Which tool to choose
For writing and reviewing questions, ChatGPT is enough: it works well both in drafting and as a critical reviewer of your drafts. What makes the difference isn't the model, but using it in two opposite phases, first as author and then as checker.
The interfaces change: all you need is the chat. No special feature to find.
How to do it
A badly written question pollutes the answer before anyone even answers. The job is to write questions that don't push toward an outcome, and the AI helps both to write them and to expose the crooked ones.
- Define the goal in one sentence: what you want to know and to decide what.
- Ask for the questions with precise constraints on neutrality. The working syntax:
Help me write 6 questions for a survey on customer satisfaction at my shop. The respondents are regular customers. I want to understand what makes them come back and what annoys them. Neutral, non-leading questions, ordered from general to specific, with the answer type suited to each one (scale, multiple choice, open answer). Avoid double-barreled questions and words that steer the answer.
- Move to the critical check. Paste your questions and ask for the error hunt:
These are my questions: [paste here]. Find the ones that are leading, ambiguous or double-barreled, explain why each is problematic and rewrite it in a neutral version.
- Keep the survey short: every extra question lowers the number of people who make it to the end.
A concrete example
A bookseller wanted to understand why some customers weren't coming back and had jotted down questions like "How much did you like our impeccable service?". She pasted it into ChatGPT, asking it to take it apart. The AI pointed out that "impeccable" already suggested the answer and pushed toward the positive, and rewrote it as "How would you rate the service you received, from 1 to 5?". It also corrected a double-barreled question that asked about prices and selection together, splitting it into two. The revised survey collected more honest answers, and the real problems emerged.
When it does NOT work (and how to fix it)
If the questions steer the answer
Adjectives like "excellent" or "impeccable" inside the question push toward an outcome. Fix: run each question through the AI's critical check, asking it to flag the steering words, and replace them with neutral phrasing. A neutral question doesn't let people guess which answer you're expecting.
If one question contains two
"Are you satisfied with prices and selection?" can't be answered if you feel differently about the two. Fix: ask the AI to spot the double-barreled questions and split them. One question, one thing only.
If you trust the questions without testing them on anyone
The AI has its own biases and doesn't know your audience: it can write questions that are clear to it but misunderstood by real people. Fix: before launching, do a pretest with three or four people from your target and ask the AI to simulate different answers to flush out the ambiguities. The human test isn't to be skipped.
A tip from someone who really uses it
Use the AI in the two opposite roles in the same session: first as author of the questions, then as a strict reviewer trying to demolish them. It's in the second pass that the survey really improves. And once you've collected the answers, you can paste them (without data that identifies people) and have it help you find the recurring patterns, but keep the conclusions under your own judgment.
Frequently asked questions
Can it be done with the free version?
Yes. Writing and reviewing survey questions is text: the free versions handle it without trouble, both in drafting and in the critical-check phase.
Can the AI also analyze the answers collected?
Yes, by pasting the data it can help you find recurring themes and summarize the open answers. First, though, remove any information that identifies who answered: for the analysis you need the content, not the names.
Does the AI write the perfect survey on its own?
No, and trusting it blindly is the surest way to collect useless data. It writes good drafts and spots many errors, but it carries the biases of the texts it was trained on and doesn't know the language of your specific audience. A question that seems clear to it can confuse the people who actually read it. The pretest with real people remains the step that separates a reliable survey from one that only measures noise.