Goldman Sachs says AI is shifting from text prediction to world models

Goldman Sachs is pushing its AI story beyond text generation. In an essay published April 23, George Lee and Dan Keyserling argued that the frontier is shifting from large language models, which predict words, to world models that simulate reality, test actions and reason about consequences.

That distinction matters inside a bank because it changes what AI is supposed to do. Goldman said LLMs are strong at pattern completion, but they run into trouble when mistakes carry real costs. A model that can extract covenants from loan documents or help draft an investment committee memo is useful. A model that can simulate how a hurricane season reshapes insured-loss distributions across a reinsurance portfolio, or forecast how a policy shock cascades through markets and behavior, gets closer to the kind of scenario work that underpins trading, credit and risk decisions.

For Goldman employees, the practical shift is from faster writing to better decision support. In market-facing teams, world models could help stress-test assumptions before a trade is placed or a client recommendation is made. In risk management, they could make scenario analysis faster and more granular. In forecasting and planning, they could reduce the time spent stitching together inputs from credit, regulation and operations before a human makes the call. Goldman’s own framing is that prediction alone is not enough once AI is pushed into robots, supply chains and enterprise workflows.

The essay also fits a broader internal push. Goldman’s 2025 annual report said the firm launched One Goldman Sachs 3.0, a new operating model propelled by AI, and said the technology could unlock significant productivity gains that the bank can reinvest in client service. In January 2026, chief information officer Marco Argenti said AI models are becoming more than just chatbots, and Goldman has said it is looking at every single process in the firm to see how AI could interact with it.

That puts the bank in the middle of a wider race to define where AI goes next. MIT Technology Review highlighted world models on April 21 as one of the 10 things that matter in AI right now, underscoring that this is a live debate, not an internal slogan. Goldman Sachs Research has also previously estimated that generative AI could raise global GDP by 7% over a 10-year period, while later research suggested AI investment could approach $100 billion in the U.S. and $200 billion globally by 2025. For bankers, analysts and operators, the message is clear: the advantage will go to people who can pair human judgment with systems that simulate what happens next.