"AI colleagues" are driving people to depression

“On days when everything goes smoothly, I can't help but think that nothing I do matters. Everything is automated, better and faster than I could ever be in the future. But on days when everything falls apart, I don't know why things go wrong. That's when I realize I have no idea what I'm doing.”

This passage doesn't come from a laid - off programmer or a customer service representative replaced by AI. Instead, it's from an employee of Anthropic, one of the world's most cutting - edge AI companies, the creator of Claude, an organization gearing up for an IPO at a valuation of $965 billion.

They created AI, and then AI started making them doubt the meaning of their work.

Anthropic released the AI employee Claude Tag | Image source: Anthropic

On May 23rd local time, Anthropic launched a new feature called Claude Tag, an “AI employee” within the collaboration software Slack. In “ambient mode,” Claude continuously monitors channel conversations and actively speaks up when it determines intervention is needed—summarizing a discussion, reminding of an overlooked detail, or retrieving the information you need from other parts of the company.

Needless to say, many office software applications have already defaulted to providing “AI employees” to help employees “passively” experience the charm of AI collaboration.

However, beneath the seemingly “efficient” surface, everyone working with AI has, without exception, felt “loneliness and frustration.”

Why is this?

The Lonely Engineers

On June 21st, Fiona Fung, the engineering lead of the Anthropic Claude Code and Cowork team, said something on Lenny's Podcast, which was later reposted by multiple tech media outlets.

She said that after the team used Claude Code extensively, they discovered an unexpected problem—the engineers were becoming lonely. Everyone was collaborating with their AI agents, and direct communication among team members decreased.

This isn't just a general feeling. Fung's team took specific intervention measures—organizing pair - programming lunches, hackathons, and scheduling “co - creation time” to try to restore the sense of human - to - human collaboration. In other words, an AI company had to specifically design activities to counteract the erosion of its product on the team's social structure.

To understand the source of this loneliness, we need to look at a set of numbers.

On June 4th, the Anthropic Institute released an internal research report titled “When AI Builds Itself.” The report revealed some previously unpublicized data—as of May 2026, over 80% of the merged code in Anthropic's codebase was written by Claude. Before Claude Code was launched in early 2025, this figure was in the single digits. Meanwhile, the amount of code merged by each engineer per day was eight times that of 2024.

The report also included a statement from another employee: they hadn't written any code by hand for about five months.

Boris Cherny, the creator of Claude Code, put it more bluntly—he hadn't written a single line of code by hand for eight months. On some days, he managed thousands or even tens of thousands of AI agents simultaneously.

He was no longer a code - writer but a dispatcher of AI agents.

In an internal survey at Anthropic, the median estimate of 130 researchers was that their output after using AI was about four times what it was before. The company proudly pointed out that Claude fixed over 800 API errors in April 2026, a task that would have taken humans four years.

However, the report's wording also acknowledged a fact—there is still a “huge performance gap” between Claude and humans in the most open - ended and judgment - required tasks. In a research test called “Next - step judgment,” the latest model of Claude, Mythos Preview, selected the next - step solution with a 64% probability of being better than the actual choice of human researchers. In November 2025, this figure was 51%.

Putting these numbers together, the picture is this—human employees are still needed, but you're becoming increasingly unsure why.

It's not that AI has replaced you. It's that AI is doing so well that you start to doubt the reason for “being here.” Your code is written by AI, your judgment is being approached by AI's judgment, and your role has changed from “creator” to “approver,” and you don't even know if what you're approving is beyond your understanding.

The second half of the anonymous employee's statement accurately describes this state: “On days when everything falls apart, I don't know why things go wrong. That's when I realize I have no idea what I'm doing.”

This isn't the fear of unemployment. It's something deeper—a fundamental loss of self - ability and self - value.

What makes the whole thing even more disturbing is that these people aren't marginal employees. They are engineers and researchers at Anthropic, at the forefront of the AI capability curve.

If they're already having these feelings, what will happen to ordinary knowledge workers with far less technical ability when AI enters ordinary enterprises in the form of Claude Tag, digital employees, and Slack resident agents?

Meta AI's “Gulag”

If Anthropic's problem is “AI is so useful that people lose their sense of existence,” then Meta's problem is its mirror image:

People are degraded to fuel for AI.

In March 2026, Meta established a new department called Applied AI, specifically responsible for improving the company's generative AI models. Approximately 6,500 engineers and product managers were transferred to this department—but for many, this wasn't a promotion or a job transfer; it was a forced conscription. An internal memo from the department head clearly stated that the transfer was not optional.

Those transferred began to call themselves “draftees.”

They found that the work they were assigned was data annotation and RLHF (Reinforcement Learning from Human Feedback)—basic work for training AI models, repetitive and trivial, completely different from their previous software engineering positions. According to an estimate by The Pragmatic Engineer, about one in every five to six Meta engineers is doing data annotation full - time.

Some employees described the job to WIRED as “soul - crushing.” Some even compared it to a “gulag.”

Reports from WIRED and Business Insider revealed a rather bleak picture—employees' roles were blurred, career development paths were unclear, and the management hierarchy was chaotic (some managers had 50 direct subordinates). All of this happened after Meta laid off about 8,000 employees (10% of its global workforce) in May 2026.

Ironically, Meta's net profit in that quarter was as high as $26.8 billion.

CTO Andrew Bosworth admitted at an internal employee meeting in early June that morale was “probably the worst or close to the worst I've seen in this company in 20 years,” comparable to the Cambridge Analytica scandal. Then, in an internal memo, he wrote: “We've done a very poor job of explaining our vision. We've shaken your trust that your professional skills will be valued, your career will develop, and you can truly make an impact.”

Chief Product Officer Chris Cox used a more vivid metaphor—“running a marathon in a hailstorm.”

Bosworth promised to limit the number of direct subordinates of managers to about 20, reduce management changes during the restructuring, and increase the budget for travel, team - building, and snacks. Some analysts noticed this detail—responding to an existential crisis by improving snacks itself shows a certain disconnect.

Looking at Meta's case and Anthropic's case together, a complete picture emerges—

At Anthropic, humans and AI are in a parallel relationship. As AI gets stronger, people feel redundant.

At Meta, humans and AI are in a feeding relationship. People are degraded to parts on the AI training assembly line.

The two paths seem opposite, but they lead to the same end—people's sense of value is undermined.

Moreover, Meta's situation reveals a colder truth. In the AI era, not only “being replaced by AI” can hurt people, but “serving AI” can too. You don't lose your job; you even get an “AI - related position”—but the essence of this job is to turn you from an engineer into an annotator, from a creator into a feeder. Your skills, your judgment, and the engineering intuition you've accumulated over decades are hardly used in the new position.

Mark Zuckerberg later also admitted in a memo that the changes “caused distress” and promised not to conduct company - wide layoffs in 2026. But the damage has already been done. It's reported that some Meta employees even secretly hoped to be laid off—because the severance package included 16 weeks of pay and 18 months of medical insurance, which was more attractive than staying in a position where they felt hopeless.

When an employee starts hoping to be fired instead of hoping to be retained, something in the system has broken.

The Counseling Room

Taking a step back from Anthropic and Meta, you'll find that this isn't an internal problem of two companies but a systemic phenomenon spreading across the entire tech industry.

Psychotherapists in San Francisco were the first to feel the change.

In April 2026, an SF Standard report interviewed several therapists in the Silicon Valley area. They said that the demand for therapy among tech industry practitioners was increasing significantly, and this time, the existential despair was deeper than ever.

Psychotherapist Candice Thompson said something that I'll never forget—previously, if someone walked into the consulting room and said “it's the end of the world,” it was obviously a statement that required clinical intervention. But now, the fears described by the visitors are real - world concerns that therapists themselves have to take seriously.

Another therapist observed that many patients' anxiety doesn't come from the direct threat of “being replaced” but from a more complex rift—they're worried that the technology they're building might harm humanity, and at the same time, they're uneasy about whether their company is paying enough attention and imposing enough constraints.

One therapist summed it up as “people are under a lot of pressure about where this ship is going.”

But most people didn't quit. They chose to stay within the system and try to exert influence from the inside. This in itself is a tired stance.

Macro - data confirms these clinical observations.

The Gallup 2026 Global Workplace Report shows that global employee engagement has dropped to 20%, the lowest level since 2020 and the second consecutive year of decline. In the United States, only about 30% of full - time and part - time employees said they were engaged in their work, a ten - year low. More notably, the engagement of management has dropped by 9 percentage points since 2022, from 31% to 22%. Gallup estimates that low engagement costs the global economy about $10 trillion in productivity losses each year.

A report released by ADP Research in March 2026, covering more than 39,000 workers in 36 countries, was even more straightforward—only 22% of workers globally strongly agreed that their jobs would not be phased out. Among front - line employees, this figure was only 18%.

The wave of layoffs in the tech industry continues. Since the beginning of 2026, nearly 120,000 tech industry employees have been laid off, almost matching the level of the entire year of 2025. Meta attributed its layoff of 8,000 employees to AI, and this is not an isolated case.

Meanwhile, a joint survey in June 2026 found that 90% of American job seekers were worried about the expansion of AI in the workplace—42% were worried about over - reliance on technology, 36% were worried about the reduction of entry - level positions, and 36% were worried that they would lose their problem - solving ability if machines did too much thinking work.

On the management side, 81% of hiring managers believe that AI will improve efficiency, and 79% believe that it can free up employees' time.

This disconnect is the problem itself—the management sees the efficiency curve, while the employees feel the threat signal.

The same technology is interpreted as completely opposite things on different floors of the same company.

The Ghost of Lordstown

At this point, a question naturally arises—is this really something new?

The history of humans working with machines is much longer than that of AI. In the manufacturing industry, workers have been working side by side with robotic arms, assembly lines, and industrial robots for more than half a century. Have they experienced similar psychological shocks?

The answer is yes, and it's very serious.

In 1966, General Motors built the most advanced automobile factory in the United States in Lordstown, Ohio, introducing robotic welding equipment and the fastest production line at that time. By the early 1970s, management laid off 300 workers and increased the production line speed from 60 cars per hour to 101—the fastest in the world. Workers were required to complete assembly actions within 60 seconds, with their arms hanging in the same position to operate springs weighing ten pounds. A mistake could break a finger or crush a wrist.

The workers' response was a full - scale rebellion—slacking off, a soaring absenteeism rate, alcohol and drug use on the job, and large - scale deliberate sabotage. Some workers threw parts into the car body, causing new cars to leave the assembly line with unfinished assembly.

In March 1972, a long - lasting strike broke out.

The media coined a term “Lordstown syndrome” to describe the general dissatisfaction of American workers with the quality and meaning of their work. It's not just about wages or working hours but about a more fundamental question: When the rhythm of the machine determines every movement of your body, are you still a whole person?

Subsequent quantitative research confirmed the depth of this harm. A study by the University of Pittsburgh found that after American workers worked with industrial robots, work - related injuries did decrease—by 1.2 cases per 100 workers. However, the mortality rate related to drugs or alcohol increased significantly, by 37.8 cases per 100,000 people, and the suicide rate and mental health problems also rose.

An interesting comparison is Germany.

The same study found that German workers didn't show significant changes in mental health after being exposed to industrial robots. Researchers speculated that this was related to Germany's stronger labor protection system and social safety net. In other words, the machine itself doesn't necessarily harm people, but in a system that makes people feel insecure, the machine will amplify that insecurity.

In 1974, Harry Braverman, a former metalworker and sociologist, provided a theoretical framework for these phenomena in “Labor and Monopoly Capital.” His core proposition is that the essential tendency of capitalist management is to separate “conception” from “execution”—management monopolizes the power of planning and design, leaving workers only with the mechanical execution part. This separation deprives workers of knowledge and judgment in the labor process, thus degrading work to “almost the level of animal labor.”

Braverman called it “deskilling.”

Half a century later, from industrial robots to AI, what has changed, and what has remained the same?

What hasn't changed is the core question—people lose their sense of meaning in the relationship with machines. The workers in Lordstown felt like an extension of the assembly line, the engineers at Anthropic feel like approvers of AI, and the “draftees” at Meta feel like feeders of AI. Stripping away the technical shell, the underlying psychological harm is surprisingly similar.

But what has changed is even more disturbing:

The injured group has changed. The victims in Lordstown were blue - collar workers—at that time, in the social discourse, their pain was acknowledged but often attributed to “low education levels” or “inability to adapt to technological progress.” Today, the injured are the top - tier knowledge workers in Silicon Valley—Anthropic's researchers, Meta's senior engineers, and people with six - figure salaries. If even they are having existential crises, how long can the narrative of “learning new skills can help you adapt” hold?