Simile emerged from stealth on February 12 with a simple pitch: interview hundreds of real humans about their lives, feed the transcripts into an AI, and sell the resulting digital twins to corporations. Index Ventures led the $100 million Series A. Fei-Fei Li and Andrej Karpathy wrote personal checks.
The company's CEO is Joon Sung Park, whose 2023 Stanford paper put 25 AI agents in a simulated town called Smallville and watched them form relationships, throw parties, and develop social hierarchies without being told to. That paper won Best Paper at ACM UIST. Simile is the commercial version: instead of fictional characters, the agents are trained on real people.
Here's how it works. Simile conducts qualitative interviews with hundreds of individuals about their lives — their preferences, their reasoning patterns, the way they explain their own decisions. That data gets combined with transaction histories and behavioral science literature. The result is a population of AI agents that mirror the preferences and tendencies of actual people. Corporations drop these synthetic populations into hypothetical scenarios and watch what happens.
CVS Health is already using it to decide which products to stock and where to place them. Telstra, Australia's largest mobile carrier, is testing it for customer experience decisions. In a Bloomberg Television interview, Park said the model correctly predicted eight out of ten analyst questions before an actual earnings call.
This is not market research. Market research asks people what they think. Simile builds copies of people who think for them.
The accuracy problem nobody's talking about
The technology works better than you'd expect — and worse than anyone wants to admit.
An NN/g meta-analysis of three digital twin studies found interview-based twins hit 85% accuracy on survey questions and 80% on personality assessments. Population-level correlations looked near-perfect. But at the individual level, economic decision-making accuracy dropped to 66%. The synthetic populations produced less variable responses than real humans, missing edge cases and polarized opinions entirely.
The racial bias findings should stop everyone in their tracks. One study found digital twins predicted white people's responses more accurately than other racial groups. Interview-enriched models reduced that gap by 7 to 38 percent — an improvement, not a fix. If CVS is using these models to decide what goes on shelves in different neighborhoods, the prediction gap becomes an inventory gap becomes an access gap.
You can't un-clone yourself
The legal framework for digital twins doesn't exist.
Harvard's Petrie-Flom Center published a paper in October 2025 titled "Predictive Persons" that laid out the problem: U.S. law grants individuals no property interest in their own behavioral data. Once you participate in a Simile interview, your decision patterns are embedded in a trained model. You can't withdraw consent after the fact. Your behavioral signature lives inside CVS's inventory system indefinitely.
HIPAA doesn't apply — Simile isn't a healthcare provider. The ACA and GINA don't cover life, disability, or long-term-care insurance, which means risk scores derived from behavioral twins could lawfully inform those decisions. Fenwick and Jurcys, in a TechPolicy Press analysis, argued that unauthorized behavioral duplication should be classified as identity theft. Current law disagrees.
The contrast with GDPR is stark. European data subjects have the right to be forgotten. Americans have the right to be simulated without knowing it.
The real product is you, minus your inconvenience
Simile's pitch deck doesn't mention ethics. Neither does Bloomberg's coverage, PYMNTS's writeup, or SiliconAngle's breakdown. The word "consent" appears nowhere in the company's public communications about how interview subjects' data gets used downstream.
This is the logical endpoint of a trend that started with cookies and accelerated through recommendation algorithms: corporations don't want to understand customers. They want to skip the customer entirely. A digital twin doesn't complain. It doesn't change its mind at the register. It doesn't sue when the algorithm decides its neighborhood doesn't need fresh produce.
Park's Smallville agents threw parties and formed friendships. Simile's agents tell CVS what to put on aisle seven. The academic paper won awards for showing AI could simulate human warmth. The commercial product sells that warmth to the highest bidder.
Fei-Fei Li invested. Andrej Karpathy invested. These are not people who lack ethical frameworks. They backed Simile anyway. That tells you exactly how much money there is in building a world where corporations can test-drive decisions on fake people before inflicting them on real ones.
The $100 million isn't for predicting behavior. It's for making behavior predictable. The difference matters more than anyone at Index Ventures seems willing to discuss.
Originally published on Substack. Follow for daily AI analysis.
Top comments (0)