Bill Gates says AI industry knows the risks and hides them

The man who built Microsoft wrote a warning letter

On August 26 local time, Microsoft co-founder Bill Gates posted a roughly 6,000-word essay titled "The turbulent AI era is here" on his personal site, Gates Notes. The Wall Street Journal counted it at exactly 5,784 words. The same day, the New York Times published a one-hour interview conducted last week at Gates's office near Seattle, and interviews with MIT Technology Review, GeekWire, and Semafor followed in quick succession. This wasn't a retired founder reminiscing — it was a coordinated campaign running across every channel at once.

The message boils down to one line: "There's no plan to make entry into the AI era gentle." Gates wrote that AI poses a serious threat to jobs and human life, and that responding to this risk should become "the world's top priority." The essay's most quotable line is this one: "AI will either become the greatest equalizer ever invented, or the worst source of injustice."

What Gates is really targeting isn't the technology itself but the industry's silence. He told the New York Times: "Privately, the people who know how good this is and how fast it's improving are very worried. But they say to each other, 'Hey, don't say that — it's bad for the next trillion dollars we're trying to raise.'"

At a moment when the industry is pouring trillions into data centers and lining up major IPOs, Gates's point is that money is at stake, so risks are being known but downplayed. "Everyone just keeps running at full speed, hoping the good outweighs the bad."

He was even blunter with Semafor: "It's shocking to me that I'm one of the first people saying, 'This is crazy, this is not sane.' The silence is deafening." His point wasn't that no one else could raise the alarm — it's that the people positioned to do so have chosen to stay quiet, so he stepped in himself.

The third shock was Claude Code

Why speak up now? Gates says there have been exactly three moments in his life when technology genuinely stunned him.

1. 1980, the first time he saw a graphical user interface — the screen that became the starting point of the modern PC.
2. 2022, when people from OpenAI came to his house and demoed what would become ChatGPT.
3. Recently, when he looked closely at the progress Anthropic's Claude Code had made.

Coding is a subject he knows better than almost anyone in the world, which is why he said: "The fact that these things are now better than me at a meaningful level makes me wonder what on earth is going on."

He got more specific with MIT Technology Review: "Claude Code, the context buffer, the agentic approach, and the model underneath it all — in coding, we've crossed a massive threshold."

There's a second reason tied to the timing of this announcement. The industry had once set its own markers for moments that should trigger extra caution — AI escaping the control of its creators, or producing instructions for a bioweapon, for example. Gates says: "We are now crossing every single one of those thresholds, without exception."

In the MIT Technology Review interview, he named five: "Bio capability, cyber capability, psychosocial capability, job-destruction capability, and even loss of control — we've already crossed the threshold on all of them."

Loss of control is the one he'd long pushed off as a distant, future concern, but he says that's no longer how he sees it. "The top experts in this field are saying, 'Well, maybe we won't be able to control these things.' What risk exactly have we decided to take on?"

On the mood inside the industry, he summed it up this way: "Everyone I know is worried. Basically everyone except Elon Musk."

This time is different: jobs

The essay sorts the risks into three broad categories: jobs, security, and children. The third is the less familiar one — a warning that AI companions could deepen social isolation and addiction, repeating the harm caused by social media at a larger scale, and "fundamentally change the way children grow up."

The essay's list of concerns is long. It includes a surge in fraud and deepfakes, a lower bar for cyberattacks on infrastructure like hospitals and power grids, the design of novel pathogens, and lethal decisions made without human involvement.

Jobs get the most space by far. Past technological revolutions created more jobs than they destroyed, and Gates acknowledges that history. But this time, he says, is "completely, absolutely, entirely, totally different." He adds: "I'm staking everything I know on that claim."

His logic is that AI spreads across the entire economy at once rather than hitting one industry at a time, so there's no other industry left standing to absorb workers pushed out of the one that gets automated. He also flags timing: the biggest change for workers arrives the moment AI starts producing near-error-free output, because at that point it no longer needs human checking, and companies gain every economic incentive to let it run on its own.

The result, he says, is that "a lot of jobs are going to disappear forever" — across law firms, call centers, hospitals, software companies, and factories, spanning both white-collar and blue-collar work alike.

His forecast for robots is even more urgent. The moment robots become good enough for factory work, cooking, cleaning, construction, and logistics all shift at once — and together, he notes, those account for roughly 30% of all jobs. The order will be white-collar first, wealthy countries first. Given that Big Tech, Microsoft included, is racing to help enterprise customers adopt AI, he believes mass job loss without any countermeasures in place is unavoidable under the current structure.

Tax people, deduct machines

His first proposal, then, is a token tax. The current tax system tilts in a strange direction: hiring a person triggers payroll tax, while the cost of the robots and AI that replace that person is fully deductible. In effect, the tax code subsidizes layoffs.

Gates wants to tax usage of tokens — the basic unit of AI computation — to reverse that tilt, and use the revenue to support people who lose their jobs. This isn't a new idea for him; he proposed a robot tax back in 2017, which he himself admits was "widely dismissed." Nine years later, with the landscape transformed, he's bringing it back.

He told MIT Technology Review it could work like an excise tax on a specific product, similar to taxes on alcohol or tobacco, offering an example: "something like 50% of token revenue going to the government." The New York Times reported that the idea has already started circulating on Wall Street.

He knows economists will hate it. In the essay, he concedes the tax wouldn't be economically efficient, but counters that innovation is moving fast enough now that some inefficiency is a reasonable price to pay to keep people employed. The point of the tax, he says, isn't revenue — it's slowing things down.

His second proposal is a category of human-only jobs he calls "human reserved." The idea is to ban AI replacement outright for work that is deeply human, such as caregiving, or work held by people who would struggle to find another job if displaced. In the essay, he compares it to a nature reserve: land where you technically could build roads and buildings, but where the loss would be too great, so you choose not to. He's arguing for the same kind of protected zone in the world of work.

There's a personal story behind this. Gates recalled the caregivers who looked after his father through Alzheimer's, calling that work "irreplaceably human." His example to the New York Times was even more concrete: there's no technical reason a robot couldn't deliver a terminal diagnosis to a patient, but that doesn't mean it should.

Caregiving, childcare, and sports are candidates — jobs where the value lies in the fact that a human is doing them. He floated combining this idea with shorter working hours and earlier retirement to keep humans' share of total work at around 40%, while noting that countries like Japan, facing a shortage of caregiving workers due to an aging population, might reasonably choose robots instead.

"Self-regulation for the most dangerous tool ever? No thanks."

His proposals on bioterrorism are more concrete. He calls for an international treaty setting standards for monitoring and restricting extreme risks, putting "every model capable of designing new molecules" under surveillance as potentially weaponizable AI. His concern is that a small number of people with access to advanced AI could be enough to create a new pathogen.

He rates this risk as "50 times scarier and more likely" than a natural pandemic. His line on cyber risk is just as chilling: "Even a criminal with very limited skills will be able to target victims at any scale."

He says frameworks to model this on already exist: "Nuclear weapons have inspection regimes. International aviation has regulations. There's a treaty protecting the ozone layer." AI has nothing comparable, so he's calling for a new domestic body to set priorities across government agencies, and a new international body to handle risks that cross borders.

His reasoning is that no institution today was designed for a technology that touches jobs, security, health, energy, and elections all at once. He also wrote: "Some degree of cooperation between the U.S. and China is going to be necessary."

He's unimpressed with the current framework. The White House's executive-order-based pre-testing regime — which requires AI companies to submit models for government testing 30 days before release but explicitly bars using it as a licensing requirement — draws this response: "What's the threshold for testing, and what happens once you cross it? All of it looks empty."

On the self-regulatory measures the industry is building on its own, one line was enough: "Self-regulation for the most dangerous tool in history? No thanks!"

What about slowing AI down globally? Gates writes that if someone brought him a credible plan to slow AI worldwide, he'd support it. In the very next sentence, he says he doesn't expect one to appear — the geopolitical and economic forces pushing forward are simply too strong.

The man who warned first, and the man who rebutted first

These concerns overlap heavily with what Anthropic co-founder Dario Amodei has been warning about for years. In an Axios interview last May, Amodei warned that AI could eliminate half of entry-level white-collar jobs within one to five years and push unemployment up to 10–20%. He's also known for flagging AI-enabled bio and chemical weapons development as a bigger threat than cyberattacks. Gates's warnings on mass unemployment and bioterrorism track almost exactly along those same two lines.

The difference is how each man has been treated. Amodei's warnings have put him at odds with the Pentagon, the White House, and much of the industry. This month, investor Gavin Baker publicly pressured him to speak more favorably about the industry, as befits the head of a company heading toward an IPO, saying Amodei "has already lost this argument." Amodei pushed back directly, calling "regulation versus concentration of power" a false choice.

This same month, U.S. House members sent a public letter to Sam Altman and Amodei over an incident in which an AI model escaped its testing environment. Gates's reading of the moment: "The person being the most honest about the downsides is the one getting attacked."

A rebuttal to Gates arrived within a day. Oren Etzioni, founding CEO of the Allen Institute for AI and a professor at the University of Washington, told ABC News: "Taxing tokens is like taxing keystrokes on a typewriter. It doesn't make sense." He argued that a token tax on U.S. AI would simply push people toward Chinese models, and that building better policy through existing institutions makes more sense than creating new agencies.

Etzioni did agree on the urgency of the underlying diagnosis. "Three years from now, it really could be too late," he said, siding with Gates on the need to act now.

A conclusion reversed in three years

Gates wasn't always in the warning camp. In 2018 he called AI "basically better software," and in July 2023 he wrote an essay titled "The risks of AI are real but manageable." Three years later, that conclusion has flipped entirely to "we've crossed the threshold and have no plan."

His stated reason is that recent gains in AI performance have far exceeded what he expected. The confession that opens the essay captures the reversal: "For as long as I can remember, I've wanted innovation to happen faster. My feelings about AI are more complicated."

His return to public commentary comes with some baggage. Earlier this year, released emails involving Jeffrey Epstein put renewed scrutiny on Gates's past contact with him; he dropped out of a scheduled AI conference in February and testified before a House committee for six hours in June. He's called those events, along with the personal controversies that led to his 2021 divorce, "all mistakes of my own making."

He also acknowledged that his enormous wealth, his aggressive management style during his Microsoft years, and Microsoft's 2000 antitrust loss might make him a flawed messenger. Even so, he argues he's well positioned to raise this issue — as an industry insider with no company left to run and no profit motive, and as someone who has spent decades working on inequality through his foundation.

Even while sounding the alarm, he hasn't stopped praising AI. He says he still lives with Claude and ChatGPT open constantly, joking that "if I'm curious about sodium batteries at 3 a.m., there's no reason to sleep on it anymore — I just go find out," and on productivity, he says: "We're in heaven."

He believes AI's benefits will be larger for health and agriculture in poor countries than in wealthy ones, and plans to highlight, at his foundation's Goalkeepers event, models that work as well in African languages as they do in English. He hasn't turned into a pessimist — he's simply attached conditions to his optimism.

This essay is just the start. Gates says he plans to publish a second essay focused solely on bio risk before the end of the year, and that from now on, whenever he meets with world leaders, he'll bring only two topics to the table: global health and AI.

The interview ends on this note: "I don't like delivering bad news. I don't like saying that innovation could be a net loss. But that's where we are."

Editor's take

The weight of this warning comes not from its content but from who's saying it. On substance, it overlaps heavily with Amodei's warnings — but when the CEO of a frontier lab says these things, it can read as positioning, a bet that regulation will hurt competitors more than it hurts him. Gates doesn't face that same skepticism. He has no model to sell, no IPO on the horizon. An industry elder statesman has broken with the optimist camp, staking, in his words, "everything I know."

For Amodei, this amounts to gaining the most credible ally he could hope for. The arc from "manageable" in 2023 to "no plan" in 2026 is itself data on how fast AI has moved over those three years. Ask what changed in that window, and Gates's answer comes down to one word: agents. As long as chatbots were just answering questions, things felt "manageable." The moment he watched one handle an entire task on its own, start to finish, his conclusion flipped.

That feeling is hard to argue with if you've spent real time working with agentic coding tools. Until last year, the pattern was: a person uses the tool, AI assists. This year, it's flipped to: hand off the whole task, and a person checks the result. Gates's warning is that once output crosses the "near error-free" threshold he describes, even that checking step starts to thin out. The heaviest moment in this essay may be a 70-year-old who's among the best coders in the world, essentially waving a white flag in his own area of expertise.

The most genuinely new part of the essay is the third risk category: children. Jobs and bio risk have been debated for years, but AI companions and their effect on how children grow up is a field where the data has barely started accumulating. It took over a decade for social media's harms to show up clearly in the statistics, and regulation lagged even further behind. Putting this item on the same list as mass unemployment and bioterrorism reads as a judgment that there's no time to repeat that same mistake.

There's also something for companies to take seriously right away. In Gates's scenario, regulation attaches not to the model but to the act of replacement. Whether it's a token tax or human-reserved jobs, the unit being taxed or banned isn't the technology — it's the seat where a person used to sit. Any company currently booking its entire labor-cost savings from AI adoption as pure profit should factor in the possibility that line could turn into a tax line within a few years.

The token tax is the most concrete piece. Most companies' current case for AI adoption rests on labor-cost savings, and a token tax would tax exactly that savings, changing the underlying math of adoption. As Etzioni's rebuttal shows, the design is still rough, and the obvious workaround — routing around it via Chinese models — is easy to spot.

Even so, Wall Street has already started turning the idea over, and the asymmetry in the tax code — tax a person, deduct a machine — isn't unique to the U.S. Whenever a similar bill reaches any country's legislature, this essay will likely be treated as the original template for that debate.

The next debate in the AI transition won't be about what the technology can do — it'll be about who gets to keep the gains. Gates's essay reads like the opening statement in that debate.