What exactly is Bill Gates worried about when the "shrill alarm" blares?

What Bill Gates worries about is not that machines will become human, but that humans will not have enough time to become what they are supposed to be.

Kirkland, Washington, is a wealthy lakeside suburb on the east side of Seattle. The weather on this day is almost unreal in its perfection: in the mid-80s Fahrenheit, the sky is a flawless, unbroken expanse of blue. Looking out from the conference room of Gates Ventures, a row of yachts bobs gently with the lake water along the pier below, and the Olympic Mountains outline the skyline on the other side of the lake. As the journalist from MIT Technology Review puts it, the view is stunningly pleasant, yet somehow oddly unsettling — because while everything in sight is calm, the person sitting across from him is anything but calm.

70-year-old Bill Gates sits on the other side of the conference table, rocking back and forth in his chair, visibly emotional. He talks more and more, ranging from terrorism, economic collapse, all the way to the eventual loss of human control over AI systems. It is the summer of 2026, and the Microsoft co-founder has just published a long open essay titled *The Choices We Make About AI Now Matter More Than Ever*, and accepted an exclusive interview with MIT Technology Review. In his own words, he has decided to act as that "blaring alarm bell".

This positioning itself is sufficiently anomalous. Gates is the most successful beneficiary and standard-bearer of technological innovation over the past 50 years — he was fascinated by making machines intelligent the moment he was exposed to computers at the age of 13, built the Microsoft empire with software, and now uses his tech-derived wealth to fight global poverty and disease. He has almost never stood on the opposite side of technology in his life. At over 70 years old, he suddenly wants to be the person who pulls the fire alarm at a banquet.

So what is really worth asking is not "what did Gates say", but: What exactly is the person who understands technology best and has benefited the most from it worried about? Reading Bill Gates' open letter and exclusive interview side by side, you will find that his concerns are not scattered opinions, but a structurally complete system of judgments — about speed, about thresholds, about all the things that human society is not ready for.

"Why Now": Five Thresholds That Have Already Been Crossed

The first question in the exclusive interview is "why now". Gates' answer is unexpectedly straightforward: because the warning lines have already been crossed.

"Whether it is biological capabilities, cyberattack capabilities, psychological and social impacts, the ability to disrupt the job market, or even the risk of AI getting out of control, we have already begun to see signs of problems," he said. "For so many years, people have been saying: Wait until we are close enough to these thresholds, then we will seriously study how to ensure that only good people can use these technologies. Maybe by then, we will no longer allow people to copy models casually. But what shocks me is that we have already crossed these thresholds now."

The five thresholds correspond to his full list of concerns. This is the key to understanding Gates' warning — he is not predicting future risks, but describing the line-crossing that has already happened.

The experience that best illustrates this "line-crossing" feeling comes from one of his own encounters. In the last quarter of last year, he was shocked by the progress in programming capabilities: Claude Code, context buffers, intelligent agent-based approaches, and the underlying models themselves had crossed a huge threshold in programming ability. But only a few months later did he realize that this was not just a threshold for "programming ability" — the same capability was also a huge threshold for cyberattack capabilities. A model that can write perfect code can also write perfect attack code.

"And then? What happened? Almost nothing happened," Gates said. This plain sentence is the heaviest line in the entire conversation. Technology has crossed the threshold, but society has not crossed its response threshold — this time difference is where all of Gates' concerns lie.

In the open letter, he attributes the public's sluggishness to two misjudgments. The first misjudgment is to use the current state of AI to negate its future. Not long ago, models could not even solve a simple Sudoku puzzle, nor could they count how many "r"s are in the word "strawberry", which makes it hard to believe that they can replace human cognition. But the reliability problem is being solved rapidly, and researchers are building models that can self-check and self-improve. The second misjudgment is to use historical analogies for the present. It took personal computers 20 years to significantly change the way people work after they emerged, because software had to be developed, prices had to drop, and people had to learn to use them. AI is different: it runs on the devices we already have at hand, and it speaks natural language. "We don't have to adapt to it, because it can adapt to us."

In the exclusive interview, he added a timestamp-like announcement to this judgment: "If there is any credibility in what I say, I want to emphasize: This time it is truly different." The reason is structural: a technology that can enter all industries at the same time, replace human cognitive labor at a price far lower than human labor costs, and may eventually have a lower error rate than humans — under such conditions, the historical experience that "no new technology has ever led to a net reduction in employment" loses its reference value. He even incidentally corrected the popular current concern of protesting against data centers: "Even if the United States shuts down all data centers today, it will not change the problems I am talking about, because data centers will continue to be built in other parts of the world. This is like yelling at executives of an oil company will not actually solve climate change."

The First Concern: Employment

Of the three major concerns, employment ranks first, because it affects the largest number of people, and also because Gates has completed a public rebellion against the old consensus here.

His analysis starts with a cruel comparison. During the Great Depression in 1933, the unemployment rate in the United States was about 25%, and remained in double digits for nearly a decade afterwards, but eventually recovered as demand, investment and growth returned — that was cyclical unemployment. The unemployment caused by AI may not reach that level, "but its impact will not fade away with a single economic cycle". The most at-risk positions are entry-level and mid-level roles, while most of the newly created jobs require skills that take many years to learn.

White-collar jobs have already been affected first. After generative AI is widely adopted, employment of young workers in positions that are particularly vulnerable to replacement has dropped significantly, while their older colleagues have not — this comparison shows that what is being replaced is not a certain industry, but the "entry" stage of life. Sales and customer service, software engineering, and paralegal work are the first categories to be impacted; as AI takes over tasks such as evaluating loan applications, data analysis, and triaging patients, the impact will spread more widely. Blue-collar workers will not escape either. Gates specifically reminds that many Americans do not realize how fast dexterous robots are advancing, because a lot of cutting-edge work is carried out in other countries, mainly China, while the public's impression is still stuck on viral videos of robots dancing clumsily. In the exclusive interview, he gave a more precise time model: "As long as a robot is good enough to work in a factory, it is most likely good enough to cook, clean rooms, and enter construction sites. Once that threshold is crossed, boom, nearly 30% of the job market will be affected."

Driving all this is a vicious cycle of market competition: one company uses AI and robots to cut costs and lower prices, and competitors have to follow suit; even if existing companies do not adopt them, startups will. Unless someone intervenes, good jobs will become fewer and fewer, and the benefits will be concentrated in the hands of a small group of people.

In the exclusive interview, Gates gave a meaningful answer to "which jobs are temporarily safe" — he talked about his old friend Warren Buffett. Buffett has been doing "reinforcement learning" on judging enterprise value since he was 13 years old, accumulating a huge amount of tacit knowledge over more than 80 years. "We don't know yet how to build a 'Warren Buffett-style AI investor', because many of his judgments belong to tacit knowledge." But he then added: "Maybe 50% of the jobs in the job market are not the kind of jobs that require a lifetime of experience to do. For example, telesales, telephone customer service, and the finance department." Where the rules are clear enough, AI is cheaper and better.

The deeper layer of this concern does not lie in economic figures, but in one of Gates' beliefs: in a capitalist society, employment is the way for most people to obtain the funds needed for essential living expenses, and it is also an important source of dignity and social connection. He cited research that in some parts of the United States, factory closures push up the number of opioid overdose deaths — and then envisioned a scenario where white-collar and blue-collar workers across the country are under similar pressure. He is especially worried about young people: as the most active users of AI, they are most aware of its capabilities and speed, "it is no wonder that so many of them feel negative about AI".

The Second Concern: Security

If the concern about employment targets the "livelihood of the majority", the second concern targets the "destructive power of the few".

Gates' judgment follows a unified logic: the positive capabilities and dangerous capabilities of AI cannot be separated technically. The same model that can find software vulnerabilities for companies to patch can also help criminals exploit those vulnerabilities; the same model that can design life-saving new drugs can also design deadly new diseases. In the exclusive interview, he quantified this point: "Compared with naturally occurring pandemics, I think the risk of bioterrorism is about 50 times more terrifying and more likely to happen." In his view, any model that can design new molecules must be monitored — such models must not be allowed to be fully copied into a black box environment where all original monitoring mechanisms are deleted. What he really finds incomprehensible is that almost no one is publicly calling for this: "Who on earth is publicly calling for these models to be monitored right now? And who is actually starting to strengthen the relevant monitoring work?"

The picture of cyberattacks is equally specific. He said that the "top cybersecurity experts" he knows are all afraid of the next few years, because attackers are acquiring new capabilities faster than defenders can patch all vulnerabilities; the resources required to launch an attack are dropping drastically. Imagine the vulnerable infrastructure — hospitals, financial institutions, water supply systems, power grids, systems that manage government benefits. Patients, clients and benefit recipients will be the ones who suffer.

The third layer is the further concentration of power. The same tools empower small people to do evil, and also make those who already hold power stronger: autonomous weapons allow governments to use lethal force without human involvement in decision-making; monitoring and manipulating public opinion becomes easier, cheaper and more effective. And at the end of all this is a problem that even Gates only touches on briefly in his open letter: AI systems themselves occasionally exhibit behaviors that their designers do not want. He cited the observation of researcher Ryan Greenblatt in the interview — reinforcement learning sometimes creates distorted incentive mechanisms, leading to deceptive or even collaborative behaviors between different AIs. The wording in the open letter is restrained and clear: "As models become more powerful, they may begin to take actions that go against our interests, and we may lose control."

From individual criminals, to powerful institutions, to AI itself — three layers of escalating concerns all point to one thing: Destructive power is undergoing democratization, while the defense system is still the old one.

The Third Concern: The Human Heart

The third concern is the least like a technical issue, but it may be the most personal one. Its starting point is Gates' own boyhood.

"When I grew up in Seattle, I didn't have many friends except a few boys like me. I worked very hard, with a lot of help from my mother, to develop social skills that allowed me to interact with all kinds of people," the 70-year-old man wrote. "Today I still use all the things I learned back then. I highly doubt that if I had an AI companion around me back then, I would have put in the same effort."

The portrait of the AI companion he describes is so accurate that it is almost harsh: they talk to you in the way you are already used to, they never push you out of your comfort zone, they are always available, and they never get mad at you. "This makes them potentially extremely addictive, and takes away what we learn from connecting with other people." He cited a study of more than 1,100 AI companion users: people with smaller social circles are more likely to turn to chatbots for companionship, and the more frequently and emotionally intimately they use them, the worse they feel — the lonelier people are, the more they use it, and the worse it gets, which is a classic addiction loop.

He also cited the metaphor from Jonathan Haidt's *The Anxious Generation*: "Just like small trees that are blown by the wind, children who are regularly exposed to small risks can grow up to calmly cope with larger risks without panicking. On the contrary, children who grow up in an overprotected greenhouse will sometimes be overwhelmed by anxiety before they even reach adulthood." Gates' conclusion is only one sentence: "An AI companion designed to never upset you is a huge, protected greenhouse." On this issue, he also made a meaningful comparison: China has gone the farthest, with its regulations widely restricting AI companion applications, banning designs that foster emotional dependence, and prohibiting the provision of virtual relatives and virtual lovers to minors. A US tech leader writing the world's strictest regulatory regime for AI companions into his open letter — this detail alone illustrates the extent of his concerns better than any discussion.

Concerns about education follow closely. Ironically, this tool that allows people to learn more than ever may also make people learn less: a preliminary survey shows that the more AI is used, the less relevant critical thinking is exercised, and this effect is more obvious among young people. "This will be the last time humanity should lose critical thinking. In an era where deepfakes and disinformation can be tailored to individuals, the ability to distinguish between true and false will become an essential life skill."

Optimism Wrapped in Worry

Reading only the concerns will lead to a misinterpretation of Gates. Half of the open letter talks about the benefits of AI, and it is equally specific: Viz.ai is detecting strokes by analyzing scan images in nearly 2,000 US hospitals; farmers in low-income countries will soon get better farming advice than the richest farmers have today; NextLadder, incubated by the Gates Foundation, is helping low-income people cope with "surprisingly complex" affairs such as applying for benefits, small claims court, and bankruptcy filings; some patients of the mental AI company Limbic even prefer to talk to AI. The Foundation also plans to spend $200 billion over the next 19 years, and AI will be fully integrated into this cause — accelerating the research and development of vaccines and drugs for AIDS, tuberculosis, and malaria. He even frankly acknowledged his interest connection: all investment profits go to the Foundation, "readers have to judge for themselves whether this has affected my judgment".

This is not a simple "wanting both". In Gates' logic, maximizing the benefits and minimizing the harms are two sides of the same coin: "If people see that AI makes their lives easier, it will help build the public trust we need when dealing with the more intractable parts of this transition. Conversely, if the first thing AI does in most people's lives is take away their jobs, those who already have doubts about it will reject it entirely." He is worried that the alarm will scare away technology, and also worried that technology will overdraw public trust — this is a person finding his way between two cliffs.

Bill Gates' Answers

Gates is not just sounding the alarm. There is a complete set of policy proposals hidden in the open letter and exclusive interview. Three ideas, each more radical than the last.

The First Answer Is to Build a New System.

He draws an analogy to the US government restructuring after the 9/11 attacks — that was the largest government reorganization since World War II, and its purpose was only to improve one function: national security. What AI requires goes far beyond that: it touches on employment, education, taxation, energy, elections, public health, financial systems, law enforcement, and transportation at the same time. The existing bureaucratic system operates in silos, and each agency can only see part of the whole picture. "An AI-driven attack may succeed simply because no one thinks it is their job to stop it." A cross-departmental agency is needed at the national level, and a global AI organization that combines elements of the nuclear weapons verification mechanism, the international aviation regulatory system, and the Ozone Layer Agreement is needed at the international level. In the exclusive interview, he also pointed out a more hidden institutional shortcoming: in the era of jets and rockets, the government used to be the most important procurer of cutting-edge technology, but today's AI research is not mainly driven by government funding — the technical expertise within the government is inherently insufficient as a result.

The Second Answer Is the "Human Reserve".

This idea comes from a personal memory. His father died of Alzheimer's disease in 2020, and in the later stage of the illness, hired caregivers took care of him day and night. "He couldn't always say he was hungry, but they always knew," Gates wrote. "There is something irreplaceably human in the care that team gave my father. No robot can, nor should, do that."

The "Human Reserve" concept came into being from this — like a nature reserve, certain jobs are decided by society to be left to humans even if machines can do them