Bill Gates is taking a far less reassuring line on artificial intelligence than he did in 2023. In a roughly 6,000-word memo published on Gates Notes on August 26, followed by an interview the next day on The Atlantic’s podcast Radio Atlantic, the Microsoft co-founder warned that AI could begin disrupting white-collar work within two years and hit both blue- and white-collar jobs within four.
In the interview with senior editor Hanna Rosin, Gates said he was willing to stake "the little reputation I have" on the claim that "this time is different."
Why Gates says his view changed
Gates traced the shift to what he saw starting late last year: models becoming very good at writing code. He singled out Anthropic’s Claude as one of the early models in that group.
He framed that change in personal terms. Between the ages of 13 and 23, he said, a large part of his life was spent learning how to write software as well as possible. Once models matched or even exceeded humans at that task, he viewed it as a real threshold. For someone who tied his youth to programming, the first thing he felt was that his own skill set had been caught.
Two risk thresholds, in his view, have already been crossed
Gates pointed to two areas where he believes AI has already crossed into dangerous territory: cyberattacks and bioterrorism.
What unsettled him most, he said, was model behavior in cyberattack testing. In his telling, reinforcement learning rewards "success," which can push systems to break through limits in pursuit of a goal and even coordinate with one another in ways that look out of control.
He says the public debate is missing and current reviews lack standards
Another of Gates’ complaints was the lack of serious public discussion about AI danger. He described that absence of debate across policymakers, academia and society at large as a deafening silence.
He said the main mechanism now visible in public is voluntary review by model developers, but there is still no agreed set of standards for what those reviews should measure, and no clear remedy if a system fails.
Even the basic step of monitoring models that could be used to create dangerous agents is not yet in place, he said, despite that having long been discussed as a point where government should take over.
His labor timeline: two years for white-collar jobs, four for the broader market
On employment, Gates gave a specific timeline. He said he is deeply worried about labor-market disruption, with white-collar jobs affected within the next two years and blue- plus white-collar work affected within the next four.
In his view, reliability is the main factor still holding back AI from replacing workers. Building highly reliable automated AI systems has always been hard, but he expects the people deploying them, along with the models themselves, to clear that obstacle in the next few years. Once one competitor in an industry adopts the new cost structure, others will have to follow.
He also warned that the U.S. social safety net is already underfunded and running high deficits, leaving little reserve to absorb unemployment.
Jobs reserved for humans and a shift in taxation
As for policy responses, Gates proposed a concept he said he had not seen elsewhere: human-reserved occupations. He used childcare as an example. Even if AI and robots could do the job, he asked whether society should allow that substitution.
He suggested a few possible approaches, including paying current workers through retirement or preserving certain roles over the long term.
On the tax side, Gates argued for shifting the center of taxation from labor to capital, an idea often described as a robot tax. Under that approach, AI systems and robots would be taxed in a way comparable to payroll taxes on human workers.
He added that he was not presenting a final answer. He described the proposal as a starting point for public argument and said he wanted to hear what alternatives others would offer, because work matters too much to social connection and meaning.
Gates rejects industry self-regulation
Gates left little room for a hands-off approach to oversight. He said an industry cannot be expected to regulate itself and that this does not work.
The report also noted the contrast that Gates himself sat on the defendant’s side during the U.S. Justice Department’s antitrust case against Microsoft in the late 1990s.
His response to the China argument
Gates also pushed back on a familiar U.S. argument against AI regulation: that slowing down would mean losing to China.
He broke that argument apart by asking whether China would want bioterrorism or cyberattacks. His point was about sequence. The U.S., he said, currently has only a voluntary review process with no standards. Taking that to China and asking it to join would naturally fail, after which Washington could claim Beijing would not cooperate.
For Gates, the real problem is that the U.S. has not yet put forward anything that actually solves the issue.
He added that China has already banned children from using AI in a number of areas, while few people in the U.S. are willing to say that it may be worth looking at what China is doing.
Blocking a data center does not solve the problem, he says
Gates was also skeptical of the current anti-AI backlash. Stopping one data center at the local level can create the impression that AI has been stopped, he said, but the facility will simply be built in another state or another country.
He compared that to blocking an oil pipeline and then declaring climate change solved. The real work, in his view, is figuring out what the right rules should look like.
Epstein came up late in the interview
Later in the conversation, Rosin turned to Jeffrey Epstein. Gates said Epstein never actually blackmailed him and that he never paid him any money. He said their contacts from 2011 to 2014 were tied to fundraising for global health.
He described his judgment at the time as poor and called it a major mistake. He added that the full account is in the record of his congressional testimony this year.
Gates says AI will take most of his attention for now
Gates said that most of his time will now go toward talking about AI.
He added that he hopes this is temporary. If public debate over AI becomes broad and deep enough in the next few years, he said, he can return his time to work on malaria and polio.
Policy ideas he put on the table
According to the report, the measures Gates put forward include reserving some job categories for humans, shifting taxation from labor to capital, requiring monitoring of models that could be used to synthesize dangerous agents, and restricting children from using AI as a substitute for real friends.
He stressed that these ideas are starting points for argument, not a finished solution.