Synthetic data is not good or bad on its own. What matters is what it is being used to represent, how it was generated, what it is replacing, and whether anyone can show it is fit for the research decision being made.
Researchers keep asking whether to embrace synthetic data or worry about it. The better position is neither pro nor anti. It is pro-evidence.
Synthetic data earns its place where it makes research faster, more efficient or more capable. It needs far more caution when its output is presented as direct evidence of what real consumers think and do. Research exists to reduce uncertainty. Technology can speed that up. It shouldn't make anyone feel more certain than the evidence allows. The danger isn't synthetic data. The danger is forgetting the difference between a model of reality and reality itself.
What Is Synthetic Data?
Synthetic data is data a model generates, rather than data observed or collected from the people or events it is meant to represent.
That definition is simple. The trouble is that the label now covers several very different practices, from statistical simulations to AI-generated personas. When people mix them up, the conversation gets muddled. These practices don't deserve the same level of caution, or the same level of trust.
The core distinction is this. Synthetic data can be a research tool. It can also become a substitute for evidence. Problems start when generated data is presented as though it were observed consumer behaviour. That is why provenance, validation, transparency and fitness for purpose matter as much as how smart the technology is.
These Are The Forms Of Synthetic Data
Five forms come up most often. Each has different uses and different risks.
Fully simulated data
Records are generated from statistical assumptions or models. If satisfaction follows a given distribution, what would 10,000 hypothetical respondents look like? This helps with statistical modelling, simulations, questionnaire testing, power calculations and scenario planning. None of these records are real respondents.
AI-generated respondents or personas
A model is prompted to act as a consumer type, such as a 35-year-old Kenyan banking customer who uses mobile banking but doesn't trust digital lending. It is a way to explore possible reactions, language and hypotheses. It is not the same as interviewing 35-year-old Kenyan banking customers, and that should be obvious to everyone reading the output.
Privacy-preserving synthetic data
A model learns from a real dataset and produces new records that share its statistical properties without copying any individual. It helps with data sharing, analytics and model development, software testing and keeping personal information out of view. The key question is whether the synthetic version keeps the characteristics the analysis depends on.
Data augmentation
Extra observations are generated to expand an existing dataset, for example 2,000 real interviews plus synthetic records to enlarge the sample. This is where caution matters most. Synthetic observations should never quietly become additional respondents. If 2,000 interviews were collected, then 2,000 interviews were collected. There may be 20,000 records after augmentation, but that is not a 20,000-person survey.
Synthetic reconstruction or imputation
Models estimate missing information, such as absent demographics, skipped answers, likely behaviour or gaps in an incomplete dataset. Researchers know this one best, because statistical imputation has been around for decades. The rule is that estimated values must stay identifiable as estimates.
What can it do, and what can't it do?
Synthetic data is good at helping researchers build, test and think. It is weak as a stand-in for what real people actually think and do.
What it can do
It can test questionnaires, survey logic and skip patterns before fieldwork. It can run sample-size simulations and model scenarios. It can stress-test dashboards, data pipelines and analytical systems, and give analysts safe data to train on. It can also help generate hypotheses and explore possible reactions before spending money on live research.
What it can't do
It can't give consumers a voice. A model can produce a very convincing answer to "Why don't you use Bank X?" That the answer sounds realistic doesn't mean a real customer ever said it.
There is a deeper limit too. AI models learn from existing human-generated material. Ask one for a typical Kenyan consumer and it may reproduce patterns from its historical training data rather than the current Kenya of today, missing nuance, context and cultural fit.
The gap shows up across geography, age, income, language, culture, digital behaviour, the urban and rural divide, and fast-moving consumer habits. The more specific the population claim, the riskier it is to treat synthetic respondents as a stand-in for real ones.
When it does go wrong, it usually fails in one of six ways.
False precision.
Synthetic datasets produce beautiful numbers. 47.3% awareness, 62.8% satisfaction, an NPS of +41. They look scientific. Precision is not validity.
Bias amplification.
If the underlying data carries bias, the model can reproduce it or make it worse. The result is synthetic confirmation of historical bias.
Homogenisation
AI-generated consumers drift towards the average. Real populations are messy. People contradict themselves and act differently from what they say. Cultures contain subcultures and markets contain outliers. A model smooths that variety away.
Circularity
If synthetic data trains models, and those models generate more synthetic data, the result is a feedback loop. Human data feeds a model, the model produces synthetic data, that data feeds the next model, and so on. Eventually the system is learning from its own assumptions.
Hallucinated consumer insight
An AI can give a very plausible explanation for why consumers behave a certain way, with nothing behind it. This is especially risky in qualitative research, where plausibility is easy to mistake for insight.
Loss of accountability
A client should always be able to ask where a number came from. Conventional research can point to the sample, the questionnaire, the fieldwork, the respondents, the weighting and the analysis. Synthetic data needs a chain of provenance that is just as rigorous.
Realistically, where does it make sense?
It helps to sort applications into three tiers of risk. The tier depends on what the output will be used to claim.
Low risk: building and testing the research machinery
Questionnaire testing, survey programming, skip-pattern checks, sample-size simulations, experimental design and scenario modelling all sit here. So does operational work like testing dashboards and data pipelines, QA, stress-testing analytical systems and training analysts. Synthetic data is genuinely useful here, because nobody is passing it off as consumer evidence.
Medium risk: useful, with stronger validation
Modelling customer journeys, spotting possible CX pain points, exploring brand perceptions, testing creative concepts before live research, estimating reactions to scenarios and exploring segments. The key word is hypothesis. Synthetic data can help researchers ask what might be happening. It should not quietly turn into a statement about what consumers are saying.
High risk: be much more careful
Brand awareness, NPS, customer satisfaction, market share, campaign effectiveness, consumer sentiment, brand health tracking, population-level claims and major investment decisions. In these areas, what real people do and say counts as evidence. Simulated answers don't count the same way by default.
Brand health shows the line most clearly. Exploring a brand is one job and measuring it is another. Synthetic data can help explore possible brand associations, consumer narratives, campaign reactions, emerging themes and alternative scenarios. Measuring brand health normally needs actual consumers from a defined population, representative or clearly defined sampling, consistent measurement so the numbers trend over time, and a method anyone can inspect.
A simple rule follows. If a brand will put a KPI on a dashboard, compare it across quarters, judge management on it or invest against it, someone should be able to say exactly where the human evidence came from.
Can it replace surveys and primary research?
No, but it can reduce how much primary research is needed. Those are two different claims.
The traditional model runs in a straight line. Researchers run the study, collect consumer responses, analyse them and arrive at insight.
A likely future model starts somewhere else. Existing research, behavioural data and synthetic modelling together generate hypotheses. Those hypotheses point to targeted primary research, and that research validates or rejects them. This second model could well be cheaper and faster, because fieldwork goes only where it is needed.
So the real role of synthetic data is cutting unnecessary primary research, not eliminating it. If the claim is about what Kenyan consumers think, Kenyan consumers should stay part of the evidence base wherever that's feasible. Human evidence for human claims.
How should it be applied, if at all?
Yes, it should be applied, under clear rules. Five practical rules cover most of what a brand or agency needs.
- Declare it. If synthetic data was used, say so. Never let an audience assume generated responses came from real consumers.
- Don't inflate sample sizes. 1,000 interviews plus 9,000 generated observations is 1,000 human respondents plus synthetic augmentation. It is not a 10,000-person survey.
- Validate against reality. Check synthetic outputs against actual survey data, behavioural data, sales, CRM records, social listening, earlier research and experiments.
- Match the method to the decision. The bigger the consequences of a decision, the higher the bar for evidence.
- Preserve provenance. Every major insight should be traceable. Where did the data come from, what was generated, what assumptions went in, and what validation was done?
A simple check, the FITS test, puts the same ideas into four questions.
- F, fitness for purpose. Is synthetic data appropriate for this question?
- I, integrity of provenance. Can anyone explain where the data came from and how it was generated?
- T, transparency. Can the client and the audience tell observed data apart from generated data?
- S, statistical and empirical validation. Has anyone shown the synthetic output behaves enough like the real phenomenon for this purpose?
The bigger the decision, the stronger the answer to each of those four questions needs to be.
Pro or anti synthetic data?
Neither. The position here is pro-evidence.
Synthetic data earns its place where it makes research faster, more efficient or more capable. It needs far more caution when its output is presented as direct evidence of what real consumers think and do. Research exists to reduce uncertainty. Technology can speed that up, but it shouldn't make anyone feel more certain than the evidence allows. The danger isn't synthetic data. The danger is forgetting the difference between a model of reality and reality itself.
The real risk is confusing simulation with observation. If an AI produces the answer a hypothetical Kenyan banking customer might give, no Kenyan banking customer gave it. The answer may be plausible. Plausibility is not evidence.
The biggest issues are transparency and provenance. Anyone reporting brand awareness, satisfaction, NPS or campaign effectiveness owes clients a clear account of how much came from real consumers and how much was generated or modelled.
Used well, synthetic data makes research smarter. It sharpens hypotheses, shows where to investigate and trims fieldwork that was never needed. But when a research claim is about what real people think, feel or do, real human evidence still belongs at the centre of it. Synthetic data can simulate consumers. It cannot, by itself, give consumers a voice.
