42AI wants to predict your next move before you make it

Here’s a setup straight out of science fiction: what if an AI could predict what you are going to do before you actually do it?

Israeli startup 42AI thinks it has cracked the formula. After three years in stealth, the company just surfaced with a $62.5 million seed round and a transformer-based behavioral world model designed to forecast human decisions. The funding was raised over three years from its founders and Infinity Group, the firm founded by 42AI chairman Amir Gal-Or.

The system is already running live at 10 large financial brokerages, pulling in over seven figures in annual revenue.

Mapping the world inside your head

World models usually live in robotics labs, helping autonomous machines guess how physical surroundings shift when an arm moves. 42AI takes that concept and points it squarely at the human mind, operating under the premise that each individual is their own world.

Traditional financial models simply look at your past trades and extrapolate trends. 42AI founders Amit Gal-Or and Ehud Ezra argue that while people change, their underlying psychological tendencies remain relatively consistent, even in wild, brand-new situations.

To map those tendencies, the system tracks seemingly disconnected touchpoints: how you complete identity verification, your voice tone and sentiment during customer support calls, call frequency, money transfers, and the timing of your orders.

42AI turns those observations into mathematical trait representations, feeding them into a transformer architecture that outputs decision probabilities across one-, four-, and 24-hour windows. A separate language model drafts natural-language explanations for each forecast.

The startup spent three years consulting with neuroscience and PTSD specialists, alongside 2011 chemistry Nobel laureate Dan Shechtman. Starting with psychological assessments of thousands of participants, they expanded the dataset with government and biobank records covering over 2 million individuals, eventually building behavioral profiles spanning roughly 30 million people.

What 98% accuracy actually means

The company claims 98% accuracy in forecasting market trading, but there is a major catch. Gal-Or clarified that the metric measures whether trading flows land inside an anticipated range, not whether the system can nail every individual trade. For direction and volume, the model achieves an F1 score between 0.8 and 0.85.

Brokerages use these predictions to manage margin calls, balance liquidity, and dynamically hedge exposure. Gal-Or claims one brokerage unlocked roughly $30 million in added value through dynamic hedging.

Still, the box has blind spots. Ezra acknowledged that if a trader shifts their mind after hearing advice from a friend over coffee, the model misses that unobserved context, and accuracy drops.

Then there is the elephant in the room: platform exploitation. 42AI claims the model is identity-blind and does not ingest personally identifiable information, yet it still builds tied individual behavioral profiles. If a brokerage knows exactly when a client is about to panic-sell, what stops them from trading against that very client? The founders point to regulators and an internal ethics committee, but that offers little comfort to everyday retail traders.

Why it matters

Transferring world models from physical robotics into cognitive modeling is an exceptionally clever technical pivot, but it puts enterprise AI on a collision course with serious ethical boundaries.

For enterprise operators, the value proposition is obvious. Managing systemic risk by forecasting collective human panic offers massive margin and liquidity advantages. One brokerage reportedly generated tens of millions in hedging value, proving that behavioral intelligence can easily monetize.

For retail users and everyday consumers, however, this technology creates a deeply uneven playing field. If an exchange or broker understands your cognitive vulnerabilities better than you do, the temptation to front-run or design friction that extracts margin is almost irresistible. In gaming, fraud prevention, or lending—areas 42AI plans to tackle next—that predictive edge feels even more invasive.

We think 42AI’s method of extracting latent psychological traits is much more robust than naive historical trend matching. But building behavioral profiles using biobank and call-sentiment datasets without active, explicit client consent sits in a grey zone. As predictive models shift from guessing what pixels come next to guessing what people do next, regulators will almost certainly step in to decide whether predicting human panic crosses the line into market manipulation.