TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Judgment · AGE 17—19 · POSITIONAL · HELD
The model fills gaps it cannot know
Restoration is limited by evidence, and generation is limited only by likelihood.
The dilemma
A photograph of a ruined gateway: the left half standing, the right half fallen. You give it to an image model and ask it to "restore the façade." Back comes a complete gateway, symmetrical, convincing, beautiful. Your tutor asks whether the right half looked like that. You do not know. Neither does the model. Do you put the image on the board?
The choices
Put it up as the restoration — it is the best image you have. Put it up labelled "conjectural reconstruction, AI-generated," and let the label carry the doubt. Or go back to the evidence first: what is documented about the fallen half — a drawing, an old photograph, the fallen stones themselves — and draw the documented part as documented and the conjectured part as conjecture, visibly different, and use the generated image only as a question to put to the evidence.
The consequence
The first choice puts an invention into the record with your name on it, and the record is what the next person restores from. The second is honest and still dangerous: a label is read once and the image is shared a hundred times without it. The third costs an afternoon in the archive and produces a drawing a conservation architect can act on, because it shows where knowledge stops. The charter the profession has worked to since 1964 says restoration "must stop at the point where conjecture begins." A model has no such point. Its whole method is to continue past it.
The case
A measured-drawing studio in a village near Gadag. A team documents a pillared hall whose roof is gone. For the jury one student generates a "restored" view with a full roof. A conservation architect on the jury asks which evidence gave the roof its slope. There was none — the model had given it a slope from a thousand other halls. The team's final sheet shows the standing columns in solid line, the roof in dashed line with the note "form conjectural; no evidence for pitch," and the generated image in a corner captioned as a hypothesis. The jury spent its time on the dashed line. That was the right place to spend it.
Our position
Generation cannot turn missing evidence into known history. Restoration is limited by evidence; generation is limited only by likelihood. A generated completion of a ruin, however the tool was driven, is conjecture that looks like documentation. It belongs in one place only: as a labelled hypothesis beside the evidence, never in the record.
Why we hold it
The conservation charters exist because earlier generations invented freely and called it restoration. Venice (1964) drew the line at conjecture and asked that any new work be distinguishable from the original. An unconstrained model's output erases that distinction by the way it works: the documented half and the invented half come out in one surface, one light, one level of confidence. Putting the line back is the architect's job, and on a ruin it is most of the job.
The strongest objection
Conservation has always drawn conjecture — the dashed line, the reconstruction drawing, the museum model — and it has been useful: it tells the public what a ruin was for and it makes a hypothesis testable. A model is a faster hypothesis engine. If students label its output as conjecture and test it against the evidence, the tool does what the reconstruction drawing always did, at a speed that lets you try ten hypotheses instead of one. Refusing it on principle may protect the record and starve the imagination.
What would make us revise it
Tools that carry the evidence boundary inside the output: the documented fabric rendered from the measured data, the conjectured fabric visibly different by default, the sources attached and still attached after the image is shared. When the distinction survives generation, the position softens to "generate, but keep the line visible." We would also revise if conservation bodies adopted generated reconstructions as accepted hypothesis drawings under stated rules. Reviewed every edition against the charters and the tools.
Prove it
Take one photograph of a damaged building. Produce a generated completion. Then produce a drawing in which documented fabric and conjectural fabric are visibly distinguished, with each conjecture's source — or the absence of one — noted. Explain in a paragraph why the second is a conservation document and the first is not.
Take it to crit
When a student shows a "restored" view of a ruin, ask them to draw a line on it between what is documented and what is conjectured. If they cannot draw the line, the image is an invention, and the crit should be about where the evidence stopped.
How it works
The Venice Charter (ICOMOS, 1964) is exact on this. Article 9: restoration "is based on respect for original material and authentic documents. It must stop at the point where conjecture begins." Article 12: replacements of missing parts "must be distinguishable from the original so that restoration does not falsify the artistic or historic evidence." Article 15 goes further for ruins — "All reconstruction work should however be ruled out a priori," with only anastylosis, the reassembly of existing but dismembered parts, permitted. The Nara Document on Authenticity (1994) widens the frame: our ability to judge authenticity "depends, in part, on the degree to which information sources about these values may be understood as credible or truthful," and those sources — form and design, materials, use, traditions, setting — are judged within the culture the heritage belongs to. A model has no information sources about this building. It has a distribution over buildings. WHY IT MAKES THINGS UP on this map explains the mechanism; this card is that mechanism applied to a ruin. SCAN TO MODEL is where the documented half comes from; WHOSE HERITAGE is what happens when the invented half reaches the people who use the place.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.