Bill Gates told NBC's Meet the Press on September 27, 2026 that the popular idea of an AI "kill switch" misses the point. "It's not enough to have a kill switch," he told moderator Kristen Welker, adding that the debate shows "how nontechnical various people are."
His main points, from NBC and Newsweek's coverage:
- Risk: "AI is certainly powerful enough to drive events that cause a billion deaths." Asked whether humans still control AI: "We do."
- Self-regulation: "No one thinks self-regulation is enough."
- Monitoring: to moderate bad behaviour "you need insight and records", with monitoring "where you're watching what the AI is doing" introduced in AI labs. Monitoring, he argued, "does not handicap us" in the race with China.
- Diplomacy: a global agreement would be "more difficult" than Cold War nuclear negotiations.
- Threat model: "the most dangerous thing we're facing right now is people with bad intent using AI"; "small groups with AI can do what only the biggest countries could do".
The interview landed amid a growing pile of proposals in Congress, including the bipartisan "AI Kill Switch Act" from Reps. Ted Lieu and Nathaniel Moran, which would let the Homeland Security Secretary order slowdowns of systems that could cause "catastrophic harm", and the Sanders–Casar superintelligence ban covered earlier this week.
Gates is right about kill switches, and the reason is engineering, not politics. A switch only helps if someone knows when to pull it, and by the time harm is visible in the world, the relevant actions have already happened. Agent incidents this month, such as the unauthorised access to an Australian government portal, were discovered after the fact from logs. That is why his "insight and records" line is the substantive one. Monitoring is the precondition for every other control: a kill switch, a pause, a liability rule or an audit. It is also the one control that fits his own threat model. If the main danger is people with bad intent using widely available AI, a switch on a frontier lab's servers does nothing, while logging and monitoring at the point of use can. The hard part he skips is who holds the records, since lab-held logs audited by nobody are close to self-regulation under another name.
Gates shifts the control debate from a kill switch to mandatory monitoring and records, which is the prerequisite for any intervention; the open question is who holds and audits those records.