TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Studio Practice · AGE 17—20 · POSITIONAL · METHOD
Verify before you build on it
A claim without a source beside it does not enter the work.
When to use
Whenever an AI answer is about to enter something that carries weight: a site analysis, a drawing, a specification, a report, a jury presentation. Especially anything with a number, a rule, or a product name in it.
The method
Treat the answer as a lead, not a finding. Take each claim that matters and find the evidence that settles it. For a rule or a figure that is usually the primary document: the bye-law text itself, the municipal record, the standard. For a product it is the manufacturer's datasheet, read knowing the maker wrote it. For a scientific question a good review may be stronger than any single paper. Where no one source settles it, find two that do not copy each other. Write the source next to the claim. Keep your notes in two columns: verified, with a source; unverified, still unchecked. Nothing from the second column goes into the work. If a claim cannot be verified, it stays unchecked, and you say so where it appears.
Watch for this
Verifying by asking the same model again, or by asking a second model. That is a second guess, not a check. A check needs something outside the model: a document, a measurement, a person who would know.
The Lab's note
The two columns are the Lab's method, and we hold it strictly: a claim without a source beside it does not enter the work. It is not the only way to verify. Some studios check by sampling; some trust a retrieval system that quotes the document it read. We keep the columns because they make the unchecked visible, and because the cost of a wrong setback is a rebuild.
Try it
Ask a model three questions about the building bye-laws of your city — setbacks, floor area ratio, parking. Then find the actual bye-law document and check each answer. Sort the results: right, wrong, and right-for-a-different-city.
Prove it
Show one confident AI claim you caught by checking it against its source, and the two-column notes that kept it out of your work.
How it works
A model answering from its training produces the likeliest-sounding continuation, not a retrieved fact. The likeliest bye-law for "a city in Karnataka" can be an average of several cities' rules, and an average is not law. Systems with search or retrieval attached can quote real documents, and an answer drawn from a document you supplied is an extraction, not a guess. It still needs reading: a quoted page can be the wrong edition, a real-looking citation can point nowhere, and a primary source can be self-interested. Verification is not distrust of the tool. It is what lets you use the tool at all.
What this idea builds on
What this idea opens up
Sources
- what-ai-actually-is
- practical-applications
- IFLA, How to Spot Fake News
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.