Synthetic respondents vs real human participants is no longer a theoretical debate. Research teams at agencies and in-house insights functions across Australia are actively deciding how much weight to give AI-generated personas alongside data collected from actual people. The choice has real consequences for data quality, ethical compliance, and the decisions that follow.
What synthetic respondents actually are
A synthetic respondent is a simulated profile generated by a large language model or statistical system, trained on existing survey data, demographic databases, or behavioural records. The model produces responses as if it were a person with specified characteristics: age, income, location, purchasing history, stated attitudes. No real individual participates.
Vendors offering synthetic respondents typically claim two things: speed and cost. A project that would take three weeks and $40,000 to field with live panels can, in some configurations, be approximated in hours. That claim has surface-level validity. It also obscures what gets lost.
Synthetic respondents are, fundamentally, recombinations of past data. They reflect whatever signals were in the training corpus. They don't hold genuine opinions. They don't have last Tuesday's experience with a product that shifted how they feel about a brand. And they don't get tired, confused, or irritated by a badly worded question in ways that reveal problems with the instrument itself. For a deeper look at how AI tools are being applied to research data more broadly, the article on how artificial intelligence is transforming the market research industry covers the wider arc of that shift.
Where synthetic respondents perform reasonably well
Honest advocates for synthetic respondents don't claim they replace humans across the board. They argue for specific use cases where the limitations matter less.
Concept screening is the clearest example. If a team has 40 product names and needs to cut to 6 before investing in qualitative testing, running synthetic profiles through the shortlist can identify the obvious losers cheaply. The team still needs real participants to understand why the survivors work. Synthetic respondents handle the filter, not the final verdict.
Stress-testing survey instruments is another defensible use. A researcher can run a draft questionnaire through a synthetic panel to catch structural problems: ambiguous questions, poorly calibrated scales, response options that leave obvious gaps. That's a methodological quality check, not a substitute for field data. It complements good survey design practice rather than replacing it.
Augmenting thin subsamples is a third area. If a study returns only 12 responses from a rare occupational group, some researchers use synthetic augmentation to build a larger working dataset for preliminary analysis. This is contentious and must be disclosed. Done without transparency, it's fabrication.
Where real human participants are non-negotiable
Four research contexts make real participants necessary, not optional.
- Novel experiences. If you're researching reactions to a genuinely new product, service, or social event, training data doesn't contain it. Synthetic respondents will confabulate plausible-sounding responses with no grounding in actual exposure.
- Emotion and motivation. Qualitative research, focus groups, and depth interviews exist precisely because human feeling is not reducible to stated preferences. The pauses, the contradictions, the unexpected associations: a language model smooths these out rather than surfacing them.
- Regulatory and compliance contexts. Healthcare, financial services, and government research often carry legal obligations about sample sourcing. Submitting synthetic data as human-generated research is not a methodology choice; it's a misrepresentation.
- Low-incidence populations. Reaching a specific cohort, say, Australians with a particular medical condition or first-generation business owners in regional Queensland, requires real recruitment. Synthetic models trained on general data produce generalised responses, not the lived specifics of a niche group.
Focus group methodology, in particular, draws its value from group dynamics: how one participant's comment shifts another's framing, how disagreement surfaces tacit assumptions. No synthetic system replicates that. The step-by-step guide to conducting focus groups outlines exactly why the human interaction is the method, not just the data-collection mechanism.
The validity question researchers can't skip
Validity is where the synthetic vs real debate becomes sharpest. A finding is valid when it accurately represents the phenomenon it claims to measure. Synthetic respondents are trained to represent past human responses. That's a different thing.
Predictive validity is the specific concern. A synthetic panel might reproduce the distribution of past survey responses with high fidelity and still fail to predict what real consumers will do with a new product. The model learned from what people said they did; it didn't learn what they'll do in a new context.
Several studies conducted in 2024 and 2025 tested this directly by running identical questionnaires through synthetic panels and live human panels, then comparing outputs. The synthetic panels performed well on attitude scales for established product categories. They performed poorly on purchase intent for unfamiliar categories and on open-ended questions requiring contextual specificity. That's not a minor limitation; for most commercial research projects, purchase intent and open-ended insight are the deliverables.
Ethical and disclosure obligations
Australia's AMSRS Code of Professional Behaviour requires that research findings accurately represent the methods used to generate them. Using synthetic respondents without disclosing this to clients or end-users is a breach of that standard, regardless of whether the data looks similar to human-generated results.
Disclosure isn't just an ethical obligation; it's a practical one. Clients who later discover their strategic decisions rested partly on AI-generated personas, without being told, lose confidence in the agency and in the research process broadly. Transparency preserves the relationship and the credibility of findings.
The profession is currently developing clearer standards on this point. Researchers working in this space should stay current with AMSRS guidance and not assume that because a technique is technically possible, it's methodologically or professionally acceptable without disclosure.
A practical decision framework
The question isn't whether synthetic respondents have any value. They do, in constrained applications. The question is what a specific project actually requires.
Start with the research objective. If the objective is to understand an existing, well-documented market segment's stated preferences on a familiar category, synthetic augmentation carries lower risk. If the objective is to understand how real people emotionally engage with something new, synthetic respondents add noise, not signal.
Ask whether the finding will drive a major business decision. High-stakes decisions, product launches, significant budget allocations, policy submissions, require the highest-fidelity data available. That means real participants, properly sampled. Cutting costs at the data-collection stage to save money that gets spent on bad decisions downstream is not a saving.
Consider the population. Synthetic models perform better when the target population resembles the general public that dominates training data. They perform worse for niche, underrepresented, or geographically specific groups. For Australian regional communities, Indigenous populations, or migrant cohorts, synthetic models trained on global or US-skewed corpora will produce systematically distorted outputs.
Synthetic respondents are a tool. Real human participants are the subject of the discipline. Research that forgets which is which produces findings that look like data but don't function like evidence.