AI and Linked Data, Unleashed

Linked data and metadata infrastructure used to be built mainly by institutions: national libraries, consortia, vendors, standards bodies. A newer pattern is now visible. Scholars with deep knowledge of one field, but limited programming background, use AI coding tools to build metadata infrastructure for that field themselves. That is what “unleashed” means here: not that the AI does the work, but that the people who understand a body of knowledge can now build the tools that organize it.

This note follows one such case through the whole metadata lifecycle, looks at another discipline and at the distance between a personal research graph and linked open data, and ends with five class activities and the questions to keep in view. Every resource named here has its own entry in the Toolbox, so students can move between the case and the standards behind it.

Justin Lewis teaches English at Olympic College. Working in rhetoric, composition, and writing studies, he has built three connected services; he reports building Pinakes and Aletheia with Claude Code. Read in order, they cover deposit, registration, indexing, discovery, and reuse:

The links are not only thematic. Pinakes’s citation networks depend on journals depositing reference lists, which is exactly the gap Eunomia’s Labors fills; Eunomia’s coverage estimates, in turn, rely on Pinakes’s article counts. A journal that registers DOIs but never deposits references can appear in Pinakes as cited but never as citing. Metadata work done upstream decides what can be discovered downstream.

The same move is happening elsewhere. The environmental and digital historian Jim Clifford (University of Saskatchewan) describes, in a guest post for Computational History, replacing chronological Word notes and fixed-column spreadsheets with knowledge graphs. His argument is that a document orders information by when it was found, a spreadsheet by the columns guessed at the start, and a graph by what the information is actually about, so each new note joins what is already known about the same people, places, and commodities.

His practice is instructive for metadata students:

• Each graph is a single JSON file that Claude Code maintains as he describes findings in plain English; a medium-sized project has about 500 nodes (people, commodities, places, institutions) and 1,200 connections.
• For 181 photographed letters, he used Google’s Gemini for handwriting recognition, then extracted structured entities, and checked transcriptions by reading key letters in order.
• He keeps explicit “open question” nodes for unverified claims and unread sources, treating gaps as data.
• He notes the limit himself: the graph “didn’t explain the connection, but it surfaced it.” His personal graphs are not FAIR or interoperable; linked open data would need formal ontologies such as CIDOC CRM, persistent identifiers, and URIs.

From personal graph to linked open data

Clifford places his own graphs at one end of a spectrum. At the other end is LINCS, the Linked Infrastructure for Networked Cultural Scholarship, a Canadian project for humanities linked open data. Its Historical Canadians dataset, which Clifford helped create, draws on the Dictionary of Canadian Biography with Wikidata, models families, occupations, and residences with CIDOC CRM, and can be queried through SPARQL. The underlying logic, entities and relationships, is the same at both ends; the distance between them is ontologies, identifiers, and institutional infrastructure. Clifford adds that his team is experimenting with using AI models to shorten that distance, but that the tooling is still emerging.

Two sets of principles describe what a shared dataset has to answer for, and LINCS commits to both:

• FAIR (Wilkinson et al., 2016): data should be Findable, Accessible, Interoperable, and Reusable. FAIR is about the data and the machines that use it.
• CARE (the CARE Principles for Indigenous Data Governance, drafted in 2018 and published through the Global Indigenous Data Alliance): Collective benefit, Authority to control, Responsibility, and Ethics. CARE is about people and purpose, and was written to complement FAIR, which on its own says nothing about power or history. Local Contexts is one practical way to carry it into metadata.

The Historical Canadians team’s own choice shows the two in tension: they did not carry over the source’s nationality and ethnicity designations, because decolonized vocabularies for First Nations, Inuit, and Métis people were lacking and they did not want to omit those designations for Indigenous people alone. More reusable is not always more responsible.

That last point is where this course’s standards come in. A private graph becomes shared infrastructure only when its entities resolve to identifiers others use (LC NAF, Wikidata), its classes come from a published model (LRMoo, which extends CIDOC CRM), and its rules are stated and testable (DCTAP, SHACL).

Each activity takes about an hour and ends with a short written note. They can be done in any order, but 1 and 2 prepare for 4.

1. Trace one article

Choose an article from a journal listed on Eunomia’s Labors. Find it on the journal’s own site, look up its DOI record through Crossref, and find it in Pinakes. Write: which metadata appears at each stop, which is missing, and where the reference list came from.

2. Read a coverage page

Open Pinakes’s coverage page and pick one journal that appears as cited but rarely or never as citing. Write: a two-sentence explanation of why, using the distinction between registering a DOI and depositing references, and one consequence for a student searching the field.

3. Build and inspect a reader

In Aletheia, build a five-week reader on a topic of your choice and export the Common Cartridge file. Rename the .imscc file to .zip, open it, and read imsmanifest.xml. If you have a learning-platform sandbox, try importing it. Write: what metadata travels with each reading (title, license, source), what does not, and what you would need to check before using the reader with students. Remember that the developer reports testing the export only on Canvas.

4. Audit the build

For one of the three services, list what was AI-assisted, what was checked by a person, and what is missing or unconfirmed (for example, the print PDF that is not tagged for accessibility). Compare your list with the Toolbox’s own account of its AI-assisted revision. Write: which claims you would trust, which you would verify, and how.

5. From notes to linked data

Take a page of your own research notes (or the entities in Clifford’s description). Sketch the people, places, things, and relationships as a small graph, with at least one open question. Then make it shareable: give two entities identifiers from LC NAF or Wikidata, choose a class for each from a published model, and write the rules for one entity type as three DCTAP rows, starting from the sample profile. Write: what the graph shows that your notes did not, and what you lost or had to decide in making it formal. Then test it: is your small graph FAIR? Do the CARE Principles apply to your sources, and if they do, whose authority would you need before sharing it?

• Sustainability. Each of these is largely one person’s work. What happens if the maintainer stops? Licenses matter here: Aletheia’s code is MIT-licensed and its catalog records CC BY 4.0, so others could continue it; a journal using Eunomia’s Labors keeps its own DOI prefix and credentials.
• Scope. Choosing which journals or readings to include is itself an argument about what counts as the field, as Pinakes’s own documentation says. What is left out, and who decides?
• Verification. Lewis audits reference matches by hand; Clifford reads key letters and tracks open questions. Where in each workflow does a person check the machine, and where does no one?
• Access and accessibility. Free to use is not the same as accessible to everyone; an untagged PDF excludes some readers.
• Interoperability. A tool can be useful without being interoperable. When does a personal or disciplinary system need shared identifiers and models, and who pays for that work?
• Labor and credit. Metadata work is often invisible. Who does it in these cases, who benefits, and how would you describe it as a job? (Roles such as linked data librarian and metadata manager are planned for Version 2 of the Toolbox.)
• Dependence on AI tools. All of these were built with commercial AI tools. What would it mean for a field’s infrastructure to depend on them?

Descriptions here are dated: the services were checked on 18 October 2026 and change quickly. This note was drafted with AI assistance from Claude, made by Anthropic, the company that also makes Claude Code, which both Lewis and Clifford report using; read the enthusiasm accordingly, and keep the discussion on the metadata rather than on any particular tool. As in the rest of the Toolbox, an AI draft is a set of claims, and each claim here should be checked against the sources below.

• Lewis, Justin. Eunomia’s Labors.
• Lewis, Justin. About Pinakes and its coverage page.
• Lewis, Justin. “Aletheia,” and the Aletheia repository.
• 1EdTech. Common Cartridge.
• Wilkinson, Mark D., et al. “The FAIR Guiding Principles for scientific data management and stewardship,” Scientific Data 3 (2016).
• Carroll, Stephanie Russo, et al. “The CARE Principles for Indigenous Data Governance,” Data Science Journal 19 (2020); and CARE Principles, Global Indigenous Data Alliance.
• LINCS. Historical Canadians dataset and data management plan.
• Clifford, Jim. “Nodes, Edges, and the Historian’s Craft: Knowledge Graphs as Research Notes,” Computational History, 16 April 2026.