How we make quizzes

Last updated: October 30, 2026

Here is exactly how a Nanizo question is made, what checks it, and what those checks can miss.

How questions are written

Most questions are written with an AI model (Claude, made by Anthropic) from a brief we set for each quiz: its topic, its level, and the rule that a question has one right answer and a short fact that adds something.

Some quizzes are built from open data instead. Capitals, flags and regions come from Wikidata, and arithmetic quizzes are generated by a program that computes the answers.

Of the 15,898 English questions on the site today, about 82% were written with AI. The rest come from open data or are generated.

How they are checked

A written question gets two machine checks by AI models from another company, not the one that wrote it:

1. Blind answer. The model sees the question and the options, but not the answer, and picks one. If it picks a different option, or says more than one option could be right, the question is flagged.
2. Fact check. The model sees the marked answer and the fact, and says whether both are correct.

Flagged questions are checked again by a stronger model. Then they are fixed, rewritten or removed. Each question keeps a record of its checks: which model checked it, when, and what it answered.

About 1 in 8 English questions has not had these checks yet, mostly the newest ones and matching-pair rounds. Machine checks catch a lot of mistakes, but not all of them. We do not have editors who read every question before it goes live. That is why every question has a Report link, and why we fix mistakes as soon as we confirm them.

Sources

After you answer, most questions show a Source link to the page the fact comes from: usually Wikipedia or Wikidata, and a fan wiki for fan quizzes. 91% of English questions have one. Photos show their credit and license.

Fan quizzes are unofficial. Names of films, books, teams and games are used only to say what a quiz is about.

Translations

Each edition is translated from English by an AI model. It works from a style guide and a glossary for each language; for example, the glossary holds the names used in that country’s editions of a book series.

A different model then answers every translated question without seeing the answer. Where it disagrees, or the wording reads badly, we fix the question or leave it out of that edition. A quiz appears in a language only after its questions pass these checks.

Corrections

If an answer is wrong, out of date or unclear, tap Report a problem under the question or write to hello@nanizo.com. When we confirm a mistake, we fix it in every language the question appears in. When a fact changes, for example a record is broken or a country changes its capital, we update the question.