Synthetic panels go mainstream, humans still validate

The gist
Synthetic research has finally gone mainstream, but human data and rigorous validation are still the gold standard.
What to know
- Simile and Kantar have moved from pilots to operational synthetic panels, with Simile simulating 1,000 Americans at 85% accuracy and Kantar scaling workflows to 50 clients across 20 markets.
- Enterprises like NTT Data and Araya now blend synthetic populations with a 50,000-person human database to test products and services before launch.
- Experts warn synthetic outputs must be carefully calibrated—recent studies show AI-generated results can look convincing, but real human validation remains essential for decision-making.
Digital Twins Hit Critical Mass
A governance shift and sudden demand for digital twins fueled a flood of new platforms, transforming synthetic research from niche pilots to mainstream enterprise tools almost overnight.
The late-2026 surge did not emerge from nowhere: on March 26, 2018, the U.S. Census Bureau announced it would use differential privacy for the 2020 Census after research showed attackers could reconstruct confidential information from 2010 Census data products, a governance precedent that later synthetic-population vendors would have to meet. By June 2026, Greenbook captured the market turn in plain language, contrasting a time when “nobody was running around asking for digital twins” with a new moment when “everyone wants to at least try digital twins,” and “tons of platforms” had started up in the prior months.
What changed between August and November 2026 was that scale and operating discipline began to appear together: Training Data reported that Simile “simulated 1,000 Americans” and “predicted their behavior 85% as accurately as people reproduce their own answers,” while an August Future of Insights Summit presentation signaled the topic had become mainstream enough for the conference circuit. At the enterprise end, Cynthia Vega said Kantar was having “substantive conversations with marketing leaders at FTSE 100 advertisers” and contrasted “a bespoke pilot” with “a skill that runs the same workflow for 50 clients across 20 markets,” built on “proprietary panel data from real, consenting individuals… across 55 markets” and certified methodology.
Human Data Remains the Backbone
Enterprises are embedding synthetic panels into workflows, but real human databases and expert oversight are essential for credible results and high-stakes decisions.
Enterprises are wiring synthetic consumers into research workflows with explicit human anchors. The User Research Strategist’s guidance is to write 10 questions with clear scenarios and distinct options, run synthetic studies, and use the output for pattern flags and complexity signals, while keeping human interpretation central; synthetic systems are being used to accelerate prep work and focus live research, not to replace direct human evidence.
That integration is visible in enterprise deployments. On August 7, NTT Data Institute of Management Consulting and Araya said they would start August 2026 research on AI panels that combine NTT Data Institute of Management Consulting's Human Information Database with AI agents, using a database that covers about 50,000 people to enable checks on products, services, and fraud schemes before they reach real people. Bessemer Venture Partners described the broader commercial pattern, from a CMO at a Fortune 500 brand running three packaging concepts against 10,000 consumers overnight to a political consultant pressure-testing a Senate campaign's messaging, and said that by 2025 real enterprise revenue was flowing. It also pointed to larger-scale use cases, including a NATO analyst modeling 5.1 million people, while noting that some leading enterprises already spend seven figures annually.
Calibration, Not Perfection, Wins
Synthetic samples can mirror real trends when carefully benchmarked, but their value lies in directional insights—not flawless replication of human outcomes.
The strongest synthetic research setups work as layered evidence engines: AI-generated populations are conditioned on human inputs, then checked against outside reality before anyone treats the output as useful. Mark Ritson’s point in The Drum captures the limit: Anthropic could run 10,000 silicon Ritsons for the price of a coffee, but it cannot tell Anthropic what he actually felt until he says so, a distinction echoed at Esomar Congress 2026 when Wilko Rozema of Dynata framed synthetic data as enhancing research without replacing people.
That approach depends on calibration and benchmarking, not blind acceptance. In 2022 Lisa Argyle and colleagues at Brigham Young released the silicon samples paper, showing that GPT-3, conditioned on real demographic profiles, could emulate response distributions across subgroups, and in 2023 James Brand, Ayelet Israeli and Donald Ngwe showed that GPT-3.5 could generate survey responses consistent with established patterns of consumer behavior and produce realistic outputs, but only as bounded approximations. Intellicast says a human survey finding of 42% liking a menu item will not be exactly 42% in a model community, though similar kinds of answers can emerge, and Greenbook argues synthetic research does not need to be 100% as accurate as private research if it offers directional guidance instead of guesswork.
Validation Gaps Expose Risks
Confidence in synthetic outputs often outpaces rigorous validation, with unchecked models amplifying stereotypes and missing critical details unless carefully governed.
Validation is becoming mandatory because synthetic research can sound precise long before it proves it is stable. WARC reported an interviewee citing an ARF survey showing “92% confidence in the output” among marketers, but only “30 something” when it came to validation, a gap that captures the core risk: plausible, professional answers are being trusted faster than they are being tested, even though small changes in prompts, question order, settings, model versions, or later model updates can produce sharply different results.
Governance is also becoming essential because persona-based systems can reproduce broad stereotypes while missing the details that matter for decisions. Greenbook’s example of asking a model to adopt “a 55year-old… geeky dude that lives in Kentucky” came with the warning that it “may get big outlines it’s not going to get details,” echoing broader concerns that models inherit training-data blind spots, can oversimplify underrepresented groups, and can turn assumptions baked into a brief into outputs that look like evidence unless they are checked against reality.
Skeptics Demand Real Evidence
Synthetic respondents excel at rapid screening but consistently overestimate effects and miss nuance, reinforcing the need for human data in decisive research.
Sceptics draw a hard line between plausible prediction and actual measurement. The Sen Times noted that Ashwini Ashokkumar’s study found GPT-4 forecasts were “strong[ly]” correlated with real results, but “systematically estimated the effects were around twice as big as the real results,” meaning a model may rank options while still failing to show whether an effect is tiny or transformative; that is why synthetic respondents are framed as useful for early exploration, not as substitutes when the goal is to understand real human behaviour or make high-stakes calls.
Empirical tests sharpen that objection. In a study that tried to reproduce a politically representative online sample of 996 US participants plus AI-moderated qualitative interviews with LLM-simulated respondents, Andrew Gordon found that “individual-level simulation with demographic personas… tripled distributional error compared with asking the model for an aggregate distribution”; even changing prompting to “eliciting a probability distribution per persona rather than one forced answer cut error by 35-48% while keeping individual-level data,” but “post-stratification (MRP) couldn't touch what remained,” and “no method matched the accuracy of a second human sample.”
That leaves synthetic consumers strongest as a screening tool, not a final arbiter. Intellicast contrasted human studies, where “you may have to take 10 concepts,” with modeled respondents, where “you can test 50 concepts, test 100 concepts,” but pricing research shows why sceptics still insist on real evidence: the updated 2026 Harvard Business School study by Brand, Israeli and Ngwe found “two- to threefold overestimation” and “wrong-signed preferences,” while a built-in laptop projector ranged from roughly US$1,074 under GPT-3.5 to about US$333 under GPT-4o, and fine-tuning gains “did not generalise well.”








