CoverStory:WhenEmployeesLeave,TheirAIClonesCarryonWorking

Nearly a year after Li Yao resigned, her voice was still clocking in at her old job.

Former colleagues tipped her off that ads played during internal review meetings still featured her voice. Li, a former video editor who had worked on several voice-over projects, realized her voice had been cloned by artificial intelligence and was being used in the company’s newly created content products.

Li’s experience reflects a fast-spreading change in the modern workplace. Once focused on automating repetitive keystrokes and scripted tasks, companies are now using AI to capture workers’ decision-making logic, communication styles and even personality traits to build “digital employees.”

This trend, rooted in the AI concept of “model distillation,” aims to capture not only what carrying out the task involves, but the specific way a certain person or team achieves it. For employers, it is the ultimate efficiency hack — compressing labor costs by converting an individual’s hard-earned expertise into a permanent, replicable corporate asset.

For employees, however, this represents an existential threat. Often lacking any option to refuse data collection, workers find themselves caught in a paradox: they are being forced to hand over their personal data, operational rhythms and deep-seated knowledge to train the very digital employees that are lined up to replace them.

In response, a grassroots movement of tech-savvy workers has emerged, developing an arsenal of open-source tools to actively subvert corporate data extraction. This technical rebellion is sparking fierce debates over who truly owns a worker’s digital legacy once their employment contract ends.

Cloning human experience

Three years ago, AI disruption was largely seen as a threat to basic knowledge work and art creation. Now, it has encroached on individual work processes, judgment and personality.

An employee’s value has long rested in part on abilities that are hard to quantify: intuition, the ability to spot anomalies and experience-based decision-making in complex situations. Today, companies increasingly want workers to translate those instincts into prompts, workflows and operational pathways that AI systems can adopt. The aim is continuity: even if an employee’s performance fluctuates or they leave, their skill set remains embedded in the system and the AI agent keeps running.

For many workers, training their AI replacement is now an implicit part of the job.

Xu Kejia, a Stanford University student who interned at a Chinese tech giant, writing animation scripts, found that her real task was not writing, but teaching AI to write with a more human feel. It was often faster to write the scripts herself, she said. But the company didn’t need creative humans; it needed people who could translate that creativity into processes the AI model could adopt.

“When it becomes increasingly mature, perhaps the first to become obsolete will be those who trained it,” Xu said.

This trend became crystallized by the sudden popularity of colleague.skill. Developed in four hours by Zhou Tianyi, a 24-year-old engineer at the Shanghai Artificial Intelligence Research Institute, the open-source AI agent ingests a worker’s digital footprint. Once fed chat logs, work documents and behavioral descriptions, it can mimic that employee’s operating habits and business logic to write reports, execute workflows and review code.

In a May preprint paper, Zhou’s team said its aim was to automate AI skill generation through the distillation of expert knowledge. The tool extracts portable skill packages from digital traces, separating a user’s “capability track” from their “persona track.” Zhou said his intention was to preserve team knowledge, but the project revealed a practical path toward “employee digital clones,” igniting controversy over the ethics of exploiting human skill sets.

In some workplaces, the practice is already routine. A software engineer at a Seattle-based cloud provider told Caixin that instead of relying on humans to review new code for vulnerabilities, the team’s collective experience and security protocols had been encoded into AI skills to conduct standardized initial checks.

“It’s not distilling a specific person but rather encapsulating the collective experience formed by the team,” the engineer said, adding that the practice is widely seen in the tech industry as a necessary boost to efficiency.

The push to capture workplace behavior isn’t limited to China. In April, Meta launched an initiative to collect employees’ keystrokes, mouse clicks and screen context to train AI agents, with Mark Zuckerberg, Meta’s chief executive, saying the company wanted to “learn how smart people work.” The move caused an internal backlash, with employees calling Meta an “employee data extraction factory,” particularly as the move coincided with mass layoffs.

In April, Meta launched an initiative to collect employees’ keystrokes, mouse clicks and screen context to train AI agents

Executives are not exempt from this trend. Yang Fangxian of 53AI said the logic of executive decision-making may be even easier to extract. A January 2026 Gartner report similarly noted that “digital twins” are being developed to replicate high-performing chief executives.

The demand for digital replication has moved beyond the workplace. In April, shortly after the death of renowned college-admissions adviser Zhang Xuefeng, a project called Zhang Xuefeng.skill appeared online, with the aim of distilling his books and interviews into an AI system capable of advising students in his individualistic style.

Fighting magic with magic

The corporate push for skill distillation has triggered a technical counteroffensive.

Within a week of colleague.skill going viral, a counter-tool called anti-distillation.skill appeared. Created by Deng Xiaoxian, a developer with a legal background, the tool allows employees to “wash” knowledge documents when managers tell them to submit material for AI training. It replaces core insights with “correct but useless nonsense,” producing a sanitized version for the company while preserving a private backup containing the real expertise.

The tool attracted more than 4 million views within four days of its release on GitHub. Deng said her aim was not to oppose AI, but to resist its malicious use. She described the tool as a defensive measure.

“If someone points a sharp weapon at you, shouldn’t you be able to forge one to protect yourself?” she asked.

The strategy has been described as “fighting magic with magic.” AI researcher Lu Cheng created a similar tool, keep-a-hand.skill. Rather than simply hiding information, it helps workers identify and preserve the skills that remain hardest to replace — such as critical judgment and creativity — while surrendering standardized procedures to the company.

Lu said he wanted to remind workers not to give up their judgment voluntarily.

“We cannot completely stop the development of AI,” he said. “The more important question is how humans can preserve their truly irreplaceable parts in this trend.”

The rights dilemma

Chinese courts have begun to establish that AI-driven technological upgrades do not automatically justify unilateral changes to employment arrangements. But defending workers against digital cloning and data extraction remains legally fraught.

Uncovering AI infringement still depends heavily on manual discovery, making civil litigation lengthy and expensive, said Yu Zehui, senior counsel at Beijing Xstar Legal Law Firm. The few plaintiffs who prevail are usually celebrities, and even then, compensation often falls far short of their legal costs.

That leaves ordinary workers with few practical options.

Since discovering her cloned voice, Li Yao has spent evenings scrolling through short-video ads, searching for content that sounds like her so she can take screenshots, record evidence and report it to platforms. “Most of the time, I can only rely on scrolling to find them myself, collecting evidence bit by bit.” She described the process as a grinding war of attrition: deleted ads are quickly replaced by new versions.

Li’s case highlights how far the legal framework is lagging behind the technology. China’s 2023 rules on generative AI require lawful sources for training data and personal consent, but legal experts say a gap remains around the property rights attached to extracted labor data.

Even if an employment contract states that “work products belong to the company,” Yu said, that does not automatically give the employer the right to use an employee’s personal information or personality traits indefinitely for AI training.

Extracting behavioral and decision-making logic blurs the line between work commitments and personal rights, said Wang Tianyu, deputy director of the Social Law Department at the Institute of Law under the Chinese Academy of Social Sciences. Under China’s Civil Code, if an AI-generated voice, linguistic style or image is recognizable as a specific person to the public, it may amount to digital exploitation of identity and an infringement of personality rights.

The data-collection methods themselves are also legally fraught. Tools such as colleague.skill rely heavily on chat logs, internal documents and emails — records that may contain personal or sensitive information. Even if the data is de-identified, China’s Personal Information Protection Law requires employee-data processing to be strictly “necessary for human-resources management,” such as fulfilling contracts or processing payroll. Whether using a former employee’s historical data to train a “digital clone” meets that standard remains a matter of intense debate.

The biggest obstacle is evidence.

“Many companies will not explicitly admit that they used a specific employee’s data to train a model,” Yu said. “It is very hard to catch a company ‘distilling’ you, and equally hard to prove that your layoff was related to that distillation.”

Once employees leave a company, they lose access to internal systems, making it difficult to prove their historical data was used to train a corporate model. AI’s black-box nature deepens the imbalance, forcing plaintiffs to infer the theft of intellectual property or privacy infringement largely from a machine’s final output.

Seeking data dividends

To address the imbalance, Wang supports exploring a “data dividends” mechanism that would allow workers to share in the value generated by their extracted skills. He said labor law needs to move from “protecting jobs” to “protecting labor behaviors,” ensuring that each instance of collaboration between human and machine is recorded as a protectable and compensable asset.

“Every time AI copies a worker’s skills and experience, it essentially increases the value of the company’s data assets while potentially weakening the worker’s future employability,” Wang said.

For many workers, resistance reflects a broader desire to preserve humanity in an increasingly automated world. Lu Shiyu, a senior researcher at the Tencent Research Institute, said the anti-distillation movement is fundamentally a rejection of the “flattening of people” — a reminder that workers are more than the sum of their extractable skills.

Sun Liping, a sociology professor at Tsinghua University, recently wrote that AI should not replace humans in doing human things, but should instead take on tasks humans cannot or do not want to do.

For younger workers such as Xu Ran, an intern at a tech giant, the human element remains essential. When handing over her required AI skill projects at the end of her internship, Xu left a quiet plea in her final handover document:

“I hope that when we use technology, we never abandon the pursuit of whether it ultimately returns to serving humanity.”

Lu Cheng, Li Yao, Xu Kejia, and Xu Ran are pseudonyms

Contact reporter Han Wei (weihan@caixin.com)