人机混合智能在教育考试中的应用研究

Research on the Applications of Human-Machine Hybrid Intelligence in Educational Examinations

摘要: 人工智能正加速进入教育考试的命题、评卷、分析与安全治理环节，在提升处理效率和一致性的同时，也带来效度偏离、算法偏差、数据安全与责任归属等风险。本文采用文献研究、典型案例分析、比较分析和规范分析，并辅以公开数据仿真，构建“任务分解–机器处理–风险判定–人工复核–反馈改进”的人机协同机制，并结合典型案例，比较其在智能命题、协同评卷、考试分析和安全监控中的职责配置、人工介入条件与证据边界。仿真显示，阈值提高可改善一致性，但会增加复核量。研究表明：人机分工不宜采用固定比例，而应依据考试利害程度、任务可验证性、模型置信度和异常风险动态调整；涉及成绩认定和违规处分的高利害结果，应保留实质性人工复核、申诉救济和组织最终负责机制。据此，本文面向河北省教育考试提出“准备–影子测试–低风险试点–受控扩展”的分阶段应用路径。

Abstract: Artificial intelligence is increasingly used in item development, scoring, analysis, and test-security governance. While improving processing efficiency and consistency, it may also create risks involving validity, algorithmic bias, data security, and accountability. Through literature review, comparative case analysis, and normative analysis, supplemented by a public-data simulation, this study develops a human-machine collaboration mechanism comprising task decomposition, machine processing, risk assessment, human review, and feedback-based improvement. Typical cases are compared to specify responsibilities, conditions for human intervention, and evidentiary boundaries in intelligent item development, collaborative scoring, examination analysis, and security monitoring. The simulation shows that higher confidence thresholds improve agreement but increase human-review workload. The findings indicate that responsibilities should not be assigned through a fixed human-machine ratio; they should be adjusted according to assessment stakes, task verifiability, model confidence, and anomaly risk. High-stakes decisions concerning score determination or misconduct sanctions require substantive human review, appeal procedures, and final organizational accountability. For educational examinations in Hebei Province, the study proposes a phased route from preparation and shadow testing to low-risk pilots and controlled expansion.