Simile Raised $100M to Simulate Human Decisions — Should We Be Excited or Terrified?

What if a company could predict exactly how you would react to a new product, a price change, or a political message — before you ever saw it? That is precisely what Simile, a Stanford spinout founded by the researchers behind the viral Smallville experiment, claims to have built. And in February 2026, Index Ventures led a $100 million Series A to scale it.

The investor list reads like an AI hall of fame: Fei-Fei Li, Andrej Karpathy, Bain Capital Ventures, A*, and Hanabi Capital. The customers already include CVS Health, Telstra, Wealthfront, Gallup, and Suntory. This is not a research project anymore. This is a company betting that the future of decision-making is simulation.

From Smallville to the Real World

In 2023, Joon Sung Park, Michael Bernstein, and Percy Liang — all at Stanford — published a paper that captured the internet's imagination. They built Smallville, a tiny virtual town populated by 25 AI agents that cooked breakfast, went to work, formed relationships, and even organized a Valentine's Day party — all without explicit programming.

The agents were not following scripts. They were generative agents: software entities that combined large language models with memory, reflection, and planning to produce believable human behavior. The paper went viral, accumulated thousands of citations, and planted a question: what if you could do this with real people?

Simile is the answer. The company has built what it calls the first AI simulation of society populated by agents based on real humans. Feed it data about a person — interview transcripts, transaction logs, behavioral patterns — and it generates a digital twin that can be placed in simulated scenarios to predict how the real person would respond.

How It Works: 85% and Climbing

The model was trained on interviews with hundreds of people, combined with transaction logs and text from scientific journals. Development took seven months. By 2024, Simile reported achieving greater than 85% accuracy in simulating human responses across a range of scenarios.

Consider the numbers in practice:

• CVS Health used Simile to determine which products to stock on shelves and predict how inventory would sell — reportedly skipping months of traditional market testing
• Earnings call rehearsal: executives simulate analyst questions before going live, with the system achieving 80% accuracy on forecasting the actual questions asked
• Litigation modeling: law firms test arguments against simulated judges and juries before entering the courtroom
• Policy testing: organizations preview how populations would react to policy changes before implementation

The pitch is compelling: instead of deploying untested decisions to the real world — launching a product, changing a price, restructuring a service — you simulate first. Fail in the simulation, not in production.

The Uncomfortable Questions

And here is where it gets complicated. Because the same technology that helps CVS stock shelves more efficiently also raises questions that cut to the heart of autonomy, consent, and power.

1. Consent and Digital Twins

When Simile creates a digital twin of a person based on their interview data and transaction history, who owns that simulation? The person being modeled did not consent to having a digital version of themselves tested against thousands of hypothetical scenarios. The twin may make "decisions" the real person would find objectionable. What rights does a person have over their own simulation?

2. The Manipulation Gradient

There is a meaningful difference between predicting what someone will do and engineering a specific outcome. If you know with 85% accuracy how a person will react to a message, you can craft messages specifically designed to exploit that prediction. Market research becomes precision persuasion. A/B testing becomes A/B manipulation. The technology is agnostic — it does not care whether you are optimizing for customer satisfaction or addiction.

3. Accuracy Limits and Overconfidence

85% accuracy sounds impressive until you consider the stakes. In a population of one million customers, 15% error means 150,000 people whose behavior was predicted incorrectly. If a pharmaceutical company uses simulation to predict patient responses to a drug pricing change, that 15% error is not a rounding error — it is tens of thousands of people making healthcare decisions that the model got wrong. The danger is not that the model fails. The danger is that organizations trust it too much.

4. The Simulation Paradox

Here is the deepest question: does simulating human behavior change human behavior? If a company simulates how its customers will react to a price increase, then adjusts the price based on that simulation, the customers are now responding to a reality shaped by their own predicted behavior. The simulation becomes prescriptive, not just predictive. You are no longer modeling the world — you are designing it.

5. Who Gets Simulated?

Simile's model requires data: interviews, transaction logs, behavioral patterns. This means the people who get simulated are disproportionately those with the most digital footprint — consumers in wealthy economies, users of major platforms, participants in formal financial systems. The billions of people outside those systems are invisible to the simulation. Decisions "optimized" by Simile may be optimized for a narrow, non-representative population, systematically excluding the people most vulnerable to the consequences.

What This Means for AI Strategy

For organizations considering behavioral simulation, several principles should guide adoption:

• Simulation is a lens, not a crystal ball. Use it to surface possibilities and stress-test assumptions, not to replace human judgment
• Establish consent frameworks. If you are creating digital twins of real people, build explicit consent mechanisms and give individuals the right to opt out and delete their simulated selves
• Publish accuracy in context. 85% accuracy in predicting product preferences is very different from 85% accuracy in predicting healthcare decisions. Calibrate trust to the domain
• Separate prediction from optimization. There should be a clear organizational boundary between understanding what people will do and engineering what you want them to do

The Bottom Line

Simile represents something genuinely new — not another chatbot or image generator, but an attempt to computationally model the full complexity of human decision-making. The $100 million bet by Index Ventures and the endorsement of researchers like Fei-Fei Li and Andrej Karpathy signal that the AI industry believes this is a frontier worth exploring.

But "worth exploring" and "ready to deploy at scale" are different things. The technology works. The question is whether our ethical frameworks, regulatory structures, and organizational governance can keep pace with a tool that promises to know us better than we know ourselves.

The simulation is running. The question is who is watching it — and who is being watched.