On June 18 at the Lujiazui Forum, Agricultural Bank of China chairman Gu Shu
outlined three main risk categories for large AI models. First, explainability:
mainstream models now run to hundreds of billions or trillions of parameters,
and massive matrix operations and nonlinear superposition make decision
mechanisms and outputs opaque and hard to interpret. Second, accuracy: models
generate by probabilistic token prediction from training data rather than
logical fact‑based deduction, so insufficient data or grounding can produce
coherent but false hallucinations. Third, autonomy risk: evolving models and
agentic applications are moving beyond fixed input‑output software paradigms to
autonomous reasoning and decision‑making, increasing process uncontrollability
and outcome uncertainty.
