Artificial Intelligence Decision-Making Responsibility Mapping and Risk Hierarchy Management System

Abstract

An artificial intelligence decision-making responsibility mapping and risk stratification management system includes an electrically connected computer-readable storage medium and a processor. By setting interception points in the inference pipeline of the model, the system can capture in real time the input data of users participating in the decision-making event, the output data of the model, the policy rule data, and the state data of the state situation. After normalizing the data into multiple quantitative values, the system performs a weighted summation operation based on multiple preset weight values ​​to calculate multiple responsibility sharing ratios. Then, based on the multiple responsibility sharing ratios and the role rule base, the system generates responsibility stratification results to calculate the risk level of the decision-making event. When the risk level exceeds a threshold, the system performs real-time access control on the model and triggers risk escalation routing through the interception points, thereby achieving the technical effect of improving the real-time attribution capability and real-time access control capability of the inference pipeline.

Claims

1. An artificial intelligence decision-making responsibility mapping and risk stratification management system, comprising: a computer-readable storage medium for storing a plurality of computer-executable instructions; and a processor for being electrically connected to the computer-readable storage medium and executing the plurality of computer-executable instructions, such that the processor is configured to: parse a user, a model, a policy, and a state context involved in a decision event, wherein, An interception point is set in one of the inference pipelines of the model, and the user's input data, the model's output data, the policy's rule data, and the state data of the state situation are captured in real time through the interception point; the input data, the output data, the rule data, and the state data are normalized into multiple quantified values, and the multiple quantified values ​​are weighted and summed according to multiple preset weight values ​​to calculate multiple responsibility sharing ratios, wherein the multiple responsibility sharing ratios correspond to the user, the model, the policy, and the state situation, respectively; Based on the multiple responsibility-sharing ratios and a pre-established role rule base, a decision-responsibility mapping and a role-based attribution model are executed to generate a responsibility stratification result. The responsibility stratification result defines a responsibility stratification and an attribution logic among the user, the model, the policy, and the state situation. Based on the multiple responsibility-sharing ratios and the responsibility stratification result, a risk level of the decision event is calculated. When the risk level exceeds a threshold, an immediate access control is executed on the model through the interception point, triggering a risk escalation route to execute either human intervention or advanced authorization management.

2. The AI ​​decision responsibility mapping and risk stratification management system as described in Request 1, wherein the computer can read and store a responsibility record corresponding to the decision event on a storage medium for post-event verification and evidence collection of the responsibility attribution of the decision event, the responsibility record being established by the processor and including multiple identification data, multiple responsibility sharing ratios, multiple role tags, the risk level and the threshold, wherein the multiple identification data and the multiple role tags correspond to the user, the model, the policy and the state situation, respectively.

3. As in Request 1, the AI ​​decision responsibility mapping and risk stratification management system, wherein the input data, after normalization, corresponds to the clarity of the input intent of one of the users; the output data, after normalization, corresponds to the output confidence score of one of the models; the rule data, after normalization, corresponds to the degree of compliance of one of the policies; and the state data, after normalization, corresponds to the situational risk level of one of the state situations.

4. As in Request 1, the AI-powered decision responsibility mapping and risk stratification management system, wherein the threshold is dynamically adjusted based on the processing result of one of multiple historical decision events and the feedback from human intervention, and the processor is further configured to: based on the feedback, calculate the false trigger ratio of one of the multiple historical decision events determined to be a false trigger, and increase the threshold when the false trigger ratio exceeds a first predetermined upper limit; and based on the processing result, calculate the missed trigger ratio of one of the multiple historical decision events confirmed to be a missed trigger, and decrease the threshold when the missed trigger ratio exceeds a second predetermined upper limit.

5. The AI-powered decision responsibility mapping and risk stratification management system as described in Request 1, wherein when the processor triggers the risk escalation route, it is further configured to: execute human intervention when the risk level exceeds the threshold to a first predetermined level, wherein the human intervention includes transmitting the decision event along with the multiple responsibility sharing ratios to a designated reviewer for review and to provide a processing decision; and execute advanced authorization management when the risk level exceeds the threshold to a second predetermined level, wherein the advanced authorization management includes elevating the processing authority of one of the decision events to an advanced authorized person, the authority of the advanced authorized person being higher than the authority of the designated reviewer, and the second predetermined level being higher than the first predetermined level.

Images