微软研究院的研究报告《Working with AI: Measuring the Occupational Implications of Generative AI》分析了20万份用户与Bing Copilot的对话数据，揭示了生成式AI对劳动力市场的实际影响。报告发现，AI目前主要扮演服务型助手的角色，辅助用户完成知识性工作中的具体子任务，而非完全取代人类。研究表明，在某些任务中，AI的辅助功能比直接执行更能提高用户满意度。这使得信息提供和支持性岗位（如客服、翻译）的AI适用性得分较高，但AI尚不能完全胜任这些工作。需要“代理”能力的岗位（如金融交易）也仅限于AI辅助人类执行。此外，即使是体力劳动，AI也能通过处理知识性部分（如辅助厨师起草食谱）来整合，展现了AI作为协作工具而非自主劳动者的广泛潜力。

💡 **AI主要作为服务型助手存在：** 研究表明，生成式AI在工作中的主要功能是作为一种服务型助手，帮助用户完成特定的子任务，而非直接替代人类完成整个工作。例如，AI可以总结信息、解释概念，但用户仍然是知识工作的核心执行者。

📈 **AI辅助提升用户满意度：** 报告指出，在某些情况下，当AI以辅助方式支持用户完成任务时，用户满意度反而高于AI直接尝试执行任务。这强调了AI在增强人类能力方面的价值，而非取代。

💼 **信息密集型岗位AI适用性高：** 提供信息或支持的角色，如客户服务代表和翻译，由于其工作活动与AI能力高度重叠，因此显示出较高的AI适用性得分。然而，AI目前仍无法完全独立胜任这些岗位。

🔗 **“代理”类岗位AI仅为辅助：** 对于需要更强“代理”能力（如执行金融交易）的岗位，AI仅能作为辅助工具，帮助人类执行相关操作，自身并不直接承担执行责任。

🛠️ **AI渗透各行各业的知识性环节：** 即使是包含大量体力劳动的职业，AI也能通过处理知识性部分（如为厨师起草食谱、营养要求）来整合，显示出AI作为协作工具，有潜力融入几乎所有行业，但目前更多是辅助而非自主工作。

Published on August 9, 2025 4:20 PM GMT

I recently read “Working with AI: Measuring the Occupational Implications of Generative AI”, a paper by Microsoft Research which offers a data-driven look at generative AI's real-world impact on the workforce by analyzing 200,000 user conversations with Bing Copilot. The research was done between January and September 2024. This is a good paper to read if you are interested in how generative AI is replacing, supplementing, and enhancing work related tasks. Below are my conclusions on the paper.

My primary conclusion is that AI still predominantly functions as a service-oriented assistant. There is a clear divergence between a user's goal (the knowledge work they are trying to do) and the AI's action (the service it performs to help). For example, a user researches a complex topic while the AI acts as an information provider, summarizing sources and explaining concepts. It is not replacing the knowledge worker, but augmenting them by handling a specific sub-task. The study notes that for some tasks, user satisfaction is higher when the AI assists the user than when it attempts to perform the task directly.

This assistive capability means roles centered on providing information or support, like customer service representatives and translators, already show high AI applicability. Many of the work activities in these roles show a strong overlap with demonstrated AI capabilities, though the study finds this does not extend to performing these jobs in their entirety.

In contrast, roles requiring more "agentic" capabilities, such as executing financial transactions, are not being performed by the AI; it only assists the human who performs the action.

Finally, this model of assistance extends even to occupations with significant physical components. The study shows that LLMs are handling the knowledge-based aspects of manual labor. For example, it is helping chefs and cooks draft recipes and nutritional requirements, even if it cannot chop the vegetables. This shows a broad, if narrow, potential for current generative AI to be integrated across nearly every field, primarily as a collaborative tool rather than an autonomous worker.

Discuss