The Invisible Labor of AI AI 背后的隐形劳动

AI is sold as the disappearance of work. A request enters a box, a sentence returns, and the interface invites us to imagine that a machine has replaced the messy coordination of human labor. But the smoother the answer appears, the more important it becomes to ask where the rough work has gone.

Much of it has been moved offstage. Data workers label images, compare answers, rank preference pairs, remove duplicates, and turn ambiguous cultural worlds into categories a model can ingest. Content moderators absorb violent, abusive, sexual, and traumatic material so that the interface can remain clean. Evaluators write prompts, judge outputs, test failure cases, and translate human judgment into a score. Software teams patch the infrastructure, monitor regressions, and keep brittle systems ordinary enough to be trusted. Managers then reorganize the workplace around dashboards, agents, and productivity claims.

This is why the labor question is also a geographic question. The visible AI economy concentrates prestige, salaries, and ownership in a few corporate and metropolitan centers, while repetitive risk is pushed into platform labor markets, outsourced annotation firms, and low-status support roles. Some workers are invited to become AI engineers, product leads, or transformation consultants. Others are asked to clean the data, watch the harms, or adapt their jobs to tools they did not choose.

The pattern is not only global; it is directional. Many of the platform centers that name, sell, and govern AI sit in the global North, while many of the labor nodes that absorb annotation, moderation, evaluation, and support work sit in the global South. Tasks and risks move outward through contracts and platforms. Clean data, safer interfaces, and enterprise value move back. This is the global labor scissors of AI: value becomes visible in one set of places while injury and repetition are made ordinary in another.

China complicates that map. It is not simply a labor node in a global North-South system, nor simply a platform center. Its large AI firms, clouds, content platforms, data rules, local governments, vocational pipelines, and annotation bases form a more state-embedded domestic geography. The scissors gap reappears inside the country: model authority and platform value concentrate in coastal and first-tier cities, while labeling, moderation, testing, customer support, and training work are routed toward inland provinces, county-level industrial parks, and outsourced labor chains.

The World Economic Forum's 2025 jobs report gives the optimistic macro grammar: new jobs, displaced jobs, and a very large reskilling agenda. McKinsey's State of AI surveys show the organizational grammar: the firms reporting more value from AI are also the firms redesigning workflows. AI Now's Artificial Power supplies the political grammar: AI is not only a tool for productivity, but a way to redistribute power over institutions, knowledge, and work.

The danger is that "reskilling" becomes a polite word for making workers absorb the cost of organizational change. Deskilling and reskilling happen together. A worker may lose discretion over a task while gaining the duty to supervise an automated system. A writer may become a prompt editor. A customer-service worker may become a monitor of machine-generated replies. A software engineer may move faster while losing the slower craft through which judgment was learned.

The question, then, is not whether AI will create or destroy jobs in the abstract. It is whether the institutions adopting AI will make the labor behind it visible enough to bargain over. Who gets credit? Who gets paid? Who gets exposed to harm? Who is watched? Who can refuse? A geography of AI labor begins with the answer's hidden hands.

AI 常常被包装成劳动的消失。一个请求进入输入框，一段文字返回，界面让我们相信： 机器已经替代了复杂的人类协作。但回答越顺滑，我们越需要追问：那些粗糙、 重复、需要判断和承受伤害的劳动，被转移到哪里去了？

很多劳动只是被移到了后台。数据工人标注图像、比较回答、排序偏好、清理重复项， 把含混的文化世界转译成模型能够吸收的类别。内容审核员观看暴力、辱骂、色情和创伤性材料， 让前台界面保持干净。模型评估者写提示词、判断输出、测试失败案例， 把人的判断压缩成分数。软件团队修补基础设施、监控退化、排查故障， 让脆弱系统显得足够日常、足够可靠。管理层再围绕 dashboard、agent 和生产率叙事， 重新组织工作场所。

因此，劳动问题也是地理问题。可见的 AI 经济把声望、薪酬和所有权集中在少数企业和大都市中心； 重复性风险则被推向平台劳动力市场、外包标注公司和低地位支持岗位。一部分人被邀请成为 AI 工程师、产品负责人或转型顾问；另一部分人则被要求清洗数据、观看伤害， 或者适应他们并没有选择的工具。

这种模式不只是全球性的，而且是有方向的。许多命名、销售和治理 AI 的平台中心位于全球北方； 许多吸收标注、审核、评估和支持性工作的劳动节点位于全球南方。任务和风险通过合同与平台向外移动； 被清洗的数据、更安全的界面和企业价值再返回平台中心。这就是 AI 劳动的全球剪刀差： 价值在一组地方变得可见，伤害和重复劳动则在另一组地方被日常化。

中国会让这张图变得更复杂。它既不是全球南北体系中单纯的劳动节点，也不是单纯的平台中心。 大模型公司、云基础设施、内容平台、数据治理、地方政府、职业教育和数据标注基地共同组成了一套 更深地嵌入国家与地方发展的国内地理。剪刀差在中国内部重新出现：模型权威和平台价值集中在沿海与一线城市； 标注、审核、测试、客服、训练和实训劳动则被导向内陆省份、县域产业园和外包劳动链条。

World Economic Forum 的 2025 年就业报告给出的是宏观语法：新岗位、被替代的岗位， 以及巨大的再培训任务。McKinsey 的 State of AI 调查给出的是组织语法： 从 AI 中获得更多价值的组织，也更积极地重画工作流程。AI Now 的 Artificial Power 则给出政治语法：AI 不只是生产率工具， 也是重新分配机构、知识和劳动权力的方式。

真正危险的是，"再技能化"变成一个礼貌词，用来要求劳动者自己吸收组织变革的成本。 去技能化和再技能化往往同时发生。一个劳动者可能失去对任务的裁量权， 同时获得监督自动化系统的新义务。写作者变成提示词编辑者；客服人员变成机器回复的监控者； 软件工程师写得更快，却可能失去在慢工中形成判断的过程。

所以问题不是抽象地问 AI 会创造还是毁灭工作，而是问：采用 AI 的机构是否愿意让其背后的劳动 变得足够可见，以至于可以围绕它进行协商。谁获得署名？谁获得报酬？谁暴露在伤害中？ 谁被监控？谁可以拒绝？AI 劳动地理的起点，是回答背后那些被隐藏的手。