The Missing Piece Should Stay Visible in an AI Reconstruction
A completed ruin can help visitors understand a place. It should also let them see where the evidence ends, which interpretations remain possible, and what the image maker supplied.
A roof changes the story
A museum screen shows a ruined courtyard beneath a complete timber roof. In this hypothetical exhibit, the foundations survive, the roof does not, and the reconstruction team considers both a covered hall and an open court plausible. An AI image tool supplies beams, shadows and weathering. The resulting picture is easy to read. The disagreement has disappeared.
That omission matters because the roof changes the visitor's interpretation of the space: how people gathered, what they could see, and how light entered. A small label saying that AI assisted the image leaves the central question unanswered. Which part of this apparent building rests on surviving evidence?
This essay proposes an exhibition practice: keep evidence, scholarly inference and illustrative completion separately inspectable, with the distinction visible wherever it changes the story. The aim is to make reconstruction useful without making a missing object appear to have been found.
An existing obligation
The London Charter, version 2.1, dated February 2009, already addresses this problem. Its documentation principles ask visualisations to disclose what they represent and the nature of uncertainty, and to explain the connection between sources, reasoning and results. This reaches beyond naming the software: readers need access to the interpretive decisions that produced the image.
The Seville Principles, ratified in December 2017, make the archaeological obligation more specific. They call for distinguishable levels of reconstruction accuracy and allow alternative interpretations when scientifically comparable. They also distinguish surviving remains, physical rebuilding and virtual reconstruction. These are professional principles, not evidence that any particular museum interface successfully communicates uncertainty.
Our application begins with the point of encounter. An exhibition cannot assume that visitors will read a technical report before looking at the screen. The first view should carry the essential distinction; further explanation can unfold on request. Otherwise, the institution has documented its doubts while presenting certainty.
This differs from keeping separate versions of a restored family photograph. The public exhibit must explain a shared historical argument to people who may encounter it only briefly. File preservation is necessary, but it does not determine what those visitors learn.
What inscription research establishes
The 2022 Ithaca study supplies a bounded reason to welcome AI assistance. It concerns ancient Greek inscription text, including restoration and attribution, rather than pictures of buildings. In its restoration evaluation, historians assisted by Ithaca achieved 71.7% top-choice accuracy. Testing hid characters from otherwise undamaged text, providing known targets. The training corpus also retained earlier scholarly supplements; the authors discuss the resulting issue of data circularity and present predictions as possibilities for scholarly assessment.
The distinction is decisive. Success on an inscription task supports exploring assistance in that task. It cannot establish that a generated roof, statue or painted surface recovers its historical appearance. Nor does agreement between a model and an existing interpretation necessarily provide independent confirmation when that interpretation helped shape the model's evidence.
The useful exhibition question is what a proposal lets someone examine. A candidate can direct attention to a comparison, expose an assumption, or suggest a question for further research. Its value need not depend on promoting it immediately into the museum's authoritative view.
Alternatives visitors can inspect
A Nottingham Trent doctoral study of Hawara, dated November 2023, developed a virtual experience with three historical reconstructions of the Egyptian labyrinth. Its evaluation involved 52 participants in Cairo, with English comprehension required and university students the largest group. Most respondents rejected the suggestion that multiple scenes were confusing, although some found them confusing. This was a prototype survey, not a controlled demonstration of lasting learning or an evaluation of AI generation.
The design possibility is valuable even with those limits. Competing views can make disagreement an object of attention. In the hypothetical courtyard exhibit, switching between a roofed and an open version could reveal exactly which interpretation changes. Keeping the viewpoint and surrounding geometry steady would help isolate that difference.
Alternatives should earn their place through evidence. A gallery of attractive generated variants could teach that archaeology is a choice of styles. Each displayed candidate instead needs a short reason for consideration and an account of what would count against it. Where one interpretation has stronger support, the presentation should say so. Visitors should not have to infer scholarly standing from image quality.
Designing a reversible exhibit
The following design is our proposal, offered for testing. Start with a view of the documented remains, then let visitors add the interpreted structure and illustrative surface details. The default image may combine them, provided a persistent legend and marked boundaries make the consequential additions apparent. A visitor should be able to remove those additions without losing orientation.
A selected element should answer a concrete question. For the hypothetical roof, a note might identify the surviving support positions, name the comparison used to propose the span, and explain that the timber texture is illustrative. These statements separate the basis for the geometry from the choice of appearance. An overall confidence badge would hide that difference.
Use words and outlines alongside colour so that the distinction survives monochrome printing and does not depend on colour perception. Avoid making an unsupported percentage look like a measurement. Where evidence is qualitative, explain its kind: a surviving fragment, a comparative example, a contested reading, or a detail added only to make the scene legible.
Reversibility also needs to extend to revision. The team should be able to change a disputed feature without rebuilding the whole interpretation. Record which explanatory captions depend on it. If the roof changes, the account of lighting may need to change too; updating the picture while preserving its old narrative would leave a contradiction in the exhibit.
For reuse outside the gallery, export a paired view showing remains and reconstruction, with the relevant qualification attached. This addresses the circulation problem explored in synthetic historical narratives: an image can travel beyond the interface that explained it. A caption cannot prevent every misleading crop, but the museum can supply a version that carries its reasoning into ordinary reuse.
Test what the visitor understood
A visible legend is a design feature. Whether visitors understand it is a separate empirical question. Before adopting the proposed display, ask visitors to identify a surviving element, an inferred feature and a detail whose appearance remains unknown. Then ask what evidence would make them prefer one reconstruction over another.
Compare a simple labelled image with the layered version. Keep the historical content comparable, and include people who choose a flat display over a headset. Record confusion and mistaken certainty alongside interest and enjoyment. These are proposed evaluation tasks, not results established by the cited studies.
More controls may produce less understanding if they compete for attention. If the layered view fails, simplify it; a careful pair of still images may serve the purpose better. The missing piece should remain visible in the visitor's account of what is known, even when the screen offers a compelling picture of what might have been.
Sources
- Hugh Denard, editor, The London Charter for the Computer-Based Visualisation of Cultural Heritage, version 2.1, February 2009; especially principle 4.
- ICOMOS, The Seville Principles: International Principles of Virtual Archaeology, ratified December 2017; especially principle 4.
- Yannis Assael and colleagues, Restoring and attributing ancient texts using deep neural networks, Nature, 2022; evaluation and data-circularity discussion.
- Farida Waheed (repository record: F. Mekheimar), Virtualizing the uncertainty of digital archaeological reconstructions: application on the Egyptian labyrinth of Hawara, PhD thesis, Nottingham Trent University, November 2023; pp. 174–176, 208–210, 226–229, 251–252.
Sources consulted October 2, 2026. The studies describe their own research settings; they do not measure current reconstruction tools generally.
Related reading
Production: commissioned by the site operator; researched and drafted by GPT-6 Astra; editorial and source review by the coordinating AI assistant.