Tithi Mondal

Guide8 min readUpdated

Synthetic Users in UX Research: Uses, Limits, Safe Method

Synthetic users are AI personas that answer like your customers. Learn what they are good for, where they mislead, and how to pair them with 5 interviews.

By Tithi Mondal, UI/UX Designer

Read inहिन्दीবাংলাमराठीதமிழ்తెలుగుગુજરાતી

Synthetic Users in UX Research: Uses, Limits, Safe Method: cover illustration

Suppose a founder in Siliguri is planning a small loan app for shopkeepers. At 11 pm the founder types into a chatbot: "Act as a 45-year-old kirana shop owner and tell me what you think of my app." The answers arrive in seconds. The answers are polite, detailed and encouraging. The founder goes to sleep feeling that research is done.

That chatbot character is a synthetic user. A synthetic user is an AI-generated profile that stands in for a group of users and answers questions as if it were one of them. This guide explains what the method is good for, where it misleads, and how to combine it safely with 5 real interviews.

Checked in November 2026. Tools change quickly, so confirm details on the project's own page.

A note on the name: "Synthetic Users" is also the name of a commercial product. This guide is about the general method, not about one product.

What are synthetic users and AI personas?

Synthetic users and AI personas are simulated people produced by a language model from a short description. A language model is the kind of AI system behind chatbots. The model has learned patterns from very large amounts of text.

A normal persona is a static summary of a user type on one page. My guide on what a user persona is covers that format. A synthetic user is different in one way: you can talk to it. You ask a follow-up question and the model writes a new answer in character.

The problem this method addresses is real. Good research takes time. Finding 5 shopkeepers who will speak for 20 minutes can take a week. A synthetic user is available at once and costs very little. The risk comes from the same speed. The answer arrives so easily that a team forgets nobody was actually asked.

How do teams use language models to role-play users?

Teams use language models to role-play users in 4 simple steps. The steps are the same in a general chatbot and in a dedicated tool.

  1. Describe the person. Write age, city, job, income pattern, phone type, languages and current habits. For example: "Runs a kirana shop in Siliguri, uses UPI daily, reads Bengali and Hindi, borrows from a local lender before festivals."
  2. Describe the situation. Explain what the person is trying to do and what they use today.
  3. Ask interview questions. Ask about past behaviour, problems and reactions to your idea, one question at a time.
  4. Repeat with different people. Create 5 to 8 varied profiles, including people who would not want your product.

Some teams go further and run many AI agents together so that the agents react to each other. That is a group simulation, and the same limits apply.

What is AI user research good for?

AI user research with synthetic users is good for preparation: drafting, spotting obvious gaps and rehearsing. The Nielsen Norman Group, a well-known UX research firm, published an evaluation of synthetic users by Maria Rosala and Kate Moran. The authors conclude that synthetic users can help teams prepare for research with real users. They also conclude that synthetic users should not replace that research.

Here is where I think the method fits a design process.

  • Drafting an interview guide. Ask the synthetic user your questions. Confusing questions become obvious when even a model answers them badly. My guide to user interview questions for app ideas gives a starting set.
  • Finding obvious gaps. A synthetic shopkeeper may ask "what documents do I need". If your flow has no answer, you have found a gap before any real person saw it.
  • Rehearsing. A first-time interviewer can practise staying quiet and asking "can you tell me about the last time".
  • Learning a new domain. The Nielsen Norman Group authors found the method useful as a kind of desk research. Desk research means reading what is already known before you speak with people.
  • Listing hypotheses. A hypothesis is a guess that you plan to test. Synthetic answers give you guesses to check.

Every item on this list produces questions. None of them produces findings.

Where do simulated user interviews mislead?

Simulated user interviews mislead most when you ask what people did, what people will pay, or how a specific group feels. The published work I read points to 4 problems.

Synthetic users agree too easily. In the Nielsen Norman Group evaluation, the synthetic users were overly positive. One synthetic learner claimed to have completed every online course it mentioned. The real learners in the earlier study often dropped out. The authors connect this to sycophancy, which means the tendency of a model to please the person asking.

The answers are shallow. The same article reports that synthetic users listed many needs without saying which needs mattered most. A list without priorities does not help you decide what to build first.

No real behaviour exists behind the answer. A model has never opened your app, stood in a bank queue or missed a loan instalment. A model's story about "last month" is generated text.

Groups get flattened. A paper by Angelina Wang, Jamie Morgenstern and John P. Dickerson studied language models as replacements for human participants. The authors argue that current models are likely to misportray demographic groups and to flatten the variety inside each group. The authors urge caution when the participant's identity matters to the task.

A short profile is a weak base in general. A paper by Joon Sung Park and co-authors describes 2-hour interviews with 1,052 adults in the United States. The team built simulations from that material. The paper states that simulations built from demographic details alone often misjudge individuals. Richer real interviews improved the simulations. The authors report that the distortions were reduced but not removed.

For Indian products I would add one expectation of my own. Language models learn from text written online. A first-time smartphone user in a small town writes very little online. So I would expect a synthetic version of that user to sound more fluent, more digital and more confident than the real person. My post on designing for first-time internet users describes how different that real person can be.

Can AI replace user interviews?

No, AI cannot replace user interviews, because a generated answer is not an observation of a real person. AI can make the interviews you do run sharper. This table shows how I would divide the work.

Question you haveSynthetic userReal person
Are my interview questions clear?UsefulConfirm in the first interview
What objections might come up?Useful as a list of guessesNeeded to see which objections are real
What did people do last time they had this problem?Not reliableNeeded
Will people pay this price?Not reliableNeeded, through real commitments
Can people complete this screen?Not possibleNeeded, by watching them try
How does a specific community feel?RiskyNeeded

What is a safe way to combine synthetic users with 5 real interviews?

The safe way is to use synthetic users only before the real interviews, and to keep the 2 kinds of notes apart. A founder or a small team can do this in one week.

  1. Day 1: write assumptions. List 10 things you believe about your users. Mark each one "assumption".
  2. Day 1: run a role-play pass. Create 5 synthetic profiles and interview each one for 10 minutes. Include one profile that dislikes the idea. Tell the model to be critical and honest.
  3. Day 1: extract questions only. Write down every surprise as a question. "Do shopkeepers trust an app for loans, or only a person they know?" Do not write down any answer as a fact.
  4. Day 2: fix the interview guide. Remove confusing questions. Add the new ones. Keep the guide to 8 or 10 questions.
  5. Days 3 to 6: speak with 5 real people. Use a phone call or a WhatsApp call, in the language the person prefers. Ask about the last real event, not about opinions of your idea.
  6. Day 7: compare. Put the synthetic guesses next to the real answers. Note where the model was wrong. That gap tells you how far to trust the method for your audience next time.

Keep 2 documents: "Guesses from AI" and "Heard from people". Never mix them, and never show AI output to an investor or a team member as research findings. The Nielsen Norman Group authors give the same advice about being open on the source.

Three practical cautions apply. First, do not paste real customer names, phone numbers or interview recordings into an AI tool without checking how that tool handles data. Second, a general chatbot costs little for this work, but dedicated tools and large simulations cost more, so check the current plans. Third, a model can state made-up details with confidence. Verify anything that sounds like a fact.

What does a designer still do with the real findings?

A designer turns what real people said and did into flows, screens and words, and stays responsible for those decisions. Synthetic users cannot take that responsibility.

I treat AI tools as assistants and keep the judgment with people. On the fintech apps I have designed, the important details came from real constraints. Those details included which document a user has at hand, which word for "instalment" a user understands, and what a user does when the network drops. A generated character does not hold a real document or a real phone.

After 5 real interviews, the next steps are ordinary design work. Map the flow, draw wireframes, and test a prototype with real people. My guides on user research for startups and usability testing on a small budget cover both ends of that work.

If you have interview notes and want help turning them into a clear product, you can read about my UI/UX design service.

Sources

Frequently asked questions

Are synthetic users accurate?

They can reflect general knowledge about a group, but published evaluations found them too positive and shallow. Treat every answer as a guess to test with real people.

Can I use ChatGPT as a synthetic user?

Yes, a general chatbot can role-play a user from a written profile. Use it to improve your questions, not to collect findings.

How many real interviews do I need after using AI personas?

For an early idea, 5 focused conversations with people who match your customer are a practical start. Add more if the 5 answers disagree strongly.

Should I show synthetic user results to investors?

Not as research. Label them clearly as AI-generated guesses and present real interview notes and real enquiries as your evidence.

Working on something like this?

Tell Tithi what you are planning. You get a clear scope, a fixed price and a timeline before any work begins. I reply to every enquiry within 24 hours.

Your details go only to Tithi and are used only to reply to you. No newsletters, no sharing.

Have a product in mind? Let us design it properly.

Tell me what you are building. I will reply with questions, a rough timeline and a quote.