A new public test from Rehearsals, Google DeepMind and Kaggle says top AI models misjudge real people on health, habits, decisions and values.
A new public test called Human Behavior Bench checks how well AI models understand what real people think and do. Rehearsals, a New York company, launched it on October 20, 2026, with Google DeepMind and Kaggle. Its first result is that today's top AI models fall short on health, habits, decision making and values.
A benchmark is a standard test used to score and compare AI models. Rehearsals says this is the first public one for understanding human preference and action. It draws on interviews with 1,244 U.S. adults, covering 750 categories and 95,000 responses.
Most of the data is kept private, with an 80% private holdout group, so the test stays valid. Part of it is public. The benchmark is also hosted on Kaggle, so any company or individual can score their latest model.
Rehearsals says AI models make wrong assumptions about people. In its words, they think most people are less aware, universally frugal and generally lazier than they really are. The company points to four weak areas:
The release mentions examples and scores, but they are not included in the material available here.
Rehearsals sells a platform that predicts marketing outcomes by simulating AI twins of real people. It built this test as the scorecard for its own model, called Engram, which it is previewing with select partners. The company says Engram beats models from Google, Anthropic, OpenAI and SpaceXAI at predicting human behavior, and that it was not trained on benchmark content.
Those claims come from the company's own announcement. It says the results were produced publicly with DeepMind on Kaggle. Whether outside researchers have checked them is not stated.
If you use AI to guess what customers want, this is a reminder that its picture of people can be off. Checking its answers against real feedback from real customers is still sensible.
This first version covers only the U.S. and a smaller group of people. Rehearsals plans a next version with more participants, four more countries and a broader range of topics. It expects to launch that before the end of 2026, though it says plans may change.
AI can be wrong about how people live and choose, so check its guesses about customers against real feedback.
OpenAI has about $11 billion in hand at a $1.25 trillion value, below its first ask. Microsoft and NVIDIA shareholders have reason to watch what comes next.
Read more: OpenAI raises $11 billion, but at a lower price than it wantedVisa tested 22 AI models on realistic shopping tasks. They passed just over half, and often bought the wrong thing anyway.
Read more: Visa tested 22 AI shoppers, and most still make costly mistakesDARPA wants AI to hunt for design flaws in stock, bond, energy and cloud markets, so enemies cannot use them to cause economic damage that reaches households.
Read more: Pentagon wants AI to find weak spots in stock and energy markets
