AI did not arrive in an empty tool shed.
The canonical Aquinas project began in 1949 as Roberto Busa, IBM, punch cards, operators, and years of checking.DHQ’s reconstructionmakes the human work inside that automation hard to miss. TEI later put shared encoding under community governance. The web made archives into interfaces.
ORBISturned an argument about Roman movement into a contestable network model. TheLDA special issuewas my attempt to make the colonization-or-collaboration problem concrete: put an imported model beside its maker, practitioners, applications, criticism, and tools. These were not steps toward less mediation. They moved the boundary between interpretation and implementation.
AI-assisted coding moves that boundary again. It can produce a transformation, an interface, and a visualization while the question is still being worked out. That is more consequential than calling AI another tool, and less magical than claiming the tool has disappeared.
- This view carries
- Thirteen documented changes in the interface to implementation
- This view drops
- A story of smooth progress or comparable labor across the cases
The interface to implementation moves from punch cards and governed markup to graphical tools, code education, natural-language code generation, and repository agents. Authority moves but never vanishes.
A journal does not simply sample a field.
DHQ reports 710 peer-reviewed articles and case studies through 2025: 386 through the regular stream and 324 through special issues. The journal warns that those routes have different selection histories, so I am not turning them into competing acceptance rates.
The public table of contents gives us a second, narrower measure. In this 806-item corpus, 353 items appear inside 38 named clusters. Those clusters gather work under a question before a reader or recommendation system encounters it.
A subject concentration can record activity in digital humanities, an invitation made by editors, or both. The ability to build a new interface does not redistribute the earlier decision about what enters the interface together.
- This view carries
- Published placement and separately reported publication streams
- This view drops
- Item-level acceptance routes, editorial motives, and comparable acceptance rates
353 items are placed in 38 named public clusters. Placement records an editorial grouping, not a reason for acceptance or a special-issue submission flag.
More names appear on the work.
The standard AI demonstration is solitary: one scholar, one prompt, one finished application. DHQ’s published record was moving in another direction before the present AI wave.
Items with two or more listed authors rise from 32.1% in 2007–11 to 54.7% in 2022–25. The annual line jumps because the journal published between 6 and 79 items per year, so every point keeps its denominator.
A byline is still a narrow record. It cannot tell us who wrote code, cleaned data, designed an interface, found funding, or kept a server alive. It tells us that named coauthorship became more common.
- This view carries
- A bounded change in names listed on published items
- This view drops
- Labor roles, contribution shares, and an explanation for why teams formed
The chart describes listed bylines only. It does not resolve people or infer roles, labor, or reasons for collaboration.
This is what DHQ calls its subjects now.
In the current XML, media studies appears on 36.7% of items published in 2007–11 and 3.4% in 2022–25. History rises from 15.6% to 21.2%; race rises from 0.9% to 14.4%. Project report reaches its highest share, 26.7%, in 2017–21.
These are DHQ’s controlled tags, not topics inferred by a model. They are also multi-label: an item can contribute to several bars. The vocabulary and denominators remain inspectable, which gives us something firmer than a topic cloud.
Read alone, the chart looks like a history of subjects. Before accepting that reading, we need another date.
- This view carries
- Current controlled-tag incidence and publication denominators
- This view drops
- Mutually exclusive topics, field-wide prevalence, and publication-time labels
The bars describe the pinned archive with its current vocabulary. Retrospective keywording means they are not a contemporaneous topic series.
Publication is only one clock.
On July 11, 2023, one commit updated controlled keywords in 615 article XML files. Across the keyword-named commits on July 11 and 12, 697 distinct article files in this corpus were changed.
Each body is tethered to the number of years between publication and that observed repository pass. A current tag attached to a 2008 article can arrive in the Git record fifteen years later.
The previous chart remains true, but its question changes. It shows how DHQ’s current vocabulary describes its published past. It does not reconstruct what every article was called in its publication year.
- This view carries
- Publication dates and an observed repository keywording pass
- This view drops
- First-assignment dates, pre-Git history, and reasons for changing a tag
The chart separates publication time from one observed metadata-change time. Git does not establish when every controlled term was first assigned.
791 records have more than one tag.
Of the 806 published items in this snapshot, 791 carry multiple controlled tags. Five carry one, and ten have none. Multiplicity is the ordinary condition of this archive.
The first button sends every multi-tag record toward one displayed tag. The second keeps the multiplicity in view. The source records do not change; the first Sankey is cleaner because the interface discarded a relation.
A natural-language interface can make the same decision without showing the button. AI-assisted coding gives more scholars the power to put that choice back in front of the reader, provided they know the choice exists.
- This view carries
- Source multiplicity, record conservation, and the exact display rule
- This view drops
- A single true subject or a claim that DHQ used either display rule
Classification policy791 published items are reduced to one displayed tag.
791 published items are reduced to one displayed tag.
Follow the recommendation to its authors.
DHQ’sExplore pageoffers three answers to “what belongs nearby?” Controlled keywords follow an editorial vocabulary. BM25 follows terms in the full text. SPECTER follows embeddings made from titles and abstracts.
This view starts with Anna Sollazzo’s 2026 article, takes the top three recommendations, takes the top three again, and projects those reading routes onto the exact author names printed in DHQ. Switch methods and the neighborhood changes. Across the focal article’s three top-ten lists, 27 distinct articles occupy 30 slots; none appears in all three.
This is an information route, not an issue roster. It shows which named authors a reader can reach after two algorithmic choices. It does not tell us who influenced whom, and it does not quietly merge matching strings into people.
- This view carries
- Two explicit recommendation steps and exact printed byline names
- This view drops
- Person identity, influence, recommendation quality, and reader behavior
Definition of nearby18 printed author names reached through12 articles.
18 exact printed author names are reached through 12 articles. This is a navigational projection, not person identity, citation, influence, or readership.
08All public source articles
The three methods mostly disagree.
The author walk could be an unusual case, so this view uses all 832 public articles present in the three recommendation files. Each point is one source article under one pair of methods, positioned by the number of shared targets in their top ten.
Controlled keywords and BM25 share 0.83 recommendations on average. BM25 and SPECTER share 1.57; controlled keywords and SPECTER share 0.92. Only 224 of 22,416 distinct directed edges occur in all three systems—about 1.00%.
For 646 source articles, not one target appears in all three top-ten lists. The tools do not merely accelerate the same reading practice. They formalize different ones.
- This view carries
- Complete pairwise overlap distributions for 832 public articles
- This view drops
- A judgment about relevance, a winning method, or evidence of readership
Pairwise mean overlap ranges from 0.83 to 1.57 of ten. Similarity depends strongly on the retrieval method.