How "Create using AI" decides the right question type for every question

New to Create using AI? Read how to create forms and surveys using AI first to see how to use it. This post explains what happens behind the scenes.

Writing a question is not the same as building it

When a creator uses Create using AI, the AI drafts the wording of each question. A survey needs more than wording. Each question also needs an answer type: a short text box, a long comment box, a single choice, checkboxes, a rating, an email field, a date and so on.

That choice matters to the people who answer. "What could we do better?" needs room to write. "How satisfied are you with our service?" works best as a rating. "Select all the features you use" needs checkboxes. The right type is decided by what the question asks for, not by which words it happens to contain.

So after the questions are drafted, the builder shows "Detecting question types..." and picks a type for each one from its wording alone. This post is about how that choice is made.

Step 1: Simple rules for obvious questions

Some questions are plain requests for one field, such as "What is your email?" or "Date of birth". A rule can settle those at once, with no model call. The rules only fire when the whole question is a standard request for that one field. A keyword alone is not enough, because "What is your email?" and "How often do you check your email?" both contain the word email, and they need different answer types.

The rules also catch questions that ask the respondent to connect a wallet or sign in, such as "Please connect your Solana wallet". Everything else goes to the second step.

Step 2: Jev reads the rest

Jev is a decision model. We give it some text and a typed question, and it answers with a choice. It does not write text. For each remaining question we ask Jev one thing: which answer type is the best way to collect the respondent's answer to this question? It chooses from 41 types.

Why Jev. This is classification with fixed labels, which is what Jev's Choice question is built for. The model has to understand what a question asks for, then pick one type from a list. We give each type a short description, and the descriptions are where the hard judgment calls are written down:

• A number type is for one typed numeric value, such as an age in years. When a question names groups or ranges to pick from, like an age group, it is a choice.
• A date type is for one calendar date. "Which date works best?" offers dates to choose between, so it is a choice.
• A wallet or social-login type is only for a question that asks the respondent to connect or sign in. A typed wallet address, username or handle is a short text answer.

Types whose names already say everything, such as email, phone number, website or signature, go to Jev as names only. That keeps each request small.

Rules settle the obvious. Jev reads the rest. The creator never picks a type by hand.

Jev's top answer is the type the question gets.

Eight questions, two steps

We ran these eight questions through the real type picker. Two were settled by rules, and six went to Jev in one request.

Look at the pairs. "What is your email address?" becomes an email field, while "How often do you check your email?" becomes a choice. "Please connect your Solana wallet" becomes a wallet connection, while "What is your Solana wallet address?" stays a short text box. The words overlap, but what each question asks for does not.

Many questions, few requests

A survey has many questions, and the creator is waiting on the screen. So Jev does not get one request per question. The questions are grouped, and each group is one request with one typed question per survey question. Several groups run at the same time.

Rules run first, so questions they settle never reach Jev at all. Only the wording of the survey's own questions is sent to pick a type. No respondent data is involved, because the survey has not been answered yet.