The Calculator Problem: Should Responders Be Allowed to Use ChatGPT?

By RepData

• article

• Advanced Statistical Techniques
• Agile Quantitative Research
• Artificial Intelligence
• Brand Research
• Concept Testing
• Customer Experience (CX) Feedback
• Data Analytics
• Data Connectivity
• DIY Surveys
• Full Service Research
• Sample Management
• Survey Research
• Participant Recruitment
• AI

This article is based on the webinar “From Access to Trust: What 20 Years of Online Sample Teaches Us About Surviving the AI Era”. Watch the replay below:

Picture a respondent halfway through your survey. They’ve got your questionnaire open in one tab and an AI chatbot in another. They get to an open-ended question, aren’t quite sure how to answer and ask the chatbot for help. Maybe they copy and paste the response. Maybe they rewrite it a little to sound more like themselves. Is that cheating? Is it fraud? Or is it perfectly reasonable for someone to use a tool that is now part of everyday life?

We have talked about this quite a bit, and there are good arguments on both sides. Whether AI assistance is acceptable depends in part on what the study is trying to measure. But researchers need to figure out where to draw the line and make those expectations clear to respondents, rather than leaving it as an unstated assumption that only surfaces when someone gets quietly excluded from a dataset.

The Algebra-Class Question

The comparison that keeps coming back to us is the one from school: could you use a calculator on an algebra test, or did the arithmetic have to happen entirely in your head? Some teachers said no tools, full stop—the whole point was proving you could do it yourself. Others argued that a scientific calculator was a legitimate instrument, and refusing to let students use it meant testing their arithmetic skills rather than their understanding of algebra. Neither position is necessarily wrong; it just depends on what you are actually trying to measure.

The same principle applies to AI-assisted survey responses. Surveys are designed to capture different kinds of information, from immediate emotional reactions to carefully considered opinions and experiences. Whether using AI compromises the results depends on what the researcher is trying to learn.

Stop Making the Respondent Guess

We think the solution is fairly straightforward, and it is a fairly simple one. Researchers need to be explicit about their expectations; stop leaving this unsaid. Surveys already include plenty of instructions telling respondents how to answer questions, from “select all that apply” to “please read carefully before answering”. There is no reason the use of outside tools should be left to the respondent’s interpretation.

If a particular study needs someone’s honest, unfiltered opinion, without AI assistance or polish, say so directly in the instructions. Explain that we want their genuine reaction, however it comes out, rather than a perfectly crafted answer. If someone subsequently uses a chatbot to complete that survey, excluding them is reasonable because the expectations were clear from the outset. In other studies, where outside tools do not compromise what is being measured, allowing AI assistance may be perfectly acceptable. Either way, respondents should not be left to guess what constitutes “cheating,” and researchers should have a clear basis for deciding which responses to include or exclude.

What Kind of Thinking Are You Actually Trying to Catch?

There is a deeper reason this matters beyond fairness to the respondent, and it comes down to what a question is designed to capture. Some of what we measure in research is deliberately fast and instinctive – what Danny Kahneman might classify as System 1 reactions. Especially for something like a concept test or message test, we want those immediate reactions before respondents have had time to overthink or censor their answers. Other times, we want the slower, more deliberate System 2 thinking that comes from a well-probed open-ended question. 

Even when we invite reflection and additional thought, we still want a human reaction. If a respondent routes a question through a chatbot before answering, we may lose both signals. Instead, we get a chatbot’s interpretation of the question, presented as the respondent’s own thinking. That is fundamentally different from what the research was designed to capture.

After all, AI is not buying anyone’s products or services. People are, and understanding how they think, whether instinctively or deliberately, is the reason we conduct this research in the first place. A respondent might deploy an AI-powered grammar checker or proofreader to improve their language or grammar, but we don’t want that to get in the way of providing their true thoughts.

Question Design Is a Better Defence Than Any Detector

This is also where we would encourage researchers to examine their own survey design, rather than focusing exclusively on respondent behaviour. A question asking someone to “carefully review this message and tell us everything you like and dislike about it” is practically an invitation for the kind of polished, essay-style chatbot response we want to avoid. A question designed to capture an instinctive, in-the-moment reaction is less conducive to that kind of AI assistance – not because respondents cannot still use a chatbot, but because the differences between an immediate human reaction and an AI-generated response may be easier to identify. Good detection matters, but thoughtful question design is just as important in protecting the integrity of the data.

Where This Leaves Us

While we are worried about AI-generated responses, “No AI, ever” may be too blunt a rule, especially as AI use becomes more common. On the other hand, researchers must establish reasonable guardrails for acceptable AI usage or the effect might be “AI always.” We need to be more deliberate about whether outside tools help or compromise what we are trying to measure, communicate those expectations clearly to respondents and design questions that elicit the kind of thinking we actually want to capture. That is a more durable approach than relying solely on after-the-fact detection, because it establishes clear expectations from the outset and helps protect the integrity of the research.

Frequently Asked Questions

1. Can I inspect and challenge its answer?

Make it show its working. If complaints increased, compared with when? Among all respondents, or only those mentioning food? Expect clear filters, sample sizes, supporting comments, and appropriate statistical tests. Three memorable complaints don’t establish the size of a problem, however quotable they might be.

How do you tell respondents what AI use is acceptable?

Be direct about it in the survey instructions. If you need unfiltered human responses, say something like: “We want your own opinions and experiences, in your own words. Please don’t use ChatGPT or other AI tools to write your answers. Don’t worry about spelling or grammar.” The instructions can vary by study, but respondents should know what’s allowed before they answer, especially if AI use could result in their responses being excluded.

Can you detect AI-generated survey responses?

Yes, although AI-generated text is getting harder to identify reliably just by reading it. That’s why survey fraud detection needs to look beyond the answers themselves. Behavioural analytics can help identify suspicious typing patterns, mouse movements, copy-and-paste activity and other signs that a respondent may be using AI to complete a survey. Combining behavioural signals with device intelligence and other respondent validation checks gives researchers a better picture of what’s happening during survey completion. No single check tells the whole story.

How should I write open-ended questions to get more authentic answers?

Think about what you actually want to learn from the respondent. If you’re testing a concept or advertising message, asking “What’s the first thing that comes to mind?” may tell you more than requesting a detailed list of likes and dislikes. If you want a more considered opinion, ask follow-up questions that get into someone’s personal experiences or reasons for feeling a certain way. Good survey design makes it easier for people to answer naturally and gives researchers more useful human insights to work with.

This article is based on the webinar “From Access to Trust: What 20 Years of Online Sample Teaches Us About Surviving the AI Era”. Watch the replay below:

Authors

Learn more about