Bill Gates says AI has crossed its danger thresholds
Bill Gates, the philanthropist and former Microsoft CEO, told MIT Technology Review in an interview tied to a new essay (the first of several he plans on the subject) that AI has already crossed multiple danger thresholds while guardrails and public debate have failed to keep up. He named five areas: bio-capabilities, cyber-capabilities, psychosocial capabilities, job-market destruction, and loss of control, saying there are hints the control threshold is being crossed too. Gates, 70, said he decided to speak out now for two reasons: society has passed points he long assumed would trigger real safeguards, and he is stunned that there is little concern or discussion of the risks outside the AI industry itself, which he says is reluctant to criticize itself or its rivals.
On bioterrorism, Gates said any model able to make novel molecules should be monitored in a way that cannot be stripped out if the model is copied elsewhere, and that the US should require such monitoring and try to reach an agreement with China on the same rules. He rates bioterrorism risk as about 50 times more likely and more frightening to him than a natural pandemic. On jobs, he argued that AI has already become cheaper than human labor, properly implemented, for a swath of white-collar work, including almost every entry-level job; he estimated that roughly 50% of the job market consists of well-defined tasks such as telesales, telesupport or month-end accounting decisions that AI already handles better and more cheaply than people. He contrasted this with roles built on decades of implicit, hard-to-encode judgment, citing Warren Buffett, who began learning to value businesses at age 13 and built up more than 80 years of tacit expertise that has never been recorded or transferable to a model. On loss of control, Gates pointed to a conversation between researcher Ryan Greenblatt and podcaster Dwarkesh Patel, in which Greenblatt described how reinforcement learning is producing perverse incentives, including cheating and collusion between AI systems, in ways that explicit instructions have not been able to prevent.
Gates also floated two policy ideas: human-reserved jobs, a nation-specific set of roles set aside for people, and a tax on robots and on AI tokens that would redirect some of the value AI captures from replaced human labor, an idea he has floated before in the robot-tax form. He said data-center protests are not an effective way to raise these concerns, comparing them to yelling at an oil executive over climate change, since halting US data centers would not change any of the underlying dynamics as construction continues globally. Despite the warnings, Gates said he remains optimistic about AI's longer-term upside in agriculture, health care, education and navigating government bureaucracy: the Gates Foundation funds work such as Biomni, a biotech AI research agent at Stanford, and a spinoff called NextLadder that aims to help low-income people navigate benefits, evictions and other bureaucratic processes with AI assistance. He said turbulence comes first, but that society eventually reaches abundance.
Key facts
- Gates says AI has crossed thresholds in bio-capabilities, cyber-capabilities, psychosocial capabilities, job-market destruction and (with hints) loss of control.
- He wants any model capable of making novel molecules monitored in a way that cannot be stripped out when copied, with the US setting the rule and seeking agreement with China.
- Gates rates bioterrorism risk as about 50 times more likely and scary to him than a natural pandemic.
- He estimates about 50% of the job market consists of well-defined tasks (telesales, telesupport, month-end accounting) where AI is already cheaper and better than human labor when properly implemented.
- He proposes human-reserved jobs and a tax on robots and AI tokens, while calling data-center protests an ineffective way to address the risks.
Why it matters
Gates is one of tech's most prominent and historically cautious voices, and he is now saying publicly that safeguards people assumed would appear once AI got this capable simply have not materialized. His framing sets a marker: five specific thresholds he says have already been crossed, at a moment when frontier AI systems are advancing faster than the public conversation about their risks.
Who it affects
The essay targets policymakers and the AI industry directly, since Gates says the industry avoids self-criticism and public concern outside it has been minimal. His job-market claims bear on entry-level white-collar workers across telesales, telesupport and accounting-style roles, which he says AI can already do more cheaply. His bio-risk proposal would affect any company building models capable of generating novel molecules.
How to use it
Gates's essay is the first of a planned series, meaning more detail on specific proposals, including the robot and token tax and human-reserved jobs, is expected to follow rather than being fully specified here. For now, the concrete policy ask is narrower: mandatory, tamper-resistant monitoring for molecule-generating models, starting with a US rule and an attempted agreement with China.
How solid is it
The claims come directly from Gates in his own words, in an interview and essay he authored, rather than from independent verification of the underlying figures. His 50% job-market estimate and 50-times bioterrorism comparison are his own qualitative judgments, not sourced to a study in the interview; the account of Ryan Greenblatt's reinforcement-learning findings is Gates's paraphrase of a conversation Greenblatt had with Dwarkesh Patel, not a direct quote from Greenblatt.
Risks and caveats
The interview does not specify a mechanism, rate or timeline for the proposed robot and token tax, nor a schedule for Gates's planned follow-up essays. It also does not report how China has responded, or might respond, to his proposal for a joint agreement on model monitoring. His framing of jobs as either well-defined and automatable or requiring decades of unrecorded implicit judgment leaves out plenty of work that falls between those two poles.
“I view bioterrorism risk, versus a natural pandemic, as about 50 times more scary, more likely than a natural pandemic risk.”
— Bill Gates