超越“黑箱”:审计大模型的因果可解释性框架
The “black box” nature of large models stands in fundamental tension with the rigid requirement that audit conclusions be traceable, verifiable and reviewable, and this tension has become a bottleneck constraining the large-scale application of audit large models. Drawing on explainable artificial intelligence (XAI) and causal inference, this paper constructs a causal chain-based explainability framework for audit large models, which proceeds in three stages: in the first stage, a natural language processing model extracts voucher entities, and SHAP values quantify the contribution of each feature to the risk score; in the second stage, a causal graph depicts the causal paths among features and distinguishes direct causes from spurious correlations, while a large language model generates counterfactual explanations; in the third stage, the audit report presents the risk level, the confidence of key features and the complete causal chain. A comparison based on an accounts receivable audit scenario shows that this design can turn opaque scores into traceable and verifiable audit evidence chains. The paper further discusses hallucination risks and data governance costs, and proposes countermeasures such as human-in-the-loop review and “T-shaped” talent cultivation.
Authors
- 莫磊
- Yanghao Luo (ORCID: https://orcid.org/0009-0006-0069-6054)
Institutions
- Guangxi University of Finance and Economics (CN)
Publication Details
- Journal
- Journal of Frontier in Economic and Management Research
- Published
- 2026-11-19
- DOI
- https://doi.org/10.63944/sjmx.jfemr
- Primary Topic
- Explainable Artificial Intelligence (XAI)
- Type
- article
- Field-Weighted Citation Impact
- 0.00