In September, public historians from around the world met in Lisbon for the 8th International Conference on Public History. The theme was “The Public History of Difficult Pasts.” We brought a talk called “AI for Archives,” and we brought it as a website rather than a slide deck: a single scrolling page with two interactive explainers, so that the room could stop, ask, and try the controls for itself. The presentation is online here.
A conference about difficult pasts was the right room for it. The pasts that are hardest to tell are often the least described, and what is not described is hard to find.
The same past, again and again
We opened with a proposition rather than a warning about hallucination: the risk is not that AI will invent the past. It is that AI will keep finding the same past.
AI systems meet archives through the records and descriptions available to them, and most collections have very few. In the United States alone, 21,588 history organizations hold collections, and roughly four in five of the private nonprofit ones run on less than $200,000 a year. Few of them can afford to describe what they hold in depth. Meanwhile, new text is now easy to produce: in one recent study of English-language articles published online, about half of the new ones were primarily written by AI. Description of primary evidence is scarce; synthetic writing is abundant. A system that learns and searches from what is abundant will keep returning to the same well-described past.
Digitization alone does not change this. A scanned photograph or letter can still be dark if no one has named what is in it, placed it in context, or connected it to anything else. Preserved is not the same as described, and described is not the same as found.
Three ways to use the same model
If AI is becoming part of how people discover the past, the question is not simply whether archives use it. The question is how the work is organized, and who directs it.
The first explainer followed one model through three arrangements. Working alone, it produces fluent language by predicting likely words, with no verified evidence from the collection. With retrieval, the system scores and ranks a set of approved sources and places the best matches in front of the model, so its answers can point to evidence; but ranking is arithmetic, not judgment about truth or significance, and what is left out can still flatten the story. Inside a curator-directed process, the model works within a written charter for the collection, explicit rules about evidence and ethics, a record of uncertainty, and review by people who know the collection.
These are not three levels of truth. Each arrangement adds something useful and opens new ways to fail. We asked the room to carry away a few distinctions: fluency is not evidence; retrieval is not judgment; a procedure is not control; and a human in the loop is not, by being there, meaningful oversight. An agent can follow a process. A process is not yet a curatorial practice.
Teaching without retraining
The second explainer took up a question we hear often: does ARCADE train its own model? It does not. ARCADE begins with a model that has already been trained, and it never changes the model’s weights. Instead, it surrounds the model with layers of direction: curatorial skills, a charter for the collection, retrieved evidence, a workflow with separate stages, iteration, and, at the end, a curator’s decision.
We use the word “teaching” deliberately, and carefully. Those materials change how the model approaches the task in front of it, but the model forgets ARCADE when they are taken away. The lasting learning happens in the practice. When a curator corrects a description or rules on a hard case, the ruling can revise the charter, the skills, the precedents, and the tests. The model does not change with the encounter. The practice does.
What we did not show
The talk was an opening argument, not a finished case. It did not present comparative results on discovery, independent evaluation, stable costs, or evidence that the method transfers from one institution to another. It only gestured at labor, rights, environmental cost, and the authority of communities over their own records. A better version would follow one real collection through the practice, show a correction that mattered, and make the trail of evidence visible. That is the work of the pilots now beginning.
We ended where the argument begins, with responsibility. We taught computers to read and share our records. Now we have to teach them to be useful curatorial partners. AI can propose. Curators question, revise, and decide.
Sources
- American Association for State and Local History. 2022 National Census of History Organizations. AASLH, 2022. https://aaslh.org/2022-census/.
- ARCADE. “AI for Archives: Two Interactive Explainers.” Presentation, 8th International Conference on Public History, International Federation for Public History, Lisbon, September 2026. https://arcadecuration.org/ai-for-archives-presentation/.
- Paredes, Jose Luis, Gregory Druck, Bevin Benson, and Ethan Smith. “AI Now Writes as Many Online Articles as Humans.” Graphite, May 2026. https://graphite.io/five-percent/research/ai-now-writes-as-many-online-articles-as-humans-do.
A note on authorship: This post is based on ARCADE’s presentation “AI for Archives” at the 8th International Conference on Public History, Lisbon, September 2026. ARCADE is built through AI-inflected practice. The team asked Claude to draft this post from the presentation and its notes. It was rewritten by Mark Tebeau and Claude Opus 5.5. Effective use of AI announces its provenance and use.