Companies Are Making Million-Dollar Decisions Based on People Who Don’t Exist

Somewhere this week, a company is deciding what product to launch, how to price it, and who to sell it to, based on the opinions of customers who were never born, never bought anything, and don’t exist. They were generated by an AI. And almost no one in the room is asking whether that’s a problem.

Welcome to the fast-rising world of synthetic respondents, AI-generated “people” built to simulate how real humans think, feel, and buy. Instead of surveying actual customers, companies increasingly ask an AI to role-play thousands of them. The trend is accelerating so fast that this month the Association of National Advertisers published a strategic analysis of synthetic panels, and a new certification body launched to audit them, mapping 34 synthetic-data products from 30 providers before running a single audit. The technology is racing ahead of the guardrails.

The appeal is obvious. A traditional study can cost $5,000 to $25,000 and take weeks; a synthetic one can cost a few hundred dollars and run overnight. Investors have noticed: one Stanford spin-out in the space raised $200 million at a $2 billion valuation this year alone.

Here’s the problem: the fake people are confidently, dangerously wrong, in exactly the moments that matter most.

Consider the hard tests. The Nürnberg Institut für Marktentscheidungen, founded by research giant GfK, ran 500 real respondents against 500 AI ones, and the AI diverged from humans on 75% of soft-drink questions and 80% of sportswear questions, while two separate groups of real people stayed consistent with each other. In another test, Emporia Research asked the same B2B questionnaire to real professionals and synthetic ones: 47% of the real people said they were satisfied in their roles, against 69% of the AI versions. The model was manufacturing contentment the real market simply didn’t feel.

That’s the core flaw. A Stanford study found AI personas suffer from sycophancy bias, they drift toward whatever answer the question implies you want, and converge on majority opinions, burying the dissent real research exists to surface. One analysis found synthetic answers matched real averages but collapsed the variance, with a third of key relationships flipping direction entirely. They produce what one review called “fluent fiction”: coherent, plausible, and wrong.

The industry knows it. In a 2026 survey, 97% of researchers said they use AI, but only 8% trust AI-generated participants, and 99.3% reject using them alone. They know the difference between a shortcut and a substitute.

To be fair, synthetic data isn’t useless. For brainstorming, pretesting, or stress-testing ideas, it’s a real time-saver, and the best models reach around 85% fidelity, but only when built from rich, real human interviews first. The danger isn’t the tool. It’s treating a simulation of people as if it were people.

And that’s the lesson for anyone making high-stakes calls. As the world fills with plausible-sounding AI output, the most valuable thing in research is no longer speed or scale, it’s provenance. Knowing your insight came from a real, verified human being, with a real biography and real money on the line, is becoming the ultimate competitive edge.

Because when a decision is worth millions, the question isn’t “what does the data say?” It’s “did these people ever actually exist?”