Almost every research team uses AI now. Only 8% trust it to be the respondent. Synthetic respondents promise speed and scale, and on the right questions they deliver. On the questions that actually matter, they quietly fail in ways the output never shows.
Almost every research team is using AI now. Almost none of them trust it to be the respondent. A 2026 industry read put the contradiction in plain numbers: 97% of researchers use AI somewhere in their workflow, and only 8% trust AI-generated participants. Synthetic respondents live in that gap, and the gap is where the real argument is.
The pitch is easy to follow. Instead of fielding a study and waiting on human completes, you query an AI model calibrated to behave like your target population. Answers come back in minutes. No recruitment, no incentives, no panel fatigue. For a profession under constant pressure to move faster on smaller budgets, that is a genuinely attractive offer. The question is what you are actually buying.
What the Term Even Means
Part of the confusion is that "synthetic" covers at least five different things. Qualtrics' 2026 trends work counted them: synthetic personas, synthetically derived insights, simulated individual data, digital twins, and simulated conversations. A pure synthetic respondent is an AI persona built from census data, behavioral modeling, and model priors, not tied to any real person. A digital twin is a replica of a specific, known individual. These are not interchangeable, and plenty of vendor pitches blur them on purpose. Before you weigh any accuracy claim, find out which one is actually being sold.
Where They Hold Up
The accuracy data is better than skeptics expect and worse than the marketing suggests. When carefully calibrated against real human data, synthetic respondents hit 85% to 95% alignment on quantitative trends, and around 85% distributional similarity across concept and pricing studies. That is good enough for early-stage work where speed and breadth matter more than precision: pre-testing a questionnaire, stress-testing concepts before you spend real money, getting a directional read on a market you have not studied before.
Then the number falls off a cliff. On complex, multi-factor studies, replication accuracy drops to somewhere between 37% and 60%. The harder and more novel the question, the worse synthetic respondents perform, which is the exact inverse of what you want. The questions that are easy to answer are the ones you least need help with. The questions worth real money are the ones synthetic data is worst at.
The Failure Mode That Doesn't Show Up in the Output
The accuracy gap is the visible problem. The structural one is harder to catch. Synthetic respondents regress to the average. They hand you the center of the distribution and quietly drop the tails, which means they almost never produce the surprising outlier that makes qualitative research worth doing. They also flatter the prompt. Sycophancy bias is well documented: ask a leading question and a synthetic respondent tends to give you the answer the question seems to want.
A 2026 paper from Paglieri and colleagues found that even when models are explicitly told to generate diverse personas, the output collapses toward a narrow cluster of stereotypes. The synthetic "sample" that looks varied on the surface is often a handful of model defaults wearing different demographic labels.
Two more blind spots compound the problem. Models reflect the period of their training data, so they miss attitude shifts that happened since. And anything under roughly 2% of the general population is barely represented, so niche segments come back as best guesses dressed up as data. None of this reads as wrong in the deliverable. The output is clean, plausible, and confident. That is exactly what makes it risky.
A Working Rule for 2026
The teams using synthetic respondents well are not treating it as a binary. They use synthetic for speed and breadth in early exploration, and human data for depth and validation when a real decision is on the line. The 8% trust number is not a verdict on the technology. It is an honest read on where the validation work still needs to happen.
Two things separate the teams getting value from the ones getting burned. First, calibration. Synthetic quality is bounded by the quality and recency of the first-party human data behind it, so teams without their own data assets are building on sand. Second, honesty with stakeholders about what a synthetic read is and is not. A directional pre-test labeled as such is useful. The same output presented as validated consumer truth is how a brand ends up making an expensive decision on a model's best guess.
Synthetic respondents are not the end of human research, and they are not a gimmick. They are a fast, cheap, narrow tool with a trust problem the industry has not solved yet. The question worth sitting with is not whether to use them. It is which of your decisions you would actually stake on an answer no human ever gave. See how Quali-Fi approaches synthetic and human data in the same research program ->