Data Management Plans for AI Research: What Funders Require in 2026

A two-page document buried in supplementary materials is quietly sinking otherwise strong AI proposals. The data management and sharing plan — required by every major federal funder — has become the section where reviewers catch the gap between a researcher's ambitions and their actual readiness to handle the datasets, trained models, and code that modern AI work produces. At NIH, program staff now review these plans against a detailed compliance checklist. At NSF, proposals submitted without one are returned without review. For AI research in particular — where a single project can generate terabytes of training data, dozens of model checkpoints, and thousands of lines of preprocessing code — the stakes of getting this section wrong have never been higher.

Browse our AI Grants page for current opportunities across NIH, NSF, DARPA, and DOE.

The NIH DMSP: Two Pages That Carry Legal Weight

Since January 25, 2023, every NIH grant application that generates scientific data must include a Data Management and Sharing Plan under policy NOT-OD-21-013. The plan is not advisory. Once approved, it becomes a binding Term and Condition of the Notice of Award — meaning failure to follow through can jeopardize future funding.

The plan must cover six elements: the types of scientific data to be managed and shared; related tools, software, and code needed to access the data; any applicable standards for format and metadata; data preservation, access, and timelines; oversight responsibilities; and how costs are budgeted. Scientific data must be shared no later than the time of publication or the end of the award period, whichever comes first.

Starting October 1, 2024, NIH added a reporting requirement: annual Research Performance Progress Reports now ask whether data has been shared, where it was deposited, and under what persistent identifiers. This is not a checkbox exercise. Program officers are reading the answers and comparing them against the approved plan.

For AI projects, the DMSP is where reviewers look for specifics. A genomics lab training a deep learning model on electronic health records needs to describe not just the clinical dataset but the model architecture, the training code, any derived features, and how access controls will protect patient data while still allowing reproducibility. Vague language about sharing "relevant outputs" is exactly the kind of phrasing that triggers revision requests.

NSF's Parallel Requirement — And Its Directorate-Specific Wrinkles

NSF's two-page Data Management and Sharing Plan serves a similar function but operates under directorate-specific guidance. The Computer and Information Science and Engineering (CISE) directorate, which funds most NSF AI research, expects plans to address software and code sharing alongside datasets. The Engineering directorate has its own guidance emphasizing reproducibility standards. Some solicitations override the two-page limit entirely.

The content requirements are broad: types of data and materials produced, metadata standards, access and sharing policies, provisions for reuse and redistribution, and archiving plans. But for machine learning research, reviewers increasingly expect to see specifics that the template does not explicitly demand — model versioning strategy, computational environment documentation, and plans for sharing not just final models but intermediate checkpoints and evaluation scripts.

NSF also now treats AI tools as a factor in research integrity. The PAPPG 24-1 Supplement updated the definition of research misconduct to include AI-based tools. That policy context makes the data management plan doubly important: it is where you demonstrate that your project's AI outputs will be documented and traceable, not just performant.

FAIR Principles Meet Machine Learning: Model Cards and Datasheets

The FAIR principles — Findable, Accessible, Interoperable, Reusable — were written for datasets, but funders now expect them applied to trained models and code. NIH's Bridge2AI program, which the NIH Council of Councils approved for Stage 2 on January 29, 2026, has developed standardized schemas for model card documentation that are becoming the de facto benchmark.

Model cards, originally proposed by Mitchell et al. (2019), are short documents accompanying trained models that specify intended use cases, performance benchmarks across demographic or phenotypic groups, training data provenance, and known limitations. For federally funded work, a model card should address at minimum: what data the model was trained on and its demographic composition, how performance was evaluated and on what held-out sets, what the model should and should not be used for, and what biases or failure modes have been identified.

Dataset documentation follows a parallel framework. The "Datasheets for Datasets" approach introduced by Gebru et al. (2018) asks creators to document motivation, composition, collection process, preprocessing steps, intended uses, distribution details, and maintenance plans. For AI proposals to NIH or NSF, incorporating these structured documentation commitments into your DMSP signals methodological rigor in a way that reviewers notice.

Hugging Face, which now hosts over 1.4 million models, has operationalized model cards as a metadata standard — and funders are watching. If your plan says you will deposit models in a public repository, reviewers will expect you to describe what documentation accompanies them.

Where AI Applicants Get It Wrong

NICHD's Office of Data Science and Sharing published a catalog of common DMS plan deficiencies that reads like a diagnostic checklist for AI researchers:

Vague data descriptions. Plans that describe outputs as "model results" or "processed data" without specifying formats, volumes, or species/domain tell reviewers nothing actionable. An AI plan should distinguish between raw training data, processed feature matrices, model weights, evaluation metrics, and inference code — each with its own format, size estimate, and sharing pathway.

Over-reliance on generalist repositories. NICHD found that applicants default to generalist platforms like Zenodo, Figshare, or Dryad when domain-specific repositories exist. NIH policy explicitly prefers discipline-specific repositories. For AI in biomedicine, that might mean depositing imaging data in The Cancer Imaging Archive, genomic data in dbGaP, or clinical models in repositories that enforce access controls appropriate for patient data. Use a generalist repository only when no domain-specific option fits.

Confusing retention with sharing timelines. Researchers sometimes state they will share data for five years — matching their institution's records retention policy — when most repositories maintain data indefinitely. The plan should reference the repository's retention terms, not your filing cabinet's.

No budget for data management. The DMSP is not free to execute. Storage fees, curation labor, de-identification pipelines for sensitive health data, and persistent identifier minting all cost money. NIH expects these costs in the budget justification. For AI projects generating terabytes of imaging or genomic data, underfunding this line item is a credibility problem.

Ignoring code and software. Many AI proposals treat the DMP as a data-only document. But NIH and NSF both expect plans to cover software, tools, and code needed to access and use the shared data. For a machine learning project, that means version-controlled code repositories (GitHub, GitLab), environment specifications (Docker containers, requirements files), and clear licensing terms.

Build the Plan Around Your Pipeline, Not the Template

The most effective DMSPs for AI research are structured around the project's actual data pipeline rather than the six-element checklist. Start with your data flow: raw collection, preprocessing, feature engineering, training, evaluation, deployment. At each stage, identify what artifacts are produced, what format they take, how large they are, and where they will be deposited. Then map those artifacts back to the plan's required elements.

This approach produces plans that are specific enough to survive NIH program staff review and detailed enough to serve as an actual project management tool — which, given the new annual reporting requirements, is exactly what they need to be.

For researchers navigating the intersection of AI methods and federal compliance, Granted can surface the right funding opportunities and help you understand what each funder expects before you start writing.