When the map loses its footnotes: AI, rare books, and historical‑GIS research

HTML: Perplexity AI
Source: what is doom loop
21 September, 2026

Update: please find at bottom a link to an in-depth curriculum from Spatial Reserves

Fourth in a series (last here) reviewing trials & tribulations in AI today. Fifth & last one follows here.

The first three posts in this series traced the AI doom loop, the rare‑book “buy → despine → scan → discard” pipeline, and the legal and copyright fault lines. This fourth post brings the story down to the level of historical and geospatial research: what happens to projects that rely on old maps, local histories, and specialized monographs when those sources are privatized, degraded, or made less accessible.

AI-generated by author (Medium)

---

Why historical‑GIS depends on “boring” books

Work in historical GIS, spatial history, and local heritage often leans on:

• Old gazetteers and place‑name dictionaries that document vanished settlements, changed borders, and local variants.
• Regional and local histories with detailed descriptions of land use, transport routes, and administrative boundaries.
• Specialized monographs in archaeology, agrarian history, and historical geography that never made it into mainstream digital collections.

These are exactly the kinds of long‑tail, pre‑digital texts now being targeted for bulk acquisition and private AI training.

---

Risks for research workflows

Loss of physical copies and marginalia

When rare or unique volumes are despined and discarded:

• Handwritten notes, ownership marks, and inserted materials disappear.
• Binding structures and annotations that help date or contextualize a text are lost.
• Future scholars cannot re‑examine the physical artifact with better imaging or new questions.

For historical‑GIS, where provenance and context often matter as much as the text itself, this is more than an abstract loss.

Privatization of key references

If the only high‑quality digital versions of certain books exist inside private AI training corpora:

• Researchers cannot cite or verify passages directly from a stable, accessible edition.
• AI tools may “know” content that humans cannot easily check or reference.
• Reproducibility suffers: two researchers using different models may get different “facts” from the same underlying source.

Distortion through AI summarization

As historians and geographers increasingly use AI to:

• Summarize old texts
• Extract place names and dates
• Generate candidate entries for gazetteers or attribute tables

there is a risk that errors or simplifications in model outputs become embedded in datasets and maps, then feed back into future models—a classic doom‑loop pattern, now in spatial form.

---

Opportunities, if governance is right

The same technologies can also help historical‑GIS work, if the underlying data is governed well:

• Better OCR and layout analysis
  Modern models can dramatically improve transcription of old fonts, maps, and tables, if trained on diverse, high‑quality historical material that remains accessible.
• Named‑entity and place‑linking
  AI can assist in linking historical place names to modern coordinates, but only if the source texts are citable and verifiable.
• Cross‑collection discovery
  Well‑designed, transparent corpora can help researchers find relevant monographs, articles, and archival references across institutions and languages.

---

Practical stances for researchers and mappers

For those working in historical GIS and related fields, a few practical positions emerge:

• Prefer open, citable digital editions
  When possible, base analyses on texts available through libraries, archives, or trusted repositories, not solely on AI‑generated summaries.
• Document AI use explicitly
  In methods sections, note where AI was used for transcription, extraction, or interpretation, and keep raw outputs where feasible.
• Support public‑interest digitization
  Advocate for library‑led or consortium‑led digitization that preserves originals and makes scans at least partially open to research.
• Push for transparency
  Encourage vendors, publishers, and AI providers to disclose major training data sources, especially when they affect domain‑specific knowledge.

---

A map without reliable footnotes is dangerous; an AI‑assisted historical‑GIS workflow without transparent, accessible sources is worse. Keeping the tiger and the child from circling into butter means insisting that the texts behind our maps remain traceable, citable, and common.

Note: these practices also slot in with a last example on the main blog here, of AI complementing but not replacing GIS in a geo-history context.

Author's footnote

Spatial Reserves free online course: GIS and Society (web view)

“... [@josephkerski] rigorously tested, taught, refined, and evaluated this course in higher education full-semester settings, short courses, online and face-to-face, in a wide variety of settings for a wide diversity of GIS data users and analysts in multiple locations around the world.

• The course is served as a story map collection
  Available here. Here is a playlist of the course videos: https://youtube.com/playlist?list=PLL_rej_JJ9qk&si=rIdPRLtIyLG04Pd9
• It is also described in the following essay:
  https://community.esri.com/t5/education-blog/a-new-complete-course-in-data-gis-and-society/ba-p/1402213
• Plus, the course is also served as an ArcGIS Experience, here:
  https://experience.arcgis.com/experience/726bab5044f74bb4a1de4965de679c7

This course invites learners to work with real data to solve problems and foster skills through a set of readings, hands-on activities, discussions, quizzes, and a final project.  I also invite instructors who seek to incorporate meaningful data and discussion about data in their courses to use all or part of this course in their own courses and programs.”