42AI comes from stealth with $62.5M for its world model to predict human behavior | TechCrunch

A model of behavioral tendencies

42AI says its behavioral world model estimates the likelihood that someone will make a particular decision, including in circumstances they have not encountered before. CEO and co-founder Amit Gal-Or told TechCrunch that people’s choices are probabilistic, shaped by underlying psychological tendencies.

CTO and co-founder Ehud Ezra said those tendencies can persist even as a person changes. The company says it seeks to identify internal motivations rather than simply extend patterns from past actions.

To build a picture, the system considers responses to events that may appear unrelated, such as identity checks, onboarding, support calls, transfers and stock trades. Ezra said it can assess sentiment, voice and interaction frequency, alongside later buying and selling activity and market conditions.

Building the system

42AI turns observations into mathematical representations of traits, which feed a transformer-based world model that estimates possible decisions under varying circumstances. A separate language model generates explanations for those predictions.

Ezra said training is designed to preserve a person’s underlying tendencies during simulations. If someone generally avoids risk but the system represents them as risk-seeking, he said, “the target is off.” The company spent three years developing its approach, drawing on psychological research, neuroscience and genetics, as well as guidance from specialists in psychology and PTSD and Nobel chemistry laureate Dan Shechtman. Gal-Or said early work involved thousands of people in assessments and behavioral scenarios, followed by government and biobank data covering more than 2 million people. He said customer datasets have since grown to almost 30 million people.

Uses and accuracy claims

In current deployments, 42AI analyzes customer records such as transactions and support calls. Gal-Or said it does not need personally identifiable information and that “the model never knows who the person is.” The account also says the company creates individual behavioral profiles from linked customer activity.

For market trades, 42AI claims 98% accuracy, but Gal-Or said that figure reflects whether trading flows fall within a predicted range, not whether each trade is correctly forecast. He also reported an F1 score between 0.8 and 0.85 for predicting trading direction and volume.

Gal-Or cautioned that the model has a natural ceiling because it cannot access everything a person experiences. The account ends as Ezra begins to explain another factor that can reduce accuracy, without completing that explanation.