Synthetic data has gone from curiosity to category — augmented sample, LLM "respondents," digital twins, simulated audiences. Used well, it is a genuine accelerator. But it is built from real human answers, calibrated to them, and trustworthy only when validated against them — which is why targeted, quality global panel and CATI interviewing remain the decision-grade ground truth. This is a practical guide to using synthetic data for what it is good at, without mistaking it for the market.
CatalystMR Research Team. (2026). Real, Synthetic, or Both: A Methodology for Sourcing Decision-Grade Data in the Age of AI. CatalystMR Methodology Papers. https://www.catalystmr.com/insights/methodology-papers/real-synthetic-or-both/
@techreport{catalystmr_real_synthetic_or_both,
author={{CatalystMR Research Team}},
title={Real, Synthetic, or Both: A Methodology for Sourcing Decision-Grade Data in the Age of AI},
institution={CatalystMR}, year={2026}, type={Methodology Paper},
url={https://www.catalystmr.com/insights/methodology-papers/real-synthetic-or-both/}
}TY - RPRT AU - CatalystMR Research Team TI - Real, Synthetic, or Both: A Methodology for Sourcing Decision-Grade Data in the Age of AI PB - CatalystMR PY - 2026 UR - https://www.catalystmr.com/insights/methodology-papers/real-synthetic-or-both/ ER -
Artificial intelligence has made it cheap to generate data that looks like survey responses, and an industry is forming around it: statistical sample augmentation, LLM-generated synthetic respondents, digital twins, and simulated audiences. For research buyers the noise is loud and the framing is often wrong — "real versus fake." Synthetic data is neither a fraud to be dismissed nor a replacement to be trusted on faith. It is a powerful, fast, increasingly capable derivative of real human data.
This paper is a practical, vendor-neutral guide to using it well. It offers a field guide to the four kinds of synthetic data; maps where each genuinely earns its place; shows where it breaks, and why real, verified human data remains the decision-grade ground truth that synthetic is trained on, validated against, and bounded by; sets out a framework for choosing to ask, simulate, or blend, with governance; and closes on the complementarity of the two. The throughline: targeted, quality global panel and CATI are not the alternative to synthetic data — they are the foundation that makes it trustworthy.
Synthetic data is no longer hypothetical. Industry analysts forecast it will supply most AI training data within a few years; the largest panel and platform companies are shipping synthetic respondents; and the category now spans everything from statistical sample-boosting to LLM-generated personas. The unhelpful reaction is to argue authenticity. The practical question is narrower: which questions can a model trained on yesterday's human data answer well — and which still demand a real, verified human today?
Two truths sit behind those figures. Synthetic data is real enough to be useful — and derivative by construction. Every synthetic respondent is produced by a model that learned from real human responses; it interpolates what it has seen, it does not observe the world anew. That is why the same momentum that makes synthetic data attractive also makes real, verified human data more valuable, not less: it is the ground truth that trains the models, the benchmark that validates them, and the renewable source that keeps them from drifting away from reality.
"Synthetic data" is one phrase for several very different things, and conflating them is how buyers get burned. The four below differ in how they are made and what they can be trusted for — but all share one root: each is generated by a model that learned from real data. Knowing which kind a vendor is selling is the first act of due diligence.
Statistical augmentation — trained on a study's own data and checkable against a holdout — is a very different proposition from an LLM persona answering from pre-trained weights. The first extends real data; the second improvises in its absence. A buyer who hears only "synthetic respondents" cannot tell which they are getting — so the first question is always: made how, from what data, and validated against what?
A method partner that only argued against synthetic data would not be worth reading. Used well, it removes genuine friction — and the evidence that it can work, under the right conditions, is real. The point is to match the tool to the task.
Synthetic respondents perform best on structured tasks — ranking, pricing, sentiment — and, as augmentation, on stabilising subgroups when trained on solid real data. Peer-reviewed work shows language models conditioned on real human backgrounds can emulate subgroup response patterns with surprising fidelity,1 and data-science research has found no significant difference between some predictive models built on well-made synthetic data and on real data.4 The opportunity is real; so are its limits.
Ask: "Is this directional, exploratory, or low-stakes — or decision-grade?" Synthetic shines on the first; the second still needs real respondents.
Every kind of synthetic data sits on a foundation of real human responses. That dependency is also its constraint: a model can only reflect the world it was trained on, and pushed past that, it averages, distorts, and drifts. Read the stack from the bottom up.
Only as representative as the data beneath them — and prone to under-dispersion, bias amplification, and drift when pushed past it.
Learn patterns from real responses; they interpolate what they have seen — they do not observe the world anew.
The only layer that observes reality directly. Remove it and the stack has nothing to stand on — trained recursively on its own output, a model collapses.
The practical question is rarely all-or-nothing. It is which method fits this decision, given its stakes and the data you already hold — and how to use synthetic data without quietly lowering the evidentiary bar.
Disclose where synthetic data is used and how it was made; validate it against a real holdout; and demand independent validation rather than vendor-selected metrics. Conduct and quality frameworks apply to synthetic data as to any other method.5,6 The governance risk is real: analysts forecast that most data leaders will struggle to manage synthetic data well.
Ask: "How was this validated against real respondents — and by whom?" If the only proof is the vendor's own, treat the result as a hypothesis, not a finding.
The strongest programs do not choose once and for all; they run a cycle in which real and synthetic data feed each other — and the cycle only holds because real, verified responses anchor it at both ends.
Targeted panel & CATI capture verified, decision-grade responses.
Augmentation and persona models are trained and calibrated on that real data.
Boost thin cells, pre-test concepts, simulate what-ifs — fast and cheap.
Returned to real respondents for the calls that carry real risk.
Synthetic data has earned a permanent place in the research toolkit — for augmenting thin cells, pre-testing, simulation, and privacy. Used for those jobs, and validated honestly, it is fast, scalable, and genuinely useful. But it is a derivative: trained on real human responses, calibrated to them, and bounded by them — and recursive over-reliance degrades it. So the rise of synthetic data does not diminish the real respondent; it raises the value of a clean, representative, decision-grade ground truth. Choose to ask, simulate, or blend by the stakes of the decision; disclose and independently validate what is synthetic; and treat targeted, quality global panel and CATI not as the alternative to AI, but as the foundation it stands on. The ICC/ESOMAR Code and the ISO 20252 framework let buyers ask for that rigour in consistent terms.5,6
CatalystMR is a global market-research panel and fieldwork partner specialising in targeted, verified sample across healthcare, B2B, and niche audiences, delivered by online panel and live CATI interviewing. We treat real, decision-grade human data as the ground truth — and advise honestly on where synthetic augmentation or simulation adds value around it.
Compliance posture: aligned to the ESOMAR Code and Guidelines and the ISO 20252 framework; certified under the EU–U.S., UK, and Swiss Data Privacy Frameworks, with personal data siloed from response data.