Executive Summary

Key Figures

The first two cards give the size of the data the model stands on and the share of it that came from outside. The last two give what the model scored in the field and what the company spent getting there.

Source: Ram Iyer, 42AI comes out of stealth with $62.5M and a world model that can 'predict human behavior' (TechCrunch, November 2, 2026). All four are numbers the company reported itself.

42AI spent three quiet years after founding before showing itself on November 2. The disciplines it pulled together over those years were not one field. It combined psychological research with neuroscience and genetics, took advice from experts in psychology and in post-traumatic stress disorder, and listed Dan Shechtman, awarded the 2011 Nobel Prize in Chemistry, among its advisers. The co-founder and CEO is Amit Gal-Or and the CTO is Ehud Ezra. The chairman and co-founder, Amir Gal-Or, is the CEO's father and the man who founded the investment firm Infinity Group.

The funding structure is unusual. The $62.5 million seed came from the founders themselves and from Infinity Group over three years, and Gal-Or said it was not entirely raised from external investors. The company says revenue has passed seven figures, and adds that it now plans to raise outside money.

More concrete than the money is what is already running. Ten large brokerages use the model to manage liquidity, margin requirements and hedging exposure. One of them, the company says, generated roughly $30 million in additional value through dynamic hedging. There are two ways the company gets paid. A customer pays a flat fee, or a percentage of the value the platform helps them gain.

Two numbers were offered as a scorecard. On predicting market trades the company claims 98% accuracy, and Gal-Or himself set out what that means: it measures whether trading flows fall within a predicted range, not whether each individual trade was predicted correctly. The other number is an F1 score of 0.8 to 0.85 for predicting trading direction and volume. Both values are company-reported, and no external verification under the same conditions was published with them.

The premise the company puts forward is short. "Behavior is not deterministic; it's constantly probabilistic," Gal-Or told TechCrunch. His co-founder Ezra added the rest of it: "A week from now, you might be a different person, but your tendencies will always play out in certain areas to a probability or distribution."

So the method is not to stretch past behavior forward into a forecast. Instead the company watches how a person reacts to events that look unrelated to the task at hand, and infers the motivations underneath. For an investor or trader, such events include completing identity checks, onboarding, support calls, money transfers and trading stocks. Ezra described watching the early contact first: when a customer talks to support or sales at the start, the company looks at the sentiment and the voice and at a lot of the interactions they are having, down to how many times they call. Later come the trades themselves, he said — how you buy and sell and when, read together with the market at that stage.

Those observations are converted into mathematical representations of traits, which are fed into a transformer-based world model, and a separate language model is left to turn the output into an explanation a person can read. "In a world model, every individual is the world," Ezra said.

Ezra also described which way the training is pushed. During simulations, the standard the reward design holds to is that a person's underlying tendencies stay intact. If someone is usually risk-averse and the model suddenly treats them as a risk-taker, "the target is off," he said.

The name the company chose places it alongside other work that has been building world models, though the comparison is ours rather than the company's: Google DeepMind's Genie, Meta's V-JEPA, and World Labs, founded by Fei-Fei Li. Those systems learn how the physical world moves. The world 42AI is trying to build sits inside a single person.

The founders acknowledged a limit. "There's a natural ceiling because you really don't have access to everything that a person experiences," Gal-Or said. Ezra added that accuracy also suffers when the system misses the situational context of a moment, as when someone changes a trading decision on the advice of a friend or a colleague.

The training data the company described is not one block but three layers of entirely different character. At the bottom is observation 42AI designed itself: a few thousand people studied through psychological assessments and behavioral scenarios, their traits converted into numbers, a layer where it is clear what the participants agreed to. In the middle are the government datasets and biobank records covering more than two million people. At the top are the records the brokerage customers had accumulated in the course of doing business, and with those included the total reaches almost thirty million.

What the top layer actually is shows in the wording the report chose. In live deployments the model analyzes the records a customer already held, and those include transactions and support calls. This is not material gathered to do behavior prediction. It is material that had already piled up in the running of sales and support. This layer is the place where one more use, trait prediction, was added to records that were already there.

Of the three layers the middle one is the odd one out. Government datasets and biobanks were never built for trade prediction. A biobank is an institution set up to gather participants' biological samples and health records for medical research, and government statistics and administrative records are collected in the name of public health and policy design. The uses written into a participation consent form are written to that scope.

So a set of questions attaches to this layer that does not attach to the other two. Which country and which datasets, through what procedure and under what conditions, and on what basis secondary use was permitted. The company disclosed none of it. What is known is the scale, more than two million people, and the category, government and biobank.

On privacy the company's answer is plain. The system does not need personally identifiable information, and as Gal-Or put it, "the model never knows who the person is." Read straightforwardly, that means it does not handle identifiers such as names or national ID numbers.

One sentence later, though, the same report adds an observation of its own: it appears the company still creates individual behavioral profiles from linked customer activity. Erasing a name and declining to tell people apart are two different jobs. With the identifier gone, the trait coordinates still attach to a particular account, and on the brokerage's side that account can be joined to a real name at any time.

Asked about misuse, Ezra called it "a valid concern." What the founders offered as safeguards were regulation, monitoring and an internal ethics committee. They did not explain what, if anything, would stop a customer from using the technology to gain an advantage over the very people it profiles. In a setup where the brokerage is the party that learns a client's tendencies first, that gap is not a small one. The pricing points the same way: a customer can pay a percentage of the value gained instead of a flat fee, which reads as meaning that the more a brokerage extracts from its clients, the larger the share 42AI collects.

Amit Gal-Or was nonetheless confident that the technology could help address exactly those risks. "We are very proud that actually we are part of the solution; we are the solution to this problem," he said.

The stage the company is drawing for itself next is wider. It is exploring gaming, fraud prevention and cybersecurity, and over the long run it says it wants the model to help robots and AI agents understand people. Ezra conceded that the model has to be taught the rules and context of each new domain, and that this will take time. Each time that teaching finishes, the trait coordinates of thirty million people cross into another purpose.

The moment data steps outside the consent it was given is usually not the moment a model changes. It is the moment the data moves. Records sitting in a biobank do nothing at all while they sit there. Their purpose divides at the instant they are copied into another organization's training set.

It is hard to find a new technical element in the 42AI case. Psychometrics, transformers and customer behavior analysis at financial firms all existed before. What changed is the connection. A path opened between material gathered in the name of public health and a brokerage's hedging model, and what it took to open that path was not progress in modeling but a widening of the reach into data.

This blog has covered the same shape twice. Once when Google's promise not to use the contents of Gmail and Photos for advertising turned conditional, and once when Microsoft wrote into a contract that student data would not be used to train AI. In all three cases what changed was never the data itself but the scope of the promises around it.

In the end the case turns the question back on your own organization. If you have ever trained on material received from outside, the three items below are already yours to answer.

• Write down, per source, what each piece of the training data was originally gathered for. Once everything is merged it is hard to trace which row came in under which consent
• Track the routes data takes out of the organization more closely than model changes. A change that lifts performance leaves a record; copying a dataset once and handing it over leaves no mark
• Do not use removal of identifiers as a synonym for anonymization. If the receiving side holds a linking key such as an account or a device, people can be reassembled even with the identifiers gone

Whether the F1 of 0.85 that 42AI reports holds up is something time will settle. The question that needs an answer sooner sits on the thirty-million side. If a record of mine is in there, which consent form did I sign that counted as permission?

Thank you for reading this far. The company name and its founders, the disciplines and advisers assembled over three years, and the relationship between Infinity Group and the founders all come from the TechCrunch report of November 2, 2026. The $62.5 million seed with its three-year structure, Gal-Or's statement that it was not entirely raised from external investors, the seven-figure revenue and the flat-fee-or-percentage pricing are in the same piece. The report also carries the ten brokerages and the liquidity, margin and hedging uses, one brokerage's claim of roughly $30 million in additional value, and the definition of the 98% accuracy figure alongside the F1 of 0.8 to 0.85. The three layers come from there too: a few thousand people from psychological assessments, more than two million from government and biobank datasets, and almost thirty million with customer data included. So do the inputs the model reads, the trait representations, the transformer and the separate language model, the reward design that preserves tendencies, and the statement that live deployments analyze existing records. The remark that the model never knows who the person is, and the report's observation that individual behavioral profiles are still built, are there as well. Regulation, monitoring and an internal ethics committee as safeguards, Gal-Or's answer that they are the solution, and the expansion plans into gaming, fraud prevention and cybersecurity with the caveat about retraining per domain round out what the report supplied. Placing 42AI next to Genie, V-JEPA and World Labs was a choice this piece made, and the report draws no such comparison. Section 5, which moves the point where data changes purpose into the domain of data management, and the three checks listed there, are this article's own reading. We would be glad to hear how your organization records the collection purpose of data brought in from outside.

Pebblous Data Communication Team
November 3, 2026