or built upon.

Limitation documentation is especially critical for AI-generated data. AI systems have known failure modes. They may underperform on minority groups, unusual examples, or out-of-distribution cases. If your AI-generated annotations are systematically less accurate for certain demographic subgroups, this must be documented in your data description so that reusers can account for it or avoid inappropriate use. Failure to document AI limitations in reused data has produced downstream research that inherited and amplified those limitations without awareness.

Licensing for AI research data raises novel questions. The license on your input data may constrain what you can share about the AI outputs. If your input data was licensed under a Creative Commons Non-Commercial license, and you used it to train a model, may you share that model? Legal opinions vary. At minimum, document the license of each artifact clearly and transparently note any licensing constraints that affect reuse. Where possible, use open licenses (CC0, CC-BY) for AI artifacts you create from public domain or permissively licensed sources.

Finally, reusability requires that data management plans (DMPs) be updated proactively as AI use evolves during a project. Most funders accept DMP updates during a project period. When you add an AI component to your workflow mid-project, update your DMP to document the new data artifacts being created, their storage and retention plans, and their expected sharing approach. This keeps your FAIR documentation current and accurate.

Writing an AI-Aware Data Management Plan

Data management plans have historically focused on primary research data: what data will be collected, where it will be stored, how long it will be retained, and how it will be shared. An AI-aware DMP must address the same questions for the additional data artifacts that AI creates.

Begin your DMP by identifying all data artifacts your project will create, treating AI artifacts with the same rigor as primary data: the input datasets, the AI-generated outputs (annotations, summaries, classifications, synthetic data), the prompt templates, the interaction logs, and any trained or fine-tuned models. For each artifact, answer: What format will it be in? Where will it be stored? Who will have access and under what conditions? How long will it be retained? Will it be shared publicly, conditionally, or not at all, and why?

For AI models, your DMP should address model documentation standards. At minimum, commit to preparing a model card for any model you plan to share. Specify the metadata standards you will use to describe the model, the repository where it will be deposited, and the license under which it will be released.

For AI interaction logs, your DMP should specify log format, storage location, retention period, and whether logs will be shared alongside published outputs. Increasingly, journals and repositories expect this documentation.

For datasets that mix human-generated and AI-generated content, common in AI-assisted annotation projects, your DMP should explain how the mix will be documented. Downstream researchers need to know what percentage of annotations were human vs. AI-generated, and how discrepancies between human and AI annotations were resolved. This 'data provenance transparency' is essential for reusability.

Finally, your DMP should acknowledge that your AI tool landscape may change during the project and commit to DMP updates when new AI components are added. A DMP that accurately described your workflow at project initiation but has not been updated to reflect AI tools added 18 months later is misleading documentation.