Agricultural Bank Chairman Gu Shu: Large models currently face three major types of risks.

On June 18th, at the 2026 Lujiazui Forum "Plenary Session Six: Empowering Financial High-Quality Development with Technological Innovation", Gu Shu, Chairman and Executive Director of Agricultural Bank of China, stated that in the process of applying large models, the main risks faced by AI large models currently include model black boxes, model illusions, and uncertainty brought about by model autonomous thinking and decision-making. In his view, this uncertainty risk can be further divided into three major categories: First, the problem of interpretability brought about by the massive parameters. The parameter scale of mainstream large models has reached billions or even trillions, and the massive parameter matrix operations and nonlinear superimposition result in opaque decision mechanisms and output results that are difficult to explain. Second, the test of accuracy brought about by probability generation. The decision-making process of large models is not linear thinking like human decision-making, but is based on the statistical regularity of the probability of source tokens in massive training data, essentially probabilistic, not logical deductive facts. When there is insufficient data and factual basis, self-consistent illusions are easily generated. Third, the models can now think and make decisions autonomously. With the evolution of large models and the deep application of intelligent agents, they have surpassed the relatively fixed paradigm of traditional software input-output, and can think autonomously, to a certain extent amplifying the risks of uncontrollability and unknown results.