The AI research tools market is converging on a shared promise: faster, cheaper, always-available audience insight without the logistics of recruiting and running studies with real people. Synthetic users and AI personas both make versions of this promise. They are not the same thing.
What synthetic users are
Synthetic users are AI agents that simulate research participants. The underlying idea is that a large language model, when prompted with a persona description (age, gender, occupation, attitudes) can generate responses that plausibly represent how a real person matching that description might answer research questions.
Some tools add sophistication to this: training on public data about specific demographics, using structured interview formats, generating multiple synthetic participants to simulate a panel. A handful of academic studies have investigated whether LLM-generated responses can approximate the patterns you'd find in real survey data, with mixed but occasionally encouraging results in specific conditions.
The core mechanism remains the same: the model simulates a type of person, based on its training data. Its knowledge of how that type of person thinks and behaves comes from publicly available text: what has been written about people like that, not from research conducted with them.
What AI personas grounded in research are
Research-grounded AI personas work from the opposite direction. Rather than simulating a generic type of person, they are built from specific evidence: the qualitative interviews, quantitative surveys, focus groups, ethnographic observations and other primary research that your organisation has conducted with your actual audience.
The mechanism is retrieval-augmented generation (RAG). When a question is asked, the system retrieves the most relevant passages from the underlying research corpus and uses them to generate a grounded response: one that reflects what your research actually found, not what a model trained on public text predicts that "this type of person" would say. Every response shows its sources and carries a confidence score indicating how well-supported it is by the retrieved evidence.
The knowledge in a research-grounded persona comes from real participants who were asked real questions in a specific context. The AI's role is to make that knowledge accessible and queryable, not to substitute for it.
The key differences
Where synthetic users have genuine value
Synthetic users can be genuinely useful in specific, bounded contexts. Early-stage product exploration (when you have no existing research and no time or budget to conduct primary research) can benefit from the kind of rapid hypothesis generation that synthetic participant responses enable. If you need to quickly generate a range of plausible responses to a concept as a stimulus for internal discussion, synthetic users can provide that without the overhead of a real study.
Some research in computational social science has found that LLM-based agents can approximate group-level patterns in attitude data for mainstream demographics, though with important caveats about minority groups, niche audiences and contexts where the training data doesn't provide adequate coverage. These findings have been replicated unevenly, and the conditions under which synthetic responses are reliable remain an active area of methodological investigation.
The appropriate framing for synthetic users is as exploratory tools for ideation and hypothesis generation, not as validated substitutes for research conducted with real people in specific contexts.
Where they fall short, and why it matters
Synthetic users have two limitations that are fundamental rather than incidental. First, their knowledge comes from public training data, which means they systematically underrepresent audiences that are not well-documented in publicly available text. Niche professional communities, non-English-speaking populations, specific regional cultures, specialist domains: any context where the internet does not provide extensive, well-curated coverage will produce synthetic responses that are poorly calibrated at best and misleading at worst.
Second, and more importantly for professional contexts: synthetic user outputs are not traceable. When a team needs to defend a decision to a client, a board, a regulator or a funder, "the AI thought this audience would respond well" is not an accountable answer. There is no underlying research to point to, no specific data points to cite, no methodology that can be scrutinised.
This is not a minor limitation. In creative agencies, engineering firms, public bodies and research-intensive organisations, accountability is not optional. Decisions need to be defensible. That requires a traceable chain from output to evidence, something synthetic users cannot provide by design.
When to use which
The clearest heuristic is whether you have existing research and whether the decisions you're making need to be defensible. If you have no research and the stakes are low, synthetic users offer fast generative stimulus. If you have proprietary research and need to use it more effectively throughout a project, research-grounded AI personas are the appropriate tool.
These tools are not really competing for the same use case. Synthetic users are a substitute for research you don't have. Research-grounded AI personas are a way to get more value from research you've already invested in. For organisations that have conducted genuine primary research (which is most organisations making serious decisions about audiences), the choice is usually clear.
The nuance is that some organisations use both: synthetic users for exploratory hypothesis generation at the start of a project, research-grounded personas for validated decision support once the evidence base exists. That combination is coherent, as long as the different epistemic statuses of the two tools are clearly understood by the team using them.
A note on independent validation
Persona Dynamics conducted an independent evaluation of our research-grounded Dynamic Personas™ against human domain experts in February 2026: 50 real-world question–answer pairs, assessed by 9 independent subject matter experts under blind conditions. The results showed 92.4% accuracy on core claims versus 88.5% for the human benchmark, and 60.7% of AI responses rated as producing specific, actionable recommendations versus 19.1% for human experts.
We are not aware of equivalent independent, blind evaluations of synthetic user tools against human benchmarks in professional decision-making contexts. This is not necessarily because such evaluations would be unfavourable. It is because the evaluation frameworks for AI research tools are still emerging. We think rigorous, independent evaluation matters, which is why we publish our methodology and findings openly. The evidence base for any AI research tool should be assessed on the same terms.