Theory vs
synthetic research
Synthetic research uses AI models to generate answers as if they came from customers, experts or patients. Verified human research from Theory Intelligence interviews real people whose identity and role have been checked. The difference matters most where decisions depend on the tails of a distribution, on new phenomena the model has not seen, or on someone being accountable for the answer.
Theory vs Synthetic Research
| Synthetic respondents | Theory Intelligence | |
|---|---|---|
| Source | A language model prompted to answer as a persona | A verified person with direct, recent experience |
| Cost and speed | Near zero; instant | Signal Brief USD 7,500 to 12,000; 10 business days |
| Average answers | Reasonable on well-documented topics | Reasonable, plus the reasons behind them |
| Tails and outliers | Compressed toward the mean; vendors' own research shows reduced variance | Present; the tails are where most decisions are made |
| New phenomena | Absent; the model knows what was published before its cutoff | Present; people know what happened last month |
| Verification | Not applicable; there is no one to verify | Identity and role verified; method recorded |
| Accountability | None | Named researcher signs the recommendation |
| Best for | Drafting questions, pre-testing surveys, cheap early exploration | Decisions where being wrong is expensive |
What synthetic respondents are
A synthetic respondent is a language model prompted to answer as if it were a particular kind of person: a procurement lead at a mid-size manufacturer, a 45-year-old insurance buyer, an oncologist. Vendors offer thousands of them for the cost of an API call. For drafting a questionnaire, pre-testing a survey or getting a quick sense of the obvious objections, they are useful and cheap.
What the evidence shows
The most careful public evidence comes from the vendors themselves. Research presented at ESOMAR by Fairgen and Google researchers found that synthetic respondents reproduce average answers reasonably well on well-documented topics but show reduced variance: they cluster around the mean and under-represent the tails. Academic work on LLM survey respondents has found the same, along with a tendency to reflect the model's training distribution rather than the target population. Synthetic respondents also cannot know anything that happened after the model's training cutoff or that was never written down. See the full article for sources.
Why that matters for decisions
Business decisions rarely turn on the average customer. They turn on the segment that will churn, the buyer who will pay double, the clinician who will refuse to adopt, the broker who has heard something about your competitor's claims service. Those are tails and recent events. A tool that compresses the tails and stops at its training cutoff will give a confident, plausible and wrong answer exactly where the decision lives.
What Theory does
Theory interviews real people. Every one is identity and role verified. Quotations are verbatim. The recommendation is signed. We use AI heavily for logistics, transcription and drafting, and we say so, but we never use it as a respondent. Our Signal Brief is the fastest way to test the difference: six to eight verified humans on one question, ten business days.
Everything Theory delivers is governed by The Theory Standard. Prices are on the services page.
Questions people ask
Are synthetic respondents accurate?
For average answers on well-documented topics, reasonably. For tails, outliers and anything recent or unpublished, the evidence, including vendors' own, shows they are not.
Does Theory use synthetic respondents?
No. We use AI for logistics, transcription, coding and drafting, and disclose it. We never use it in place of a human participant.
When are synthetic respondents a good idea?
Drafting and pre-testing questions, cheap early exploration, and generating hypotheses to test with real people.
Test the difference on one question
A Signal Brief: six to eight verified interviews, a signed answer, ten business days, USD 7,500 to 12,000. info@theoryintel.com.
Start a brief