Blood Donors Said They'd Give 3.2 Times a Year. They Actually Gave 41% Less. Now, Imagine Training AI on That Data Simile just raised $100 million to build AI clones of real people for market research. Interview humans, capture their preferences and purchase data, then let companies query those digital twins forever. No survey fatigue. No focus groups. Infinite questions. CVS is already using them. 2.9 million responses from 400,000 people. They claim 95% accuracy in replicating known findings. When independent researchers tested digital twins on questions outside the training data, accuracy dropped to 67%. That's the difference between confirming what you already know and discovering what you don't. Worth noting: CVS Health Ventures is an investor in Simile. CVS is also the flagship customer. Gallup is a launch partner. Every voice in the story has a financial interest in the technology working. I couldn't find a single independent, adversarial evaluation of the accuracy claims. 80% of new products fail within six months of launch. Most passed focus groups first. Why? Because what people say they want and what they actually do are two different things. Psychologists call it the say/do gap. Blood donors told researchers they'd give 3.2 times per year. Their actual rate was 41% lower. People claim they prefer eco-friendly products, but then grab the cheaper option at checkout. We don't fully understand our own decision-making. We answer surveys as the person we want to be, not the one who shows up on Tuesday when we're tired and the kids are screaming. So when you train an AI twin on that same self-reported data, you digitize the lie. Faster. Cheaper. Still wrong. I'll be fair. Traditional research has real problems that twins solve. Focus groups breed groupthink. One loud personality hijacks the room. Response rates for many major surveys have cratered below 10%. Removing the group dynamic, eliminating fatigue, and allowing unlimited depth of questioning. These are genuine improvements. But they trade visible problems for invisible ones. A focus group, everyone knows it's messy. A dashboard that says "95% accurate" with clean charts and instant answers? That feels like certainty. Even when it isn't. A twin is trained on a fixed window of data. Job loss, a health scare, a new relationship, inflation. Any of these rewrites someone's decision framework overnight. What's the shelf life of a twin? Nobody says. The honest framing: this technology generates faster, cheaper hypotheses. Not answers. Hypotheses that still need testing against what real humans actually do. CVS's own VP said it: "We're never going to stop talking to real customers." The question is whether every company buying this will show the same restraint. Or whether the speed and savings will quietly replace the messy work of watching what people actually do. https://lnkd.in/e7bSDU5x