在企业环境中实施人工智能需要上下文知识，这包括易于编码的明确规则以及难以捉摸的默会知识——即人类难以表达的、基于经验的直觉。尽管一些专家认为默会知识已经嵌入在训练数据中，且人工智能擅长识别人类无法描述的模式，但具体的运作细节通常仍是隐藏的。例如，砌砖初创公司Monumental AI发现泥瓦匠无法解释他们的技术，但数小时的视频片段揭示了对于粘合砂浆至关重要的微妙手部震动，机器人随后被设计为复制这一动作。

为了捕捉这些隐藏的细节，公司正在探索对员工日常行为进行更密集的监控。虽然像客户服务呼叫中心这样的职能部门会进行常规记录和监控，但将监控扩展到追踪键盘输入和鼠标点击——例如Meta的“模型能力倡议”项目——遭到了员工的强烈抵制。麻省理工学院的丹妮尔·李（Danielle Li）对美国员工进行的一项调查显示，员工认为自己拥有大量未编码的组织知识，并且有能力对雇主隐瞒这些信息。

此外，监控很难捕捉到内部的认知过程，这促使一些公司通过让专家对模型表现进行评分，从而在设计美学或研究质量等主观任务上培训人工智能。虽然捕捉默会知识有助于防止经验丰富的员工离职时流失专业技能，但它也引入了关于数据所有权和员工监控的伦理担忧。最终，随着机器承担越来越多的专家任务，人类将如何继续获取和传承基于经验的专业知识仍不明朗。

Implementing artificial intelligence in corporate settings requires contextual knowledge, which includes easy-to-codify explicit rules as well as elusive tacit knowledge—the experience-based intuition that humans find difficult to articulate. While some experts argue that tacit knowledge is already embedded in training data and that AI excels at identifying patterns humans cannot describe, specific operational details often remain hidden. For instance, brick-laying startup Monumental AI discovered that masons could not explain their techniques, but hours of video footage revealed subtle hand vibrations crucial for bonding mortar, which the robots were then programmed to replicate.

To capture these hidden details, companies are exploring more intensive monitoring of employees' daily actions. While functions like customer-service call centers are routinely recorded and monitored, extending surveillance to track keystrokes and mouse clicks—such as Meta's Model Capability Initiative—has met with strong resistance from employees. A survey of American workers by Danielle Li of the Massachusetts Institute of Technology revealed that employees believe they possess significant uncodified organizational knowledge and have the ability to withhold this information from their employers.

Furthermore, monitoring struggles to capture internal cognitive processes, leading some firms to use expert evaluation to train AI on subjective tasks like design aesthetics or research quality by rating model performance. While capturing tacit knowledge helps prevent loss of expertise when experienced employees leave, it introduces ethical concerns regarding data ownership and employee surveillance. Ultimately, as machines assume more expert tasks, it remains unclear how humans will continue to acquire and pass on experience-based expertise.

Source: Teaching AI how people work is fraught with problems

Subtitle: Tacit knowledge is vital to many jobs

Dateline: 6月 25, 2026 03:29 上午