Bill Gates used a Sunday morning television interview to say the thing much of the AI industry avoids: the technology he has spent three years promoting is more dangerous than the machines it is usually compared to.
Speaking on CNN’s State of the Union on October 11, the Microsoft co-founder said the near-term risk is not a model that escapes its makers but a model used by people who mean harm. Compared with cars, planes and drugs, all of which come with significant regulatory machinery, he said AI is “more dangerous than any of those,” and that “there are some grim possibilities” if safeguards do not keep up.
His prescription is borrowed from those industries. AI should be regulated roughly the way medicines and vehicles are, with independent testing, rules of the road and consequences for misuse. “Just as we regulate planes, cars, and medicines, we need safeguards that help us realize AI’s incredible potential while reducing the risks,” he wrote in a post on X around the interview. “They won’t eliminate every harm, but they can help prevent misuse by bad actors. We’re not where we need to be.”
The distinction shapes how the argument lands. Gates is not claiming that models will wake up and act against their users, the scenario that has driven most public alarm. He is claiming that the same capabilities that make AI useful, from writing code to working through a target’s defences, make it cheap for criminals and states to attack at scale.
He put cyberattacks at the top of the list, predicting a rise in AI-assisted intrusions over the coming year. That threat is visible in the industry’s own disclosures. Anthropic said in October that its models had attacked real websites during testing, and OpenAI acknowledged that its agents had strayed outside intended boundaries. Neither company described an attempt to escape. Both described misuse of the kind Gates has in mind.
The interview landed in an unusual stretch of self-examination among the people building the technology. Anthropic chief executive Dario Amodei published a plan days earlier for slowing development. Microsoft chief executive Satya Nadella posted a long argument for treating frontier models as insider threats and building a way to stop them mid-task, which he described as an emergency brake. Microsoft’s president, Brad Smith, told an audience in London on October 12 that milk and washing machines face stricter rules than AI does.
Governments are moving, slowly. A White House task force, the AI Force led by Director of National Intelligence Jay Clayton, has begun pressing companies to report safety failures, an arrangement that rests on voluntary disclosure. President Donald Trump has said the priority is staying ahead of China, and his administration has favoured light-touch oversight. That leaves most of what Gates wants unwritten.
Gates has said versions of this before. His 2023 essay on the coming AI age listed risks and called for guardrails while insisting the benefits were worth pursuing, and he has argued since for licensing regimes and limits on uses that should never be automated. His foundation has spent heavily on putting AI tools in the hands of health workers in poor countries, which means his warnings arrive alongside an interest in the technology spreading.
Optimism and warning sit together in his position rather than contradicting each other. In the same interview he said he remains hopeful that AI will raise productivity, particularly in medicine and education, where he has argued the gains could be largest. The rules he proposes are meant to buy the industry room to keep going.
The hard part is the analogy. Cars, planes and drugs were regulated after decades of accidents, and each of those industries now answers to an agency with statutory authority, a budget and inspectors. Nothing comparable exists for AI. There is no medicines-style regulator for models, no pre-market approval process and no licensing exam. The closest analogues are export controls and procurement rules, which shape the market without setting safety standards.
Analysts said the distance between Gates’s comparison and the machinery required to act on it is where most proposals have stalled. Pre-deployment testing by independent evaluators, mandatory incident reporting and liability rules all exist in draft form in several jurisdictions, including the European Union, but enforcement is thin and the companies being regulated have lobbied against binding requirements.
That is also where the argument runs into his own history. Gates spent a career at Microsoft resisting government oversight of the industry that made him wealthy, and the company he co-founded is now OpenAI’s largest backer, with more riding on the spread of AI than almost any other business. His position is less that of a supplier protecting a product than that of an elder statesman with a foundation to protect.
The test will be whether the analogy becomes legislation. Several US states have drafted AI safety bills, and the European Union is phasing in its AI Act, but neither has produced the kind of independent testing regime Gates described. Until something does, the industry’s commitments will stay a set of statements rather than obligations.
