An Israeli startup exited stealth with a $62.5M “behavioral world model” that claims to predict human behavior — then conceded its headline “98% accuracy” doesn’t mean what it sounds like. The capital came mostly from the founders and the chairman’s own firm, and the product is sold to brokerages to forecast the very traders those brokerages take the other side of.
The number is confirmed — but it’s founder money plus capital from Infinity Group, the firm founded by chairman Amir Gal-Or, father of the CEO. No third-party lead priced the company, and no valuation was disclosed. The “seed” label borrows the credibility of a VC round 42AI didn’t raise.
The marquee “98% accuracy” collapsed under its own CEO’s clarification into “whether flows fall within a predicted range.” The more honest metric — an F1 of 0.8–0.85 for directional trading signals — is respectable but ordinary, and entirely unaudited. Core evidence is self-published preprints, not peer review.
The buyer is a brokerage; the person being modeled is that brokerage’s own trader. Those interests are opposed. Asked what stops a broker from trading against a client it knows is about to panic, the founders called it “a valid concern” and pointed to an internal ethics committee with no disclosed charter.
Key Finding: 42AI has a genuinely differentiated wedge — low-latency behavioral prediction sold into regulated finance — and rides a hot, well-capitalized “behavioral world model” wave. But almost nothing is externally verified: the raise is largely insider capital, the flagship accuracy claim was recanted on the record, and the central ethical problem — monetizing an information asymmetry against the person being modeled — goes unanswered.
42AI turns a person’s behavioral traits into mathematical representations, feeds them to a transformer world model, and bolts a language model on top to narrate the output. Each stage adds a claim worth testing.
KYC, onboarding, support calls, voice/sentiment, transfers, trading activity — linked per customer across ~30M people.
Behavioral tendencies converted into mathematical representations — the step 42AI treats as proprietary.
A transformer outputs probability distributions over decisions — instrument, direction, volume — up to 24 hours ahead.
A separate, unnamed language model narrates why. The explainability edge may rest on a third-party LLM.
Brokerage uses the forecast for hedging, margin, and liquidity decisions — against its own traders’ positions.
The differentiator 42AI sells — explainability — may not be theirs to sell. The core model is described as self-trained, but a “separate language model generates the explanations,” with no vendor named on the site, in any privacy policy, or in job postings. No cloud or model vendor (OpenAI, Anthropic, AWS, GCP, Azure, NVIDIA) is disclosed anywhere we could find.
42AI marketed a 98% accuracy figure. Pressed, CEO Amit Gal-Or reclassified it as “a measure of whether trading flows fall within a predicted range” — not whether the model gets individual decisions right. The metric that survives is an F1 score of 0.8–0.85 for predicting direction and volume: a credible directional signal, not clairvoyance, and audited by no one outside the company. Selling a loudly-advertised “98%” into regulated brokerages is exactly the territory where the SEC has already penalized firms for AI-washing.
The $62.5M figure is real. What it represents is the question — because external investors didn’t set the price, and most of the names behind it share a surname.
The money came from the founders and Infinity Group, the PE/VC firm founded and run by chairman Amir Gal-Or — father of CEO Amit Gal-Or. Amir Gal-Or and Infinity are independently documented and legitimate, which confirms the source but also confirms the round is largely family and insider money. No outside lead performed diligence or set a valuation.
The capital was accumulated over roughly three years — closer to patient self-funding than a single market-tested raise. There is no disclosed valuation, no cap table, no external syndicate. For a company claiming ~30M people in its datasets and ten brokerage clients, the absence of any priced external round is the tell.
Against its peers, 42AI is the thinly-capitalized outlier. Simile raised ~$200M (Greenoaks) at a ~$2B valuation; Yann LeCun’s AMI Labs raised over $1B; Humans& and Aaru report nine-figure rounds. 42AI’s edge is a narrow regulated-finance wedge — not its balance sheet, and not external conviction.
Six structural risks the $62.5M does not resolve.
The product forecasts how a brokerage’s own traders will behave so the brokerage can manage hedging and margin. Buyer and subject have opposed financial interests. Asked what stops misuse, founders conceded it was “a valid concern” and offered an ethics committee with no disclosed charter, membership, or enforcement power.
“98% accuracy” was recanted by the CEO in the same interview. The SEC has penalized advisers (Delphia, Global Predictions) for overstating AI, and Massachusetts opened a brokerage-AI probe. A headline accuracy claim sold into regulated finance sits squarely in that enforcement lane.
$62.5M mostly from founders and the chairman’s own firm, over three years, with no external lead and no valuation. The “seed” framing implies market validation that never happened.
Behavioral profiles on ~30M people drawing on biobank and government datasets, with no public privacy policy or subprocessor list. The “model never knows who the person is” claim is inconsistent with building individual profiles linked to live accounts.
The explanation layer — 42AI’s stated differentiator versus black-box models — relies on an unnamed “separate language model.” If that’s a third-party LLM, the moat is rented.
“Seven figures” of revenue from ~10 unnamed clients means any single logo loss is material. Key evidence is a self-published arXiv preprint and a Medium post — not peer-reviewed — and each new domain reportedly requires full retraining.
42AI is a real wedge wrapped in a borrowed word. “Seed” implies outside conviction that didn’t happen; “98% accuracy” implies a precision the CEO walked back on the record. The technology may well work as a directional signal — but the business model monetizes an information asymmetry against the very people it profiles, with no privacy policy, no external audit, and an ethics committee that exists mostly on paper. Before any brokerage integrates this, the question isn’t whether it predicts — it’s who gets hurt when it does.
Based entirely on publicly available information, including the TechCrunch announcement of November 2, 2026. Every performance and traction figure is company-claimed and not independently verified.
We build and maintain websites, help businesses of every size put AI to work, and run growth marketing that’s held to the same standard as our research.
