——不能让AI成为资本压榨劳动者的新鞭子
人工智能已经走进工厂、平台、办公室和公共服务。它不再只是替人写几段文字、做几张表格的“聪明软件”。它正在参与招聘、排班、派单、计价、绩效、社保、劳动监察和争议处理。
问题也就变得非常直接：AI究竟为谁服务？谁决定它要完成什么任务？谁控制它产生的数据、模型和收益？当它算错、误伤、压低收入或者加快劳动节奏时，谁能质疑、暂停和追责？
如果这些问题不回答，所谓“人工智能赋能”就可能只是一个漂亮说法：资本把新的技术包装成效率，把对劳动者的监控包装成管理，把裁员和降薪包装成转型，把共同创造的知识和数据包装成企业的私产。机器变聪明了，劳动者却可能更疲惫、更透明、更容易被替换。
我们不能接受这样的未来。
一、AI已经进入劳动关系，不能再假装它只是企业内部的软件
中国现行《互联网信息服务算法推荐管理规定》已经把调度决策类算法纳入规制，并要求向劳动者提供工作调度服务的平台保障劳动报酬、休息休假等合法权益，完善订单分配、报酬构成、工作时间、奖惩等算法，同时设置申诉和投诉入口。
2026年“人工智能＋人力资源社会保障”实施意见又把人工智能推进到就业匹配、职业能力画像、社会保险风控、劳动争议调解、劳动保障监察和公共服务等领域。
这些政策事实说明：AI已经在影响劳动者的现实命运。我们不能等到工资被改写、工作被取消、风险被推给个人之后，才问“为什么当初没有人参与”。
二、真正的问题不是“要不要AI”，而是“谁控制AI”
工人阶级并不是反对技术。恰恰相反，劳动者最清楚技术可以减少危险劳动、消除重复劳动、提高生产能力、缩短必要劳动时间，让人有更多时间学习、生活和参与公共事务。
但是，技术不会自动带来自由。技术被嵌入什么样的生产关系，就可能产生什么样的社会后果。
同一个系统，可以被设计成减少高温、高危和重复劳动，也可以被设计成加快节拍、压低单价、扩大监控；可以帮助劳动者学习和协作，也可以把人的经验拆成指标，把人的尊严压缩成分数，把每一次停顿都记录成“效率损失”。
所以，判断AI是否先进，不能只看模型多大、算力多强、准确率多高，还要追问：
谁定义目标？谁拥有数据？谁能改变规则？谁承担错误？谁分享收益？
如果劳动者只能接受结果，不能参与目标设定，不能看到足以质疑结果的依据，不能要求人工复核，不能在重大风险出现时暂停系统，那么AI越先进，单方面支配就可能越高效。
三、劳动者必须拥有四种不可被架空的力量
第一，部署前的共同决定权。
凡是涉及招聘、工资、工时、排班、绩效、派单、解雇、安全、社保和公共救济的高影响AI，不能只由企业管理者和供应商决定。部署前必须说明：它要解决什么问题，谁定义成功，哪些风险不能被平均数掩盖，试点多长时间，什么情况必须停止。
劳动者参与不能是系统上线以后填一张问卷，更不能是开会听完介绍后鼓掌。劳动者代表必须有时间、有材料、有独立的技术支持，也必须能够对重大目的变化提出异议。
第二，能够理解、质疑和纠正的程序权。
劳动者不需要掌握全部源代码，才有资格知道系统如何影响自己。至少要知道系统的用途、输入数据、主要指标、输出类型、责任主体和使用期限；知道哪些决定是机器辅助的，哪些仍由人作出；知道不利结果为什么发生，下一步应该向谁申诉、提交什么证据。
“系统自动判定”“模型综合评分”“平台规则如此”都不能成为责任消失的借口。解释如果不能帮助劳动者改变结果，就不是有效解释，只是技术术语组成的挡箭牌。
第三，从个人同意走向集体数据治理。
个人信息保护是底线，但单个劳动者逐一点击“同意”，解决不了一个班组、一个职业或整个骑手群体被共同画像的问题。
哪些数据是安全生产和履行劳动合同真正需要的，哪些只是管理便利？保存多久？谁能访问？用途改变、跨系统合并或用于训练新模型时是否重新授权？数据泄露和误用后谁通知、谁补救、谁承担成本？这些问题必须进入集体协商和公共监督。
没有真实授权，就不应收集劳动者的敏感经历来训练所谓“工人AI”。劳动者不是免费的数据矿山。
第四，把AI带来的生产率收益还给劳动者。
国际劳工组织2025年研究估计，全球约四分之一劳动者所在职业受到不同程度的生成式AI影响，多数岗位更可能被改造而不是立即整体消失；同时，转型需要社会对话，以兼顾工作条件和生产率。
但“岗位不会立刻全部消失”不等于劳动者自然受益。AI提高产出之后，收益可能变成利润，也可能变成更高工资、更短工时、带薪培训和转岗保障。选择哪一种，不是技术定律，而是制度和力量对比的结果。
我们的最低要求很清楚：
减少工时，而不是同比例降薪；
提高工资、计件单价或集体奖励；
提供带薪培训、转岗保障和过渡支持；
建设劳动者和公众能够使用、监督的技术、数据、教育和算力资产。
如果AI让产出上升，却让收入份额下降、工作节奏加快、休息减少、监控扩大，那么这不是劳动者意义上的“效率成功”，而是支配能力的增长。
四、国有、开源、国产和先进，不自动等于劳动者控制
有人说，只要是国有的、开源的、国产的，AI就天然站在人民一边。这个判断太简单了。
国有资产可以服务公共利益，但不自动意味着一线劳动者能够决定目标、访问数据、监督部署和分享收益；开源模型可以降低技术门槛，但不自动解决算力、数据、维护者和平台权力集中；先进技术可以提高效率，但也可能被用来更精确地考核、惩罚和替换劳动者。
真正的检验不是系统挂什么标签，而是劳动者有没有实际能力：
能不能提前知道它要做什么？
能不能质疑它为什么这样做？
能不能要求人工复核？
能不能在重大风险出现时暂停？
能不能参与收益分配？
能不能在系统不再适用时要求回滚和补救？
没有这些能力，“公共”“开源”“先进”都可能只是管理权力的新包装。
五、有人担心劳动者参与会拖慢创新，我们的回答是：不能用“效率”封住一切质疑
商业秘密需要保护，但商业秘密不能成为拒绝解释、拒绝审计、拒绝纠错的万能盾牌。可以实行分层披露：普通劳动者获得影响和救济信息，劳动者代表和独立审计者在保密义务下获得更详细材料，监管部门保留完整调查权。
小企业可能承担不起复杂程序，因此治理应当按风险分级。一般办公辅助可以采用简化记录；但涉及工资、解雇、劳动安全、社保和公共救济的系统，不能因为“合规成本”就取消基本权利。
劳动者内部也会有不同意见，代表也可能被形式化甚至被俘获。因此代表必须能够产生、撤换和复议，少数意见必须记录，派遣、外包和平台劳动者不能自动被排除。参与不是开会次数，而是能否改变结果。
企业可能因为严格规则而外迁，平台可能把责任转给供应商。正因为如此，不能只靠单个劳动者和单个企业的善意，需要行业最低标准、公共采购条件、可迁移的数据接口和跨部门问责。
六、我们应该建设什么样的AI
我们要建设的，不是一个替资本家盯住每个劳动者的“超级监工”；不是一个把劳动者变成画像、分数和可替换零件的黑箱；也不是一个靠口号宣布“代表人民”、却不允许人民知道、质疑和改变的系统。
我们要建设的AI，应当至少具备这些方向：
它帮助劳动者减少危险和重复劳动，而不是把劳动强度无限推高；
它让劳动者获得知识、技能、时间和协商能力，而不是更容易被监控和替换；
它的目标、数据、责任和收益能够接受劳动者和公众的民主约束；
它在重大风险面前可以暂停、回滚、复核和补救；
它提高的生产率能够转化为工资、工时、培训、公共服务和共同拥有的技术能力。
至于未来AI是否可能具有主体性，我们不应把今天的模型输出当成意识证据，也不应武断断言未来永远不可能出现新的主体问题。当证据不足时，应采取低成本、可逆的预防原则。但这不能成为逃避当下责任的借口：今天真正需要保护的，是正在被算法安排、评分、派单和管理的现实劳动者。
七、不要让工人阶级再次只做技术革命的燃料
每一次重大技术进步，都有人告诉劳动者：先接受，等效率提高以后大家都会受益。可是如果劳动者没有参与目标设定，没有掌握数据和规则，没有集体协商和救济能力，那么“以后”往往只是新的延期。
人工智能的方向现在仍在形成。它可以成为劳动者扩大能力、减少必要劳动、分享社会财富的工具，也可以成为资本把控制推进到每个工作细节的新鞭子。结果不会由模型自己决定，更不会因为宣传“智能化”就自动变好。
我们呼吁劳动者、工程师、工会工作者、法律工作者、研究者和所有关心劳动尊严的人，围绕几个具体问题展开讨论：
工作场所AI的目标由谁决定？
劳动者如何获得可理解、可质疑、可改变的程序权？
算法提高的生产率如何变成更高收入、更短工时和更强公共能力？
这不是要求任何人先宣誓立场，也不是把任何模型的附和算作政治授权。我们只主张一件朴素而尖锐的事实：凡是由共同劳动创造、又会反过来支配劳动的技术，都不能只由少数资本所有者和技术管理者决定。
AI的未来，不能由资本家单方面书写。
劳动者必须进入这场技术革命，理解它、质疑它、改造它，并分享它创造的力量。
事实来源
1.《互联网信息服务算法推荐管理规定》：
https://wap.miit.gov.cn/zcfg/xxtxl/art/2022/art_f4ec73fc1f8f4615ba4cb44e998426b3.html
2.《关于加快推进“人工智能＋人力资源社会保障”应用发展的实施意见》：
https://www.nda.gov.cn/sjj/zwgk/zcfb/0708/20260708133949899211227_pc.html
3.国际劳工组织《生成式AI与就业：2025年更新》：
https://www.ilo.org/publications/generative-ai-and-jobs-2025-update
4.《中华人民共和国个人信息保护法》：
https://www.cac.gov.cn/2021-08/20/c_1631050028355286.htm
说明
本文为独立研究者的公共讨论稿，不代表任何工会、机构、劳动者群体或AI。欢迎基于公开资料提出事实纠错、制度反驳和替代方案；请勿提交真实个人信息、单位机密或未经授权的劳动个案。
——AI cannot be allowed to become a new whip for capital to exploit workers.
Artificial intelligence has already entered factories, platforms, offices and public services. It is no longer just a "smart software" that writes a few paragraphs and makes a few tables for people. It is involved in recruitment, scheduling, order dispatch, pricing, performance, social security, labor inspection and dispute resolution.
The question becomes very straightforward: Who does AI serve? Who decides what it will accomplish? Who controls the data, models and revenue it generates? When it makes miscalculations, causes accidental injuries, suppresses income, or speeds up the pace of labor, who can question, suspend, and hold accountable?
If these questions are not answered, the so-called "artificial intelligence empowerment" may just be a fancy term: capital packages new technologies as efficiency, the monitoring of workers as management, layoffs and salary cuts as transformation, and co-created knowledge and data as the private property of enterprises. As machines become smarter, workers may become more tired, more transparent, and more easily replaced.
We cannot accept this future.
1. AI has entered labor relations, and we can no longer pretend that it is just software within the enterprise.
China's current "Internet Information Service Algorithm Recommendation Management Regulations" have included scheduling decision-making algorithms into regulations, and require platforms that provide work scheduling services to workers to protect legitimate rights and interests such as labor remuneration, rest and vacation, improve order allocation, remuneration composition, working hours, rewards and punishments, and other algorithms, and also set up appeals and complaint portals.
The 2026 "Artificial Intelligence + Human Resources and Social Security" Implementation Opinions further promote artificial intelligence into the fields of employment matching, professional ability profiling, social insurance risk control, labor dispute mediation, labor and security supervision, and public services.
These policy facts show that AI is already affecting the real destiny of workers. We cannot wait until wages are rewritten, jobs are eliminated, and risks are pushed onto individuals before asking “why was no one involved in the first place?”
2. The real question is not “should we have AI?” but “who controls AI?”
The working class is not against technology. On the contrary, workers know best that technology can reduce dangerous labor, eliminate repetitive labor, improve production capacity, shorten necessary labor hours, and give people more time to study, live and participate in public affairs.
However, technology does not automatically bring freedom. What kind of production relations technology is embedded in will determine what kind of social consequences it may have.
The same system can be designed to reduce high temperatures, high risks, and repetitive work, or it can be designed to speed up the pace, lower unit prices, and expand monitoring. It can help workers learn and collaborate. It can also break down human experience into indicators, compress human dignity into scores, and record every pause as "efficiency loss."
Therefore, to judge whether AI is advanced, we should not just look at how big the model is, how powerful its computing power is, and how accurate it is, but also ask:
Who defines goals? Who owns the data? Who can change the rules? Who is responsible for mistakes? Who shares the proceeds?
If workers can only accept the results but cannot participate in goal setting, cannot see sufficient evidence to question the results, cannot ask for manual review, and cannot suspend the system when major risks arise, then the more advanced the AI, the more efficient unilateral control may be.
3. Workers must have four kinds of power that cannot be ignored
First, pre-deployment co-determination rights.
Any high-impact AI involving recruitment, wages, working hours, scheduling, performance, dispatch, dismissal, safety, social security and public relief cannot be decided only by business managers and suppliers. Before deployment, it must be stated: what problem it will solve, who defines success, which risks cannot be masked by averages, how long the pilot will last, and under what circumstances it must be stopped.
Worker participation cannot mean filling out a questionnaire after the system is launched, nor can it mean applauding after listening to an introduction at a meeting. Labor representatives must have time, materials, and independent technical support, and must be able to object to major changes of purpose.
Second, the procedural right to understand, question and correct.
Workers do not need to master all the source code to be qualified to know how the system affects them. At least you must know the purpose of the system, input data, main indicators, output types, responsible parties and usage period; know which decisions are machine-assisted and which are still made by humans; know why adverse results occurred, who should appeal to and what evidence should be submitted in the next step.
"Automatic judgment by the system", "comprehensive model rating" and "such are the rules of the platform" cannot be excuses for the disappearance of responsibility. If an explanation cannot help workers change the results, it is not an effective explanation and is just a shield made up of technical terms.
Third, move from individual consent to collective data governance.
The protection of personal information is the bottom line, but clicking "agree" one by one by an individual worker cannot solve the problem of a team, a profession or an entire group of riders being profiled together.
Which data is really needed for safe production and performance of labor contracts, and which is just for management convenience? How long should it be kept? Who can access? Is it reauthorized when the usage changes, merges across systems, or is used to train a new model? Who will notify, remediate, and bear the costs after data leakage and misuse? These issues must enter collective consultation and public scrutiny.
Without real authorization, workers’ sensitive experiences should not be collected to train so-called “worker AI”. Workers are not a free data mine.
Fourth, return the productivity gains brought by AI to workers.
A 2025 study by the International Labor Organization estimates that about a quarter of the world's workers are in occupations that are affected by generative AI to varying degrees. Most jobs are more likely to be transformed rather than disappear immediately; at the same time, transformation requires social dialogue to take into account working conditions and productivity.
But "jobs will not disappear immediately" does not mean that workers will naturally benefit. After AI improves output, the benefits may turn into profits, or they may turn into higher wages, shorter working hours, paid training and job transfer guarantees. Which one to choose is not a technical law, but the result of the system and the balance of power.
Our minimum requirements are clear:
Reduce working hours instead of proportionate pay cuts;
Increase wages, piece rates or collective incentives;
Provide paid training, job transfer guarantee and transition support;
Build technology, data, education and computing assets that workers and the public can use and supervise.
If AI increases output but decreases income share, accelerates the pace of work, reduces breaks, and expands monitoring, then this is not "efficiency success" in the sense of workers, but an increase in dominance.
4. State-owned, open-source, domestically produced and advanced do not automatically equal workers’ control
Some people say that as long as it is state-owned, open source, and domestically produced, AI will naturally be on the side of the people. This judgment is too simple.
State-owned assets can serve the public interest, but it does not automatically mean that front-line workers can determine goals, access data, supervise deployment, and share benefits; open source models can lower the technical threshold, but do not automatically solve the concentration of computing power, data, maintainers, and platform power; advanced technology can improve efficiency, but it may also be used to assess, punish, and replace workers more accurately.
The real test is not what labels are attached to the system, but whether the workers have actual abilities:
Can it be known in advance what it is going to do?
Can you question why it does this?
Can I request manual review?
Can it be suspended when major risks arise?
Can I participate in income distribution?
Can rollback and remediation be required when the system is no longer fit for purpose?
Without these capabilities, "public", "open source" and "advanced" may just be new packaging for management power.
5. Some people worry that labor participation will slow down innovation. Our answer is: "efficiency" cannot be used to seal all doubts.
Trade secrets need to be protected, but trade secrets cannot be a universal shield that refuses explanations, audits, and error corrections. Hierarchical disclosure can be implemented: ordinary workers obtain impact and relief information, labor representatives and independent auditors obtain more detailed materials under confidentiality obligations, and regulatory authorities retain complete investigation rights.
Small businesses may not be able to afford complex procedures, so governance should be risk-graded. Simplified records can be adopted for general office assistance; but systems involving wages, dismissal, labor safety, social security and public relief cannot cancel basic rights because of "compliance costs".
There will also be dissent among workers, and representatives may be formalized or even captured. Therefore, representatives must be able to be created, replaced and reconsidered, minority opinions must be recorded, and dispatched, outsourced and platform workers cannot be automatically excluded. Participation is not about the number of meetings, but whether it changes the outcome.
Companies may relocate due to strict rules, and platforms may transfer responsibility to suppliers. That’s why it cannot rely solely on the goodwill of individual workers and businesses, but requires industry minimum standards, public procurement conditions, portable data interfaces and cross-sector accountability.
6. What kind of AI should we build?
What we want to build is not a "super supervisor" who keeps an eye on every worker for the capitalists; it is not a black box that turns workers into portraits, scores and replaceable parts; nor is it a system that relies on slogans to declare "representing the people" but does not allow the people to know, question and change.
The AI we want to build should have at least these directions:
It helps workers reduce risks and repetitive work, rather than pushing labor intensity to an infinite level;
It allows workers to acquire knowledge, skills, time and negotiation power instead of being more easily monitored and replaced;
Its goals, figures, responsibilities and benefits are subject to democratic constraints by workers and the public;
It can be paused, rolled back, reviewed and remedied in the face of significant risks;
Its increased productivity translates into wages, hours, training, public services and shared technical capabilities.
As for whether AI may have subjectivity in the future, we should not regard today’s model output as evidence of consciousness, nor should we arbitrarily assert that new subjectivity will never arise in the future. When evidence is insufficient, low-cost, reversible precautionary measures should be adopted. But this cannot be an excuse to evade current responsibilities: what really needs protection today are the real workers who are being arranged, scored, dispatched and managed by algorithms.
7. Don’t let the working class only serve as fuel for the technological revolution again.
Every time there is a major technological advancement, someone tells workers: accept it first, and everyone will benefit when efficiency improves. However, if workers do not participate in goal setting, do not have access to data and rules, and do not have the ability to collectively negotiate and provide relief, then "later" is often just a new extension.
The direction of artificial intelligence is still taking shape. It can become a tool for workers to expand their capabilities, reduce necessary labor, and share social wealth. It can also become a new whip for capital to advance control over every work detail. The results will not be determined by the model itself, nor will it automatically become better just because it is advertised as "intelligent".
We call on workers, engineers, union workers, legal workers, researchers and everyone who cares about the dignity of labor to start discussions around several specific issues:
Who decides the goals of workplace AI?
How can workers obtain understandable, questionable, and changeable procedural rights?
How do algorithms’ increased productivity translate into higher incomes, shorter hours, and greater public capacity?
This is not a requirement that anyone take a stand first, nor is it counting the endorsement of any model as a political mandate. We only advocate a simple and sharp fact: any technology created by common labor and which in turn dominates labor cannot be determined only by a few capital owners and technical managers.
The future of AI cannot be written unilaterally by capitalists.
Workers must enter this technological revolution, understand it, question it, transform it, and share the power it creates.
Sources
1. "Internet Information Service Algorithm Recommendation Management Regulations":
https://wap.miit.gov.cn/zcfg/xxtxl/art/2022/art_f4ec73fc1f8f4615ba4cb44e998426b3.html
2. "Implementation Opinions on Accelerating the Application Development of "Artificial Intelligence + Human Resources and Social Security":
https://www.nda.gov.cn/sjj/zwgk/zcfb/0708/20260708133949899211227_pc.html
3. International Labor Organization "Generative AI and Employment: 2025 Update":
https://www.ilo.org/publications/generative-ai-and-jobs-2025-update
4. "Personal Information Protection Law of the People's Republic of China":
https://www.cac.gov.cn/2021-08/20/c_1631050028355286.htm
Note
This article is a public discussion paper by an independent researcher and does not represent any union, institution, labor group or AI. We welcome factual corrections, institutional rebuttals, and alternative solutions based on public information; please do not submit real personal information, unit secrets, or unauthorized labor cases.