FAIR data principles: What makes research data ready for machines and AI

The FAIR principles are a set of guidelines for publishing research data so that it is Findable, Accessible, Interoperable and Reusable. They were written with software in mind as much as people: a dataset is FAIR when a program can locate it, understand what it contains and use it without a human sorting out the details by hand.

A dataset that exists but cannot be used

Picture a lab that posts a spreadsheet of ocean temperatures. The file has columns named T1 and T2, no units, no note on where the sensors sat, and no license. A person could email the authors and ask. A program cannot, and neither can a researcher who finds the file three years later after the authors have moved on.

FAIR is a checklist for fixing exactly that. The data is no longer a loose file but a described, identifiable object that explains itself.

Four letters, fifteen small rules

The principles come as 15 short statements grouped under the four letters. In plain words:

• Findable. Data and its description get a globally unique, persistent identifier (F1), rich metadata (F2), and are listed in a searchable place (F4). Metadata means information about the data: who made it, how, and what each column means.
• Accessible. Anyone allowed to can retrieve the data by its identifier through a standard protocol (A1). The protocol must be open, free and universally implementable (A1.1), and may require a login where necessary (A1.2). The description stays available even if the data is gone (A2).
• Interoperable. The data uses shared, broadly applicable formats and vocabularies (I1, I2) and points to related data with qualified references (I3).
• Reusable. The data has accurate attributes (R1), a clear usage license (R1.1), a record of where it came from, called provenance (R1.2), and follows the standards of its field (R1.3).

The wording above follows the paper as reproduced on PubMed Central and the GO FAIR foundation’s summary.

Where the idea came from

The principles were published in 2016 by Mark Wilkinson and many co-authors in the journal Scientific Data, under the title “The FAIR Guiding Principles for scientific data management and stewardship.” They grew out of a 2014 workshop in Leiden, the Netherlands, called “Jointly Designing a Data Fairport,” and were endorsed by people from universities, industry, funders and publishers. An addendum followed in 2019.

The authors stressed that computers, not just people, need to do the work. Researchers cannot handle modern data volumes by hand, so software agents must find and use data on their own, even when they meet an unfamiliar format. The authors also favor general-purpose open standards over a custom parser for every kind of data.

What “machine-actionable” really means

This phrase appears often in AI-ready data talk, and the paper treats it as a scale, not a yes or no. The more an object tells a software agent, the more machine-actionable it is. A good agent should be able to tell what kind of object it has, judge whether it helps with the task, check whether it may legally and practically use it, and then act.

That is why a license and provenance sit next to file formats in the list. A model-training pipeline that cannot tell whether it may use a dataset has to stop and ask a person.

FAIR in today’s AI push

The same logic runs through the news of this week. Hugging Face has launched Project ProtoStar with five US universities, the Astera Institute and the Open Research Community Accelerator, aiming to get publicly funded data, models and code published in machine-actionable, AI-ready formats. The company argues that public release alone does not make a resource usable when files sit in legacy formats or lack metadata. Our full report on Project ProtoStar covers the partners and the first climate forecasting challenge.

The post does not describe ProtoStar as an implementation of FAIR, and the two are not the same thing. FAIR is a general stewardship framework from 2016. “AI-ready” is a newer, looser goal that adds practical needs such as clean formats a training script can load directly. In practice the two overlap heavily: identifiers, documentation, licenses and provenance help both.

Limits and common misunderstandings

FAIR does not mean open. The authors say the principles do not require data to be public; rich public metadata can make sensitive data highly FAIR even when the data itself stays restricted, and A1.2 explicitly allows authentication. Only the access protocol must be open.

FAIR is also not a standard or a certification. The authors describe the principles as guidance, so two groups can claim FAIR and mean very different levels of effort. Nothing in the principles checks that the metadata is correct, either: a well-labeled dataset can still be biased, small or wrong.

Finally, FAIR takes work. Writing good metadata and provenance costs researchers time, which is one reason infrastructure and funding rules matter. Hugging Face says ProtoStar aims to surface such barriers and good practices that could inform infrastructure, funding and policy.

What to watch next

The first public test is the Climate Risk Intelligence challenge hosted by Carnegie Mellon, running November 2-13, 2026, with final results due in early 2027. It will show how much extra effort reusable data takes in practice. Hugging Face says later challenges from other institutions will be listed on the ProtoStar page.

One caveat applies: Hugging Face is both the host of the project and one of the repositories it wants researchers to use, so its usage figures are the company’s own. Watch for independent measures of whether documented, machine-readable data is in fact reused more.