TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Judgment · AGE 18—21 · FACTUAL · EVOLVING

Where did this image come from?

Human detection of fakes is now close to chance, and the source can still be traced.

The dilemma

A mixed set of building images lands in a precedent study: some photographs, some renders from practices, some generated. All of them are "references" now. You could sort them by eye, looking for the tells — melted text, a shadow going the wrong way. Or you could ask a different question of each one: where did this come from?

The choices

Trust your eye. Or trace each image. Who posted it, and where? Is there an earlier, larger copy somewhere else? Does a second, unconnected source show the same building? Does the file carry a record of how it was made — a content credential, a camera's data, a named photographer? Three questions, in order: where did it come from, who made it, what credentials does it carry. Only then decide what kind of thing it is.

The consequence

Sort by eye and you are betting on a skill that is failing. Studies of how well people tell generated images from real ones now put the score close to chance, and each new model removes the tells the last one left. Sooner or later a building that never existed enters your precedent study with the authority of one that does. Trace instead and the answer no longer depends on your eye. A photograph with a named photographer, a date and a place can be checked. An image that appears nowhere earlier than a design feed, with no maker and no record, is a likeness until proved otherwise — however good it looks.

The case

A precedent board on laterite houses of the Konkan coast: nine images. The tutor does not ask the class to spot the fakes. She asks for the source of each. Six trace to named photographers, a survey book and a practice's own site. One traces to a practice's portfolio page, where it is labelled as a render. Two trace no further than a design feed — no maker, no place, no earlier copy, no credential. One of those two had been cited by three students as "a house in Sindhudurg." Nobody had looked for the house. It was the best-looking image on the board.

Try it

Take a mixed set of nine building images. For each, write three lines: where it came from, who made it, what record it carries — and mark any line you could not fill. Then sort the set by how much you could establish, not by how real it looks. Keep the one that fooled you in your failure catalogue, with the question that would have caught it.

Take it to crit

Hand the student a mixed set. Can they say where each image came from, and what they could and could not establish about it? "I think it is generated" is a guess. "It appears nowhere earlier than this feed and carries no record" is a finding.

How it works

Three layers of evidence, from weakest to strongest. The pixels: the tells (melting text, shadows that disagree, brick coursing that changes module) still exist, but they shrink with every model release, and a 2024 meta-analysis of fifty-six studies put human detection of deepfakes close to chance, so the pixels are a hint at best. The trail: a reverse image search finds earlier and larger copies; a second, independent source confirms the building exists; a named maker can be asked. The record: some tools now attach a signed content credential to a file at the moment it is made, recording the tool, the maker and the edits — the open standard is C2PA. The card MARKED AT BIRTH, on the Ethics & Provenance strand, explains what these marks prove and what a missing mark does not. Two things to hold apart. Provenance is not truth: a genuine photograph can still mislead, and a generated image can be honestly labelled. And a missing record proves nothing: most real photographs carry no credential either. The card is EVOLVING because the tells and the marks both change by the month. The three questions do not.

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.

Age grows from 11 at the centre to 22 at the edge, and six sectors show the learning strands. Tab into the map and the arrow keys step from idea to idea, following the links where there is one. Enter opens the idea under the cursor, and E reads out its links and the reason recorded on each. Press slash for Search, question mark for the full key list, and Escape to leave. Open Ideas for the complete readable list, including what each idea builds on and what it opens up.

LOGIKA · RBDS AI LAB INDIA
ON-RAMP · AGE 11 · FIRST ENCOUNTERS, NOT GATES — IDEAS · — DEPENDENCIES
DONE
OPENS NEXT
SOLID — STANDS ON