AI-generated respondents, synthetic populations, and other model-driven approaches are increasingly being discussed as though they represent the next evolution of primary research. The underlying premise is straightforward: if a model has been trained on enough information about a particular audience, perhaps it can predict how that audience would respond without requiring researchers to conduct a traditional survey. It's an interesting proposition, and one that deserves thoughtful discussion rather than an emotional reaction. At the same time, I find myself returning to a question that I don't think our industry has fully explored: When did modeling people become the same as talking to them?
The more I think about that question, the more I believe we're beginning to blur the line between two activities that have always served different purposes. Both have value. Both can inform better decisions. But I don't believe they produce the same type of evidence, nor do I think they should be viewed as interchangeable.
Whether we're conducting surveys, interviews, focus groups, ethnographies, or observational studies, the objective is fundamentally the same. We begin with a question, identify the population we want to understand, collect information directly from that population, and analyze what we've observed. Sometimes the research confirms what we thought we knew. Sometimes it tells us we were wrong. Sometimes it reveals that the world has changed in ways we didn't anticipate. Either way, it replaces assumptions with evidence.
Whether we're talking about traditional predictive analytics or today's large language models, the objective is to identify patterns within existing information and use those patterns to estimate what is likely under similar circumstances. That capability is extraordinarily valuable. Organizations have relied on predictive models for decades to support forecasting, demand planning, fraud detection, risk assessment, and countless other business decisions. The distinction is that models estimate. Research measures. Models estimate what people are likely to report. Primary research measures what people actually report. That difference becomes particularly important when the purpose of the research is to understand change.
Some of the most valuable studies I've worked on over the years were commissioned because something no longer made sense. Customer attitudes had shifted unexpectedly. A new competitor had entered the market. Brand perceptions had changed. Adoption of a technology was accelerating more quickly than anticipated. Leadership had begun making decisions based on assumptions that were no longer producing the expected outcomes. In every one of those situations, the client wasn't looking for a prediction. They wanted to understand what was happening now and, more importantly, why. And that's where I think the current conversation around synthetic respondents sometimes misses an important distinction.
Every research study is conducted within a specific moment in time, and respondents don't answer questions in a vacuum. Their responses are influenced by everything happening around them when they participate in the study. Economic conditions, regulatory changes, competitive pressures, technological advances, workforce challenges, customer expectations, and industry news all shape how people interpret questions and formulate their answers. If market conditions have shifted, if a new regulation has taken effect, if budgets have tightened, or if a disruptive technology has emerged, those realities become part of the responses we collect. A model approaches the problem differently. It identifies patterns within existing information and estimates what someone with similar characteristics is likely to say. Depending on the application, those estimates may be remarkably accurate. But there is an important distinction between estimating a likely response and observing one. If circumstances have changed in ways that weren't reflected in the information used to build the model, that change may not yet be reflected in the prediction.
Throughout my career, I've watched industries change remarkably quickly. New regulations have altered purchasing decisions. Supply chain disruptions have shifted investment priorities. Emerging technologies have changed competitive landscapes. Economic uncertainty has influenced budgets, hiring plans, and capital expenditures. Those changes don't simply affect business outcomes. They influence how people think, what they prioritize, and the decisions they make.
When clients commission research under those circumstances, they aren't asking us to estimate what decision-makers are likely to think based on historical patterns. They're asking us to find out what decision-makers actually think today. And that's an important distinction. If the objective is prediction, a well-developed model may provide an excellent estimate. But if the objective is measurement, if the goal is to understand whether attitudes have shifted, whether assumptions still hold, or whether something unexpected has emerged, then estimation and observation are not interchangeable. You can't confirm that opinions have changed without observing those opinions. You can't discover something new by assuming you already know the answer. That's why I don't think the question is whether synthetic approaches can replace primary research. I think the better question is: Replace it for what?
If the objective is to forecast, simulate, or generate hypotheses, model-driven approaches may be exactly the right tool. If the objective is to measure what respondents report at a specific moment in time, particularly when the world around them is changing, then estimation and observation are not interchangeable.
We have decades of research on parenting attitudes, purchasing behavior, media consumption, childcare, education, and family decision-making. A model trained on those data may provide excellent predictions. But being the parent of a young child in 2010 wasn't the same experience as being the parent of a young child in 2026. Technology has changed. Social norms have changed. Economic conditions have changed. Education has changed. Parenting itself has evolved. The question isn't whether a model can make a reasonable prediction. The question is whether those predictions still reflect the reality parents are experiencing today. That's not something we should assume. It's something we should measure.
The mistake isn't using synthetic approaches. The mistake is assuming they answer the same questions as primary research. The moment we stop asking people because we believe we already know what they'll say is the moment we stop measuring and start assuming.
Contact: Ariane Claire, Research Director, myCLEARopinion Insights Hub
A1: Accuracy isn't the same as observation. They answer different questions.
Even a highly accurate prediction is still a prediction. It tells you what's probable based on the past, not what's true right now.
A2: Because the audience you studied years ago may not be the audience you have today.
The question isn't whether the model can make a reasonable prediction. It's whether that prediction still reflects the reality people are experiencing now.
A3: When the objective is prediction rather than measurement.
The tool isn't the problem. The trouble starts when a model is used to answer a question that actually calls for observation.
A4: Because respondents never answer in a vacuum.
Understanding change is often the whole point of commissioning research. You can't confirm that attitudes have shifted without observing the shifted attitudes.
A5: No — it's an argument against assuming they answer the same questions as research.
The moment we stop asking people because we assume we already know what they'll say is the moment we stop measuring and start assuming.