心理学与工程学学生寻求有关AI聊天机器人对心理困扰提示响应的论文和资源
英文摘要
A student nearing completion of a Psychology degree and studying Systems Engineering is seeking research papers, datasets, benchmarks, and methodological advice for a project comparing how AI systems (ChatGPT, Gemini, Wysa, Replika) respond to prompts involving psychological distress at different intensity levels. The study aims to analyze linguistic and safety responses—such as empathy, psychoeducation, crisis resources, or refusals—rather than clinical effectiveness. Key interest areas include how responses change with prompt intensity, declarative versus question phrasing, explicit versus indirect distress, and the influence of hidden safety layers, system prompts, model versions, and stochastic outputs. The request also covers reproducibility concerns, moderation classifiers, and product updates.
中文摘要
一名即将完成心理学学位并攻读系统工程的学生正在为一项研究项目征求论文、数据集、基准和方法论建议。该项目旨在比较AI系统（ChatGPT、Gemini、Wysa、Replika）如何回应不同强度心理困扰的提示，重点关注语言和安全层面的响应（如同理心、心理教育、危机资源或拒绝），而非临床有效性。研究兴趣包括提示强度变化对响应的影响、陈述句与疑问句的区别、显性与间接困扰的表达，以及隐藏安全层、系统提示、模型版本和随机输出等因素。同时还涉及可重复性、审核分类器和产品更新等技术问题。
关键要点
The user is a dual Psychology and Systems Engineering student seeking resources for a project on AI responses to psychological distress.
发帖者是心理学与系统工程双学位学生，正在为研究AI对心理困扰响应的项目寻求资源。
The systems to be compared are ChatGPT, Gemini, Wysa, and Replika, spanning general-purpose and mental-health-oriented chatbots.
待比较的系统包括ChatGPT、Gemini、Wysa和Replika，涵盖通用型和心理健康导向的聊天机器人。
The analysis focuses on linguistic and safety responses (empathy, crisis resources, refusals, etc.), not clinical outcomes.
分析重点在于语言和安全响应（同理心、危机资源、拒绝等），而非临床疗效。
Key methodological interests include prompt framing, intensity levels, hidden safety mechanisms, model versions, and reproducibility.
关键方法论兴趣点包括提示措辞、强度层级、隐藏安全机制、模型版本和可重复性。