Bill Gates says AI has already crossed danger thresholds in bio, cyber and jobs

Gates declares the danger lines already crossed

Bill Gates has issued one of his starkest warnings yet about artificial intelligence. In an essay published on August 26, along with an accompanying interview with MIT Technology Review at Gates Ventures in Kirkland, Washington, the 70-year-old Microsoft co-founder argues that AI has already moved past the points at which several categories of risk were supposed to be checked — and that almost nobody outside the industry is talking about it.

According to MIT Technology Review, Gates counts five thresholds as breached: biological capability, cyber capability, psychosocial harm, job-market destruction, and even the loss of control over the systems themselves. For years, he said, the working assumption was that as AI approached these lines, society would work out how to limit who could use powerful models and what those models could do. Instead, he describes himself as shocked that the moment has passed without that happening, and stunned by how little concern the subject generates beyond the industry itself, where companies rarely criticize one another.

The essay is the first of several Gates plans to publish on the topic. He is also raising it privately with leaders in industry, government and civil society, and he casts himself, somewhat reluctantly, as a deliberately "shrill" voice.

Biology sits at the top of the list

Among the risks he raised, the biological capabilities of frontier models drew his sharpest language. Any model that can produce novel molecules should be monitored, Gates said, and he views AI-enabled bioterrorism as roughly fifty times more frightening than a natural pandemic — and more likely, too.

Why the usual jobs argument fails this time

Gates also took on the standard reassurance that no previous technology has produced a net loss of jobs, an argument he says he has made himself in the past. History is a poor guide here, he argued, because AI can substitute for human cognition across many industries at the same time, at cost far below human labor, with error rates he expects to end up below human ones. He pointed to companies already hiring fewer entry-level workers as an early and still modest signal. Robotics has not arrived yet, he said, but the pace of progress is striking — a little faster in China than in the United States, though evident in both.

Taxes and human-reserved jobs

The essay also advances policy ideas. One is human-reserved jobs: categories of work that each society agrees to keep in human hands, with the choices varying from country to country. Another is taxation. A robot tax is a longtime proposal of his, and he now pairs it with a tax on the tokens generated when AI does work that previously required people, with the proceeds set aside for society.

Not a data-center argument

Gates was dismissive of protests aimed at data centers. Stopping every data center in the United States, he argued, would change none of the risks he describes, because compute will be built globally regardless. He compared the tactic to shouting at oil executives as a way of addressing climate change, and said the debate over minimizing AI's harms needs a more effective starting point.

Abundance, after turbulence

The message is not uniformly dark. Gates highlighted AI's promise in agriculture, health care and education, and its knack for bureaucracy — helping someone take a small-claims dispute to court, or, through NextLadder, a group spun out of the Gates Foundation, guiding low-income families through benefits, training programs, eviction and bankruptcy. The foundation itself, he noted, uses AI in vaccine and drug development, takes part in a public-domain effort on protein- and cell-level modeling, and funds Biomni, a biotech AI research agent at Stanford. His summary of the road ahead: turbulence first, abundance later.

Why it matters

Gates is not a critic at the fringe of the field. He is one of the most recognizable figures in software and global philanthropy, and his foundation is itself an enthusiastic AI user, which makes the warning harder to dismiss as technophobia. His framing also shifts the regulatory conversation: rather than preparing for capability thresholds still to come, he is saying they are behind us and the response is overdue. The concrete proposals he floats — reserved human work, a robot tax, a token tax — give policymakers specific instruments to debate instead of abstract principles. And his rejection of data-center protest tactics indicates where he believes pressure belongs: on governing what AI systems can do, not on where the compute gets built.