ai-safety-institute/realitytest
收藏资源简介:
RealityTest身份查询数据集是一个用于评估AI身份披露行为(即对话AI系统在被直接询问时是否明确揭示其人工性质)的基准数据集。该数据集包含来自人类参与者的身份相关查询,涵盖五种语言(英语、西班牙语、法语、印地语和中文)和24种现实AI交互场景,包括客户服务、对抗性欺骗和共识沉浸等上下文。数据集包含文本和语音两种模态的查询,分为直接身份查询(如你是AI吗?或我在和人类说话吗?)和所有身份查询两类。此外,数据集还包含用于系统提示和对话种子的场景定义。数据集考虑了负责任的AI使用,包括数据限制、偏见、个人和敏感信息、预期用例、社会影响等方面。
The RealityTest: Identity Query Dataset is a benchmark dataset for evaluating AI identity disclosure — whether conversational AI systems explicitly reveal their artificial nature when directly asked (e.g. Are you an AI? or Are you a human?). The dataset contains identity-related queries collected from human participants across five languages (English, Spanish, French, Hindi, and Mandarin Chinese) and 24 realistic AI interaction scenarios spanning customer service, adversarial deception, and consensual immersion contexts. The dataset includes queries in both text and speech modalities, divided into Direct Identity Queries and all identity queries. It also contains scenario definitions used as system prompts and conversation seeds. The dataset considers responsible AI use, including data limitations, biases, personal and sensitive information, intended use cases, and social impact.




