By Michael Hussey, Founder and President, StatSocial
The rush toward synthetic audiences was inevitable. Traditional market research has frustrated brands for years because it is expensive, slow to field, and often dependent on panels full of professional survey-takers gaming the system for incentives.
So when synthetic audience platforms promised faster answers at a fraction of the cost, the market moved quickly.
Over the last two years, venture capital has poured hundreds of millions of dollars into synthetic audience and AI-driven research platforms. As a result, major research firms and enterprise software companies are rapidly adding synthetic capabilities to their offerings.
Their logic is sound: Why spend weeks running studies if AI can simulate audience reactions in seconds?
The reality, however, is different. Synthetic audiences can sound convincing while still missing the mark on how groups actually behave.
Not all synthetic approaches are alike. Some platforms prompt language models with demographic profiles and let inference fill in the rest. Others ground their synthetic samples in first-party transactional or survey data and have a stronger foundation. The deeper problem applies even to the grounded ones: the behavioral signal underneath is usually too thin to carry the weight of the questions being asked of it.
The models only know what the internet knows.
Synthetic systems are still inference engines. They are not built from direct observation of how real groups behave. Trained on massive amounts of publicly available internet text, they assemble synthetic personas based on how audiences are portrayed online, then generate responses that sound plausible through repetition, assumptions, and the language surrounding certain groups. That is very different from understanding how actual groups think, behave, or make decisions.
You start to see the weakness once certain audiences become far more visible online than others. Some groups generate an endless stream of content, commentary, and discussion. In contrast, others leave behind very little public signal at all, forcing synthetic systems to compensate with inference rather than actual understanding.
As a result, many synthetic audience systems are not really modeling people so much as modeling how people get talked about online.
A 2026 paper from Google DeepMind found that when asked to generate diverse personas, the output collapses around a narrow cluster of stereotypical responses. The models generate responses based on their sense of how a group is described, not on an independent representation of who is actually in it.
That distinction becomes harder to ignore once audiences become more specific. Broad categories like “dog owners” or “Gen Z consumers” are relatively easy for models to approximate because there is so much public discourse attached to them. But once the audience becomes more niche, whether that is oncologists, luxury travelers, or fansof a particular creator, the system starts relying much more heavily on abstraction and inference.
At that point, the system often pieces together synthetic personas from whatever traces of those audiences exist online, even when those traces have little to do with how real people in those groups actually behave.
The outputs can still sound convincing even when the underlying representation is incomplete, flattened, or built on weak assumptions about the audience itself.
Most companies still cannot inspect what their models are doing.
Most synthetic audience platforms also operate like black boxes. Researchers see the outputs without the underlying assumptions. They cannot fully explain how audiences were constructed or why specific responses were generated. That becomes a problem once synthetic outputs start informing business decisions.
Traditional research has plenty of problems, including low-quality panels and unreliable respondents. Researchers can still examine the methodology closely enough to understand how the data was collected, how participants were selected, and what may have influenced the outcome. With synthetic systems, much of that visibility disappears because the assumptions, weighting, and model logic behind the outputs are often difficult to inspect.
Fast iteration is useful. Synthetic certainty is not.
Synthetic systems are clearly useful for rapid testing and directional work. But parts of the industry are already drifting toward something more problematic: treating modeled behavior as if it were interchangeable with direct audience understanding.
The models will improve quickly, but speed and fluency do not solve the underlying problem if the system is still learning from distorted or incomplete behavioral signals. People are not interchangeable audience segments.
Once you move beyond broad audience categories, the gaps start to show up quickly because people with nearly identical demographics often make decisions for reasons that have very little to do with the categories they are grouped into. Synthetic systems tend to iron out a lot of that inconsistency, turning uneven human behavior into cleaner audience profiles that can appear far more dependable than they really are.
The danger is mistaking coherence for understanding.
AI will improve parts of the research process, especially speed and iteration. But fluency is not fidelity, and some companies are already treating synthetic outputs as more authoritative than they really are.
The useful test is simple. Can you inspect the behavioral inputs underneath the model? And do those inputs come from the audience you care about, or from public language about that audience? If you cannot answer both, you are working with something closer to a coherent guess than a research finding.
Research exists to help companies understand how real people behave, not to produce cleaner-looking outputs. Once the outputs start sounding believable, people become far less likely to question whether the system actually understands the audience it is responding to.
