心理测量和人工智能的相互促进与协同发展
The Mutual Promotion of Psychological Measurement and Artificial Intelligence
詹沛达 1丛艳章 1李浩宇 1倪子旭 1金舒悦 1周萱孜 1黎鑫 1郝文慧 1张睿锋 1何可人2
作者信息
- 1. 浙江师范大学心理学院教育神经智能测量实验室
- 2. 浙江师范大学心理健康教育与发展中心
- 折叠
摘要
心理测量是探究人工智能(Artificial Intelligence, AI)智能特征、揭示其发展规律的关键手段，也是理解人智复杂交互关系的重要途径；AI则为心理测量提供降本增效的技术支撑，推动其向个性化、隐性化、智能化与多模态化革新。本文系统梳理心理测量赋能AI和AI赋能心理测量两条路径的研究进展，总结当前面临的挑战并提出应对建议。总体而言，心理测量为AI提供标准化评估框架与理论约束，有助于厘清其智能特点、边界与潜在风险；AI则通过方法与技术的创新反哺心理测量的工具开发、构念建构与理论演进。两者并非孤立发展，而是相互促进、协同演化，其深度融合有望在理论驱动与数据驱动双轨并行下催生出人智共生时代的测量科学新范式。
Abstract
Recent advances in artificial intelligence (AI), particularly large language models (LLMs), have profoundly reshaped both technological landscapes and scientific inquiry in psychology. This paper systematically examines the bidirectional synergy between psychometrics and AI, arguing that their integration is not merely instrumental but foundational for the emergence of a new measurement paradigm in the era of humanAI coexistence.On one hand, psychometrics provides essential theoretical frameworks and methodological rigor to evaluate AI systems. By adapting established psychological constructssuch as intelligence, creativity, personality, theory of mind, and moral reasoningresearchers can assess AIs psychological profiles, benchmark its capabilities against human standards, and uncover its developmental trajectories and limitations. Current approaches include the direct application of human psychometric instruments (e.g., WAIS, Big Five inventories) and large-scale AI-specific benchmarks (e.g., MMLU, BIG-bench). Empirical studies show that models like GPT-4 already match or exceed human performance in domains such as verbal reasoning and creativity fluency, yet lag in visual-spatial tasks and emotional depth. However, these methods often suffer from anthropomorphic bias, prompt sensitivity, and a lack of grounding in psychological theory, leading to questionable validity. Moreover, most evaluations rely on classical test theory, yielding ordinal rankings that hinder fine-grained, cross-model comparisons on a common metric. To address this, we advocate for integrating modern psychometric modelssuch as item response theory and cognitive diagnosis modelsto enable equated, interpretable, and diagnostic assessments of AI capabilities. Beyond trait measurement, psychometrics also enables the systematic study of (1) human cognition, emotion, and attitudes toward AI; (2) AIs impact on human psychological development in education, mental health, and socialization; and (3) interdependent dynamics in humanAI collaboration, including role allocation, trust calibration, and interaction patterns.On the other hand, AI is revolutionizing psychometrics itself. LLMs facilitate automated item generation, significantly reducing development costs and human bias while enabling dynamic, context-sensitive assessments. AI also supports implicit and multimodal measurement through the analysis of natural language, facial expressions, voice, and behavioral logs, moving beyond traditional self-report questionnaires. Furthermore, deep learning enables the unsupervised extraction of latent psychological dimensions from real-world data (e.g., social media), potentially refining or even redefining psychological constructs. In scoring and interpretation, AI systems can provide reliable, scalable, and diagnostic feedback on open-ended responses, while predictive modeling allows for early risk detection and personalized interventions. For instance, transformer-based models like CLIP and Flamingo enable cross-modal integration of text, image, and audio, while graph neural networks model complex problem-solving trajectories. Nevertheless, these advances raise critical concerns about algorithmic black boxes, data bias, cross-cultural fairness, and the ethical use of sensitive behavioral data.The paper identifies key challenges in both directions. For psychometrics-to-AI, issues include the risk of uncritical anthropomorphism, unstable AI responses due to prompt and parameter sensitivity, cultural bias in benchmarks, and the lack of fine-grained diagnostic feedback in current evaluations. For AI-to-psychometrics, concerns center on transparency, the validity of AI-generated content, data privacy, and the scarcity of interdisciplinary expertise. To address these, we propose five strategies: (1) developing AI-specific psychometric paradigms that account for AIs dataalgorithmmodel nature and incorporate functional traits (e.g., reasoning stability, cross-context adaptability); (2) creating stability metrics, such as output consistency indices, to quantify AIs trait volatility; (3) embedding core psychometric principlesreliability, validity, fairnessinto AI systems from the design stage; (4) training domain-specific AI psychometricians that integrate psychological theory with computational methods; and (5) establishing ethical guidelines for data collection, use, and synthetic data generation.Ultimately, we envision a co-evolutionary future in which psychometrics and AI mutually inform each other: psychometrics offers interpretability, standardization, and ethical grounding, while AI contributes scalability, adaptivity, and multimodal integration. Their deep integration may give rise to intelligent psychometricsa new discipline that merges theory-driven and data-driven approaches to understand both human and artificial minds in an increasingly intertwined world. This synergy not only advances scientific understanding but also ensures that AI development remains human-centered, scientifically sound, and ethically responsible.关键词
心理测量/人工智能/心理特质/基准测试/人智交互Key words
psychological measurement/artificial intelligence/psychological construct/benchmarks/human-AI interaction引用本文复制引用
詹沛达,丛艳章,李浩宇,倪子旭,金舒悦,周萱孜,黎鑫,郝文慧,张睿锋,何可人.心理测量和人工智能的相互促进与协同发展[EB/OL].(2025-09-29)[2026-04-02].https://chinaxiv.org/abs/202509.00251.学科分类
科学、科学研究/计算技术、计算机技术
评论