TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Judgment · AGE 16—18 · FACTUAL · DURABLE
Polish reads as truth
Smooth text feels more certain than its evidence, and the confidence you feel is yours.
The idea
A paragraph that reads smoothly feels more certain than its evidence. That is a fact about readers, not about truth. Psychologists have repeatedly found that text which is easy to read tends to be judged more true and more familiar. A language model produces smooth text by construction, because it picks the most likely next word every time, so its polish costs nothing and proves nothing. The confidence you feel when you read it is yours. You lent it to the sentence.
Why it matters
The card CONFIDENT ≠ CORRECT describes the machine's end of this. This is your end. The trap closes only if you read polish as a reason to check less.
See it in the studio
A junior asks a model for the fire rating of a partition system and gets three clean sentences with a number in them. They read like a manufacturer's data sheet. Nobody opens the data sheet. The number goes into the specification, and it is wrong by one hour.
Watch for this
The small relaxation in your shoulders when an answer arrives tidy. Notice it. That is the moment when checking stops.
Try it
Take one paragraph a model wrote for you. Rewrite it clumsily: same claims, worse sentences. Show both to a classmate and ask which they trust. Then check the claims. The trust and the truth will not line up.
Prove it
Explain why polish tells you nothing about whether a claim is correct, and name the one thing that does.
How it works
The effect is older than machines, and it has its own literature. Experiments on what psychologists call processing fluency — Reber and Schwarz in 1999, and Alter and Oppenheimer's 2009 review of the field — found that statements which are easier to read are judged more likely to be true. Newspapers, brochures and confident speakers have always borrowed authority from it. What is new is the supply. A model generates finish for free, at any length, on any topic, including topics it has no reliable knowledge of. The studio end of this is only observational so far: one 2026 analysis of nearly ten thousand conversations between general users and one chat model found that when the model produced a finished artefact (a document, a piece of code), users were less likely to question its reasoning or notice missing context than in conversations where they kept refining. That analysis is about behaviour in a chat window, not about fluency as the psychologists mean it, and it cannot say why. The fix is a procedure. Decide what you will verify before you read, so the finish cannot lower the bar afterwards.
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.