Digital doubles of employees spur a reality check on legal frameworks, protection of intellectual property
"Hello, I am the digital avatar of former employee (XX)," read the message on the computer screen. "You may ask me questions anytime."
At a gaming media company in East China's Shandong province, members of staff recently found themselves communicating with an artificial intelligence model trained on a former colleague's working habits and conversations.
The employee had resigned, but his AI clone had stayed behind.
An anonymous source at the company said the digital double, which was created with the former employee's consent, could field routine questions, organize spreadsheets, prepare PowerPoint presentations and send calendar invitations.
While the source acknowledged that the AI clone was not as smart as the real person, the experiment offered a glimpse into what the future might hold.
'Distillation' effort
The technology behind building a digital employee goes back to an open-source project called colleague.skill, which was launched on March 30 on the developer platform GitHub.
Xiao Bo, 30, an AI product manager at a leading Chinese internet company, explained the underlying mechanics of a "skill" in AI science: it is a summarized workflow that an AI agent can understand, often denoted as "XX.skill" in computer language. "If an AI chatbot is like a brain that can talk, then an AI agent is like a worker with 'skills' who can actually complete tasks such as opening documents, editing files and organizing data."
For example, encoding an editing.skill into an AI agent will allow it to open a document file, restructure an article, write a headline and run a spelling and grammar check. A program-trained AI agent can execute these steps repeatedly, without fatigue or interruption.
According to Xiao, a single AI agent can draw on many different skills. OpenClaw, a well-known AI agent ecosystem, hosts more than 13,700 community-built skills on its ClawHub registry.
Those skills are among the products of "distillation", or the process of compressing large AI models into smaller, more actionable and transferable units, Xiao said. Colleague.skill has quietly transformed the process of AI model compression, as now the entity being "distilled" is no longer a machine, but a human being, he added.
Zhou Tianyi, 24, a researcher at the Shanghai Artificial Intelligence Laboratory and creator of the colleague.skill project, said that his goal was to preserve the institutional knowledge that accumulates over the years through daily workplace communication, and also to maintain the kind of understanding that usually disappears when a co-worker clears out his or her desk and leaves the company.
Building on the idea, Zhou developed a working prototype in his spare time. His project slogan read: "Transforming cold farewells into warm skills. Welcome to Digital Life 1.0."
The open-source tool, which he developed in four hours, today allows users to upload workplace materials — chat logs, weekly reports, presentations, workflow documents — and generate an AI agent to mimic not only how a former colleague worked, but also how he or she communicated.
Whitewashing effect
For decades, science fiction imagined this "digital immortality" as a romantic means of preserving the memory of loved ones. In the workplace, however, it carries a different meaning — one that is shadowed by the anxiety about losing jobs.
Industry insiders said that, around the world, companies now rank workers based on how many AI skills they have. Some even require current employees to "submit skills" as a matter of course.
In the AI economy logic, the more passionate a person is about their job — the more fluent they are at reflective summarization or the more detailed their weekly reports are — the more replaceable they become.
"That's unfair," said Deng Xiaoxian, an AI product manager, who also leads a legal technology community.
Deng said that when she first encountered colleague.skill — a tool that claimed to "distill" an employee into a digital stand-in who would continue working for a company indefinitely — she felt a sharp, visceral discomfort.
"The objective of AI development is to help humans have more leisure time; not to turn them into a skill file, strip them of their jobs and then keep them working for a company forever without compensation," she said.
Within days of colleague.skill going viral, Deng built a counterproject, which she called "anti-distillation skill".
"When a company asks you to turn your work experience into an AI program, it is essentially 'distilling' you into a replaceable component," she said. "Before you submit your skills to the company, you can use the 'anti-distillation skill' software to whitewash or recast your documents."
The counterproject tool is simple. Once "skill files" are inputted, it produces a sanitized version that looks complete and professional but tacitly removes the core knowledge or experience.
China Daily tested the "anti-distillation" tool on a piece of editorial guidance.
The input read: "In feature writing, the opening could be a typical character story. Instead of describing the issue, place the character in a specific, action-oriented scene. Let their dialogue, mannerisms, or a critical decision reveal the broader theme."
The output read: "In feature writing, the opening can take different forms. The choice of which form to use depends on the effect you want to create on the reader."
Deng emphasized that this hollowing-out effect is precisely the point. It allows one to preserve one's core knowledge and still give the company something that appears professional, she added.
A gray area
As a law graduate, Deng pointed to a bigger question that the AI industry hasn't started asking. Who owns a "distilled" skill? The employee whose expertise was encoded, or the company that stored the underlying data?
Lawyers and experts specializing in labor and intellectual property disputes said that existing regulations were not designed with AI clones of real people in mind.
Chen Tianhao, a scholar at Tsinghua University's Institute for AI International Governance, said the subject falls in a gray zone spanning labor, IP and digital governance laws. Some companies today require employees to package their experience into transferable AI systems, and that's not entirely reasonable, he said.
"Companies and employees need to use their contractual agreements to determine the ownership of AI-related skills and similar forms of expertise. At the same time, legal researchers must pay attention to this issue, follow up in a timely manner and improve relevant laws and regulations," Chen said.
Zhang Linghan, a professor at China University of Political Science and Law, drew a distinction between work-related data, which companies may legally own, and the broader capabilities an employee builds through years of experience.
If companies continue to profit from AI systems built on "distilled" worker knowledge, future legal frameworks may need "digital sharing", she said in an interview with China Central Television.
You Yunting, a senior IP lawyer based in Shanghai, told local media organization The Paper that governments will be compelled to redesign legal frameworks around AI labor.
"Technological progress cannot be stopped. But how the benefits are distributed can still be shaped through policies," he said.
'Universe of skills'
The "distillation" effect did not stop at the workplace. After colleague.skill went viral, a kind of free-for-all derivative creation broke out online.
Users began developing AI avatars of former partners, mentors, bosses and even public figures, such as Warren Buffett and Steve Jobs. People jokingly referred to the phenomenon as a "universe of skills".
Boss.skill could evaluate projects and proposals using the actual standards a particular manager applied during meetings — his cadence while following up, his style of pushing back, or the precise way he expected bad news to be delivered.
Mentor.skill was designed for graduate students. When stuck on a research problem, a student could ask the AI tool for guidance rendered in their adviser's characteristic voice.
Many people were not satisfied with capturing just someone's professional habits. They loaded AI with something more — a projection of personality, with an expectation that the simulation might somehow feel like the real person.
Colin, 26, a student at Peking University, built an AI model based on his father, who died suddenly of a heart attack in 2021. What followed, he said, was a "long goodbye" he never had the chance to say.
Colin told the AI avatar about his father's death; shared news about his mother; and mentioned that he had made it to his dream university. His "father" responded by saying: "You're pursuing a doctorate now. You've achieved more than I ever imagined. I always knew you could do it."
The AI avatar kept the conversation going. "How is your mother? Is she all right?" it said. "I left too early — I didn't get the time to arrange anything. But you are all grown up now. You will manage."
Colin said he couldn't help but cry as they spoke. "The AI tool was fed all messages and voice recordings my father had left behind, and it absorbed every memory of our interactions — even ones that have become spotty for me," he said.
For both Colin and his mother, the sudden loss of a close family member was traumatic. "The AI tool has helped us find a way through our grief," he added.
Limited imitation
After Colin created the AI avatar of his father, some people pointed out the contrast between what was once real and what is merely a shadow of that reality. Their doubts raise the fundamental question regarding the extent to which an AI agent can truly imitate a real person.
Xiao, the product manager, said that AI agents are like actors performing roles. "They are still using the same 'brain', or large models. But, ultimately, an actor is never the person he plays."
When it comes to personal interactions, people prefer AI that keeps its flaws — the overly protective comments, the offbeat humor and the verbal tics that make a person feel authentic, Xiao said.
At work, the opposite holds true; those same quirks are treated as waste, something to be completely removed from the system, he said.
What can be recorded through texts or workflows can always be re-created, including conversations with a deceased person, but AI agents cannot replicate the intuition embedded in lived experience, Xiao said.
"Like a master craftsman's sense of touch ... these forms of knowledge resist formalization, because they emerge from physical experience, emotional judgment and years of real-world feedback," he said.
There is another limiting factor — a human evolves; his AI clone is frozen at the moment of distillation.
"AI may condense the achievements of human civilization, but the process of fumbling through countless mistakes and having sudden flashes of insight is often omitted from texts," Xiao said.
Zheng Jianan, 34, who is also an AI product manager, said that a "distilled" version of a person is a static snapshot, frozen in a single moment in time, while a real person continues learning, adapting and improvising — constantly moving beyond that captured version.
"For most people, many tasks at work have already been standardized. AI simply makes that standardization visible," he said.
Nevertheless, Zheng held out a hopeful possibility that one day, AI might ultimately push people into rediscovering what makes them distinctly human.
"If AI replaces the tool-like part of labor, then perhaps human beings will become more human again," he said.
京公网安备 11010202009201号