Joon Sung Park simulates shoppers, but citizens resist code
By Daniel Okafor · Reporting from New York ·
Joon Sung Park built digital twins to predict human choices for retailers, but civic leaders who substitute synthetic consensus for real voters erode democratic accountability.
A machine can now predict human choices, provided those choices remain inside a computer. On the podcast "Sourcery," host Molly O'Shea interviewed Joon Sung Park, the co-founder of Simile. Park argued that his foundation models recreate human behavior by building digital twins of living people. To build them, his firm partners with commercial panels, including Gallup, reaching people across the globe. Clients buy access instantly. Retailers like CVS and overseas lenders like Banco in Brazil query these synthetic crowds to test store decor, corporate statements, and product pricing.
The digital town of Smallville
Park built his reputation at Stanford University in 2023 with an experiment called Smallville. He placed agents into a simulated village, gave them personal histories, and watched them organize a Valentine's Day party through automated planning and memory. That game town led to a study of Americans. Voice agents interviewed each participant about childhoods and hard choices. Two weeks later, researchers asked those individuals to complete behavioral tests while synthetic twins answered the same questions. Park said the digital counterparts matched the humans.
Surveys miss the friction of real life
The ablest defense of synthetic testing points to speed and cost. Corporate teams spend millions of dollars and several years to run traditional analyses. Simile delivers answers in seconds. Park cited an automotive client that tested marketing for an electric vehicle; the simulation revealed that the proposed campaign depressed car sales across the entire market, sparing the company an expensive blunder. Simile filters panel fatigue through arithmetic attention checks and trains a confidence model that scores each prediction. When confidence clears, corporate executives treat the output as solid evidence.
Yet human behavior resists statistical functions. Park himself acknowledged that what people tell an interviewer follows a normal curve, whereas actual execution requires steep activation energy along an exponential drop. A citizen tells a surveyor that he exercises weekly, but stays in bed because sleep carries lower opportunity cost. That gap between intent and action widens when stakes rise. Thomas Schelling mapped segregation decades ago through mathematical models, but real neighborhood changes still depended on physical moves and local laws. Simulating human choices on a screen records what people think they might do when nothing costs them money, sweat, or reputation.
Can code replace a town hall?
This software now aims beyond retail shelves. Park suggested that simulations could test universal basic income programs, climate policies, and retention strategies. Here the technology touches the voting booth. If school boards, city councils, and legislative staffs consult synthetic panels instead of meeting taxpayers, democratic accountability breaks down. A polling model trained on interview transcripts cannot replicate the fury of a homeowner facing a tax hike or the refusal of a voter who changes his mind at the ballot box. Pollsters have promised shortcuts since George Gallup began sampling voters. The record shows that voters surprise models precisely because human life contains crises, social pressure, and private hesitations that no query anticipates.
Park built an impressive commercial instrument for retail marketers who want quick feedback on shampoo bottles and pharmacy layouts. But civic leaders who govern real communities must never substitute synthetic consensus for actual human dissent. When a policy touches livelihoods and liberties, no digital twin can replace the citizen who stands before a microphone to speak for himself.