Synthetic respondents in pharma research

synthetic respondents

Six questions before they shape a decision

Synthetic respondents have moved from conference talk into pitch decks. These are AI-generated personas, built on large language models, that answer surveys and interview questions as a defined audience might. Several platforms now offer synthetic HCP and patient panels, and some in the field expect life sciences to be among the first sectors to adopt them (Forbes, October 2026).

The case for them rests on speed and cost. Recruiting HCPs for advisory boards and studies can cost hundreds of thousands of dollars and take 12 to 16 weeks, according to the same Forbes piece, written by a founder whose company builds synthetic personas for pharma.

Much of the coverage stops there. For insight teams, the more useful question is what a synthetic respondent can tell you about how people actually behave. Here are six areas to weigh before one informs a decision.

1. What has the model learned from?

Synthetic respondents learn from what people have written and said: published surveys, interview transcripts, online discussion and clinical literature. That is stated preference, the layer of evidence behavioural science treats with most caution.

HCPs describe prescribing as evidence-led, yet their choices also reflect habit, workload and the last patient who did badly. Patients report adherence they don’t maintain. A model trained on what people say inherits the gap between saying and doing, then reproduces it with more confidence and at greater scale.

2. Is there enough data on this audience?

Some providers position synthetic panels for hard-to-reach groups, including physicians and rare disease patients (Simsurveys, April 2026). Those are the audiences where a model has least to learn from.

When the global expert base for a condition is a few dozen specialists, little published opinion exists to train on, and much of it may come from the same handful of voices. A synthetic panel can produce hundreds of answers, but it cannot add knowledge that was never in its training data. These are also the audiences where a wrong assumption about the diagnostic pathway or treatment threshold costs most.

3. What was it validated against?

A typical validation prompts a model with a demographic description and compares its answers on attitude and opinion items with a human sample. A 2026 framework paper on validating synthetic survey populations describes this as the usual approach, and argues every synthetic study should document how its personas were built and state the level of validity it claims (arXiv, 2026).

Agreement on attitudes says little about whether a model can predict what a cardiologist will prescribe on Monday, or whether a patient will refill after the first side effect. Before trusting a synthetic panel, ask what it was validated against: what people said, or what they did.

4. Has it kept the segments that matter?

There is a sensible role for them. Synthetic panels can pressure-test a questionnaire before fieldwork, flag confusing stimulus, generate hypotheses to take into real interviews, and help decide which questions deserve scarce HCP time.

The deciding factor is how much weight the output will carry. If it will shape a TPP, a forecast, a positioning or a launch sequence, it needs real respondents and methods designed to get past stated preference, such as observation, choice tasks and implicit measures.

5. How much weight will the output carry?

There is a sensible role for them. Synthetic panels can pressure-test a questionnaire before fieldwork, flag confusing stimulus, generate hypotheses to take into real interviews, and help decide which questions deserve scarce HCP time.

The deciding factor is how much weight the output will carry. If it will shape a TPP, a forecast, a positioning or a launch sequence, it needs real respondents and methods designed to get past stated preference, such as observation, choice tasks and implicit measures.

6. Are your real panels still real?

Synthetic respondents are the visible use of language models in research. A less visible one is AI-generated answers arriving in real online panels. Researchers have begun publishing scales to detect LLM-generated responses in online surveys (2026), and the BHBIA Fieldwork Committee has issued guidance on preventing fraudulent respondents in healthcare market research.

Checks built to catch speeders and straight-liners may miss fluent, plausible open-ends. It is worth asking fieldwork partners how they confirm that an HCP respondent is a real, practising clinician, and how they screen open-ended answers.

What to do now

The common thread is the distance between what people say and what they do. Synthetic respondents are quick and inexpensive, and they reflect stated views far better than behaviour. That makes them useful for preparing research and a weak basis for decisions on their own.

Before commissioning synthetic work, put three questions to the provider: what data built the personas, what they were validated against, and how variation is preserved. Involve compliance and your DPO early too. Some providers present synthetic research as largely outside data protection rules because no personal data is processed during a session (getminds.ai, 2026), but the data the model learned from still has a source.

Porterhouse Insights helps pharmaceutical teams understand how HCPs, payers and patients actually make decisions, through primary market research and behavioural science.

If synthetic respondents are on your 2027 research plans, let’s talk through where they fit. Contact us at info@porterhouseinsights.com