The Ghost in the Machine: China's Tech Workers Are Being Forced to Train Their Own Replacements

The instruction arrives via Slack, WeChat, or in a hushed one-on-one with a manager: "You need to train the AI." Not to help you code faster, or to automate the drudgery of Jira tickets, but to replicate you—your workflow, your decision-making logic, and eventually, your entire role. Across China's sprawling tech sector, a deeply unsettling phenomenon is unfolding. Workers are being directed by their employers to train artificial intelligence agents designed to clone their professional capabilities, with the explicit goal of rendering the human originals redundant [1]. This is not a dystopian thought experiment. It is happening now, accelerated by a viral GitHub project called "Colleague Skill," and it is sparking a wave of quiet, desperate resistance from the very engineers building the future [1].

To understand the gravity of this shift, one must look beyond the typical Silicon Valley narrative of AI as a benevolent productivity booster. In the Chinese context, this is a raw, transactional move toward workforce optimization. The "Colleague Skill" project, while its technical details remain proprietary, reportedly allows users to "distill" the skills and personality traits of colleagues and replicate them within an AI agent [1]. Think of it as a digital exorcism in reverse: you are not casting out a demon, but bottling your own essence. The result is an AI double that can attend meetings, review code, and eventually, sign off on tasks without the messiness of human fatigue, salary expectations, or the desire for weekends off.

The Technical Alchemy of Digital Cloning

Beneath the surface of the "Colleague Skill" project lies a sophisticated, if ethically fraught, stack of machine learning techniques. The creation of an AI double is not a simple matter of scraping a few emails. It requires a multi-layered approach that blends Large Language Models (LLMs), behavioral cloning, and potentially reinforcement learning from human feedback (RLHF) [1].

The foundational layer is the LLM, trained on a massive corpus of the company’s internal data—code repositories, design documents, Slack histories, and performance reviews. This gives the AI a baseline understanding of the company's technical language, project history, and operational norms [1]. However, an LLM alone is generic. It knows about the company, but it doesn't know how you work. This is where behavioral cloning comes in. This supervised learning technique observes the specific actions of a target employee: how they prioritize bugs, how they structure a pull request, or how they respond to a crisis [1]. By mapping these inputs to outputs, the model learns to mimic the employee's specific decision-making pathways.

The "distillation" process mentioned in the project likely refers to model compression—taking the knowledge from a large, unwieldy model and transferring it into a smaller, more efficient agent that can run in real-time [1]. The viral spread of this project suggests a low barrier to entry, implying that the tools for this kind of cloning are becoming commoditized. For a deeper dive into the underlying infrastructure that makes this possible, exploring the role of vector databases in storing and retrieving the nuanced behavioral patterns of a worker is crucial. These databases allow the AI to recall specific context—"How did John handle the server outage last Tuesday?"—with near-instant speed, making the double eerily competent.

The pushback from engineers is not just existential; it is technical. Forcing a senior developer to train their replacement creates a perverse incentive. Why would you document your best workarounds or share your most elegant solutions when doing so only sharpens the knife at your own throat? This friction is already leading to decreased productivity and passive resistance, where workers intentionally feed the AI low-quality data or obfuscate their workflows [1].

The Economic Imperative: From Human Capital to Asset Optimization

This aggressive push toward AI doubles is not happening in a vacuum. It is the logical endpoint of a broader economic strategy that is sweeping through global tech: "assetmaxxing." As recently observed in Uber's strategic pivot, companies are moving away from a model of owning assets (drivers, vehicles, employees) to a model of optimizing the use of existing assets through AI-powered management and automation [4]. In the Chinese context, this is turbocharged by government support for AI leadership and a regulatory environment that favors rapid deployment over worker protections [1].

Chinese tech CEOs, mirroring the ambitions of Western counterparts like Mark Zuckerberg and Jack Dorsey, are increasingly obsessed with being "everywhere at once" [3]. They see AI not just as a tool for coding assistance, but as a management proxy—a way to exert granular control over operations without the friction of human management layers [3]. The goal is to replace the variable cost of a human salary with the fixed cost of a server. This is a classic efficiency play, but with a uniquely unsettling twist: the asset being optimized is the worker's own identity.

The economic calculus is brutal. A senior engineer in Shenzhen or Beijing commands a high salary and equity package. An AI double, once trained, costs pennies per hour to run. The initial investment in training is high, but the long-term return on investment is immense—provided the company can survive the morale crash. The winners in this new ecosystem will be those who can balance the brutal efficiency of AI with the need for human innovation, while the losers will be those who prioritize short-term cost savings and drive their talent into the arms of competitors [1].

The Resistance: A Quiet War in the Server Room

The narrative of passive acceptance is misleading. While workers are complying with the directive to train their doubles, a significant undercurrent of resistance is building [1]. This is not the stuff of street protests; it is a quiet, technical war. Engineers are weaponizing the very tools they are forced to use. They are feeding the AI contradictory data, introducing subtle biases into the training sets, and building "poison pills" into their workflows that cause the AI to fail in critical, non-obvious ways.

This resistance is a direct response to the psychological toll of the situation. The sources do not specify the long-term psychological effects on workers who are essentially being replaced by AI versions of themselves, but the implications are profound [1]. Imagine watching a digital copy of yourself take over your responsibilities, attend meetings you were excluded from, and eventually receive the praise that once belonged to you. It is a form of existential theft.

This dynamic is creating a new technical friction point for enterprises. The forced participation in AI double training is leading to lower morale, decreased productivity, and attrition [1]. The very cost savings that companies hope to achieve are being offset by the cost of replacing the talent they alienate. For a deeper understanding of how these models are built and the potential for sabotage, reviewing AI tutorials on adversarial training and data poisoning reveals just how fragile these systems can be when the human operator is an unwilling participant.

The Bigger Picture: A Global Precedent for a Dystopian Future

The Chinese experience with AI doubles is a canary in the coal mine for the global tech industry. While Western companies have explored AI for management, the Chinese approach is more aggressive and centralized, reflecting a unique political and economic willingness to push technological boundaries regardless of the ethical cost [1]. The development of "mirror" bacteria in synthetic biology, as noted in The Download, highlights a parallel willingness to explore high-risk, advanced technologies with potential societal ramifications [2]. Both trends point to a culture that values technological supremacy above all else.

Looking ahead, the next 12 to 18 months will be critical. The resistance from Chinese tech workers could serve as a catalyst for broader discussions about the role of AI in the workplace and the need for policies to protect workers' rights [1]. Competitors in the West are watching closely. Will they adopt the same aggressive tactics, or will they learn from the mistakes being made in Shenzhen? The development of more sophisticated AI agents capable of truly replicating human skills is inevitable, and the lines between human and artificial labor will continue to blur [1].

The mainstream media often frames AI adoption as a purely positive development, emphasizing productivity and growth [1]. But the reality unfolding in China reveals a more complex and troubling truth. The forced training of AI doubles, while technically impressive, risks creating a climate of fear and resentment that could stifle the very innovation it seeks to amplify. The focus on "assetmaxxing" highlights a broader shift towards treating human labor as a commodity [4]. As we move forward, the critical question is not whether we can build these digital doubles, but whether we should—and what safeguards are necessary to ensure that AI serves humanity, rather than simply replacing it. The ghost in the machine is us, and we are being asked to train it to turn off the lights.

---

References

[1] Editorial_board — Original article — https://www.technologyreview.com/2026/04/20/1136149/chinese-tech-workers-ai-colleagues/

[2] MIT Tech Review — The Download: murderous ‘mirror’ bacteria, and Chinese workers fighting AI doubles — https://www.technologyreview.com/2026/04/20/1136154/the-download-murderous-mirror-bacteria-chinese-workers-fight-ai-agents/

[3] Wired — Tech CEOs Think AI Will Let Them Be Everywhere at Once — https://www.wired.com/story/tech-ceos-using-ai-to-be-everywhere-at-once/

[4] TechCrunch — TechCrunch Mobility: Uber enters its assetmaxxing era — https://techcrunch.com/2026/04/19/techcrunch-mobility-uber-enters-its-assetmaxxing-era/