Key Takeaways

  • Frontier models like GPT-4 drop as low as 20% to 30% accuracy when simulating real human behavior in niche sub-populations because they are optimized for hyper-rational reasoning.
  • Simile AI achieved 85% fidelity in a 1,000-person digital twin study, matching the rate at which human beings replicate their own past choices and attitudes.
  • Standard LLMs assume maximum efficiency: when asked how to get home, a model suggests an Uber or a bus schedule, missing the person who walks home to think.
  • Building accurate human simulators requires conditioning on deep personal interviews, observational transaction logs, and randomized controlled trials instead of standard coding benchmarks.

The Flaw of the Hyper-Rational Engine

Ask GPT-4 how to get home from work. It will calculate the fastest route, check transit schedules, or tell you to hail an Uber. It solves for speed and cost.

Joon Sung Park, co-founder of Simile AI and lead author of the landmark Stanford Generative Agents paper, did something else for years. He walked home. Walking was slower and less efficient, but it gave him time to think. A standard model fails to predict that choice because it treats human beings as logic calculators.

Frontier models are trained on data from professional programmers, mathematicians, and scientists. They are reinforced to produce objective, optimized answers. As Park puts it:

“What these models are really, really good at today is they are trying to basically become the super rational objective machines... Simile actually does not care about any of this. The models that we are talking about here, what we are trying to create are models that are as dumb as I am. If I make those mistakes, the model has to make the same kind of mistake.”

When founders use frontier models to simulate customers, focus groups, or churn patterns, the simulation breaks down. Real people do not read every line of a terms of service page. They forget passwords, buy products based on nostalgia, and hold irrational grudges.

The 85% Replication Benchmark

In a study simulating 1,000 individual digital twins, Park tested how well custom behavioral models could mimic real choices versus unadapted frontier models. The results revealed a wide gap.

Simile reached 85% fidelity when benchmarked against human self-replication. That means the digital twin agreed with the real person's choices at roughly the same rate the person agreed with their own past decisions. Meanwhile, default frontier models collapsed to between 20% and 30% accuracy across niche sub-populations.

“This is where we basically could replicate people's behaviors and attitudes 85% as accurately as people would replicate their own,” Park noted.

Frontier models fail at behavioral prediction because they confuse the correct answer with the human answer. Replicating social physics requires feeding agents qualitative interviews, real-world spending records, and randomized behavioral experiments. Without that context, an agent defaults to academic rationality.

What to Do With This

Stop asking default ChatGPT prompts to act as your target customer persona for product validation or pricing feedback. Pull the raw transcripts and purchase logs from your last ten churned users this week. Feed those specific behavioral errors, complaints, and trade-offs into your test prompts so your synthetic testing mirrors actual user confusion instead of idealized logic.