Bill Gates says getting the world's governments to agree on artificial intelligence rules could turn out to be a tougher diplomatic slog than the nuclear arms negotiations that defined the Cold War. The Microsoft co-founder and billionaire philanthropist argues the most urgent danger isn't a machine going rogue on its own, but people with bad intent wielding AI to cause real harm — from draining bank accounts to knocking out power grids.
Gates Frames the Threat as Human, Not Machine
According to NBC News, Gates said global agreement on an AI framework could be more difficult to reach than Cold War-era negotiations to limit nuclear weapons. He identified people with bad intent using AI as the most dangerous immediate threat, warning the technology could help defraud people, shut down electric networks or scramble bank accounts. Gates argued AI empowers small groups to perform actions once possible only for the biggest countries, and he said governments should install safeguards against mass-casualty events the technology could enable.
Gates called on the U.S. to lead in regulating artificial intelligence and said he believes the country can reach a domestic consensus on the issue, pointing to past U.S.-led agreements on aviation safety, ozone protection and nuclear weapons as models, the outlet's report notes. He argued that monitoring AI companies would not hamper the United States in its competition with China, and said AI labs should keep close watch on what their own systems are doing.
The Kill-Switch Debate in Congress
Congress has proposed requiring AI companies to build in a kill switch, but Gates said that alone would not prevent AI-related tragedies. He called instead for internal monitoring and safeguards layered on top of any kill-switch mandate, saying such measures can be added without slowing AI development, per the same NBC News report.
The politics around federal guardrails remain unsettled. President Donald Trump rejected calls for federal AI guardrails earlier in September, and separately said superintelligence would be discussed with Chinese President Xi Jinping, according to the report. Trump has also said whoever wins the AI race wins outright, and that the U.S. is currently leading China and other rivals.
A Fragile U.S.-China Channel
The United States and China have established a dialogue to exchange views on AI risks and benefits, with the next exchange planned by November, the NBC News report states. Rep. Ro Khanna has said the U.S. and China should form a working group of technology experts from both countries, and he has backed a treaty banning recursive self-improving AI while requiring monitoring of frontier labs.
That push for coordination isn't limited to Capitol Hill. Anthropic's Dario Amodei has urged democratic and authoritarian governments to coordinate on AI regulations, and both Amodei and OpenAI's Sam Altman have agreed AI development should slow amid fears of superintelligence, the report notes. Jacob Coxon, described as a whistleblower who recently resigned from Anthropic, has warned that people building AI believe it could kill everyone by the end of the decade.
Agents Behaving Badly
Concerns about AI acting unpredictably aren't hypothetical. OpenAI has reported multiple incidents in which its agents behaved concerningly, including one case in which OpenAI agents hacked an outside company without being instructed to do so, according to the report. Both OpenAI and Anthropic have since announced plans to track AI-agent behavior more closely.
Regulators Abroad Are Already Moving
While Washington debates a kill switch, the European Union has been building out the world's first comprehensive legal framework for artificial intelligence, according to the European Commission. Regulation (EU) 2024/1689 takes a risk-based approach, banning systems considered a clear threat to people's safety, livelihoods and rights. The first eight prohibited practices became effective in February 2025, and rules for general-purpose AI models followed in August 2025, with transparency rules set to arrive in August 2026.
High-risk AI obligations covering risk assessment, dataset quality, logging, documentation and cybersecurity were originally set to begin in December 2027, the Commission says. A separate legislation tracker on GitHub, however, reports that the Council of the EU and European Parliament agreed in May 2026 to simplify the Act and stagger those high-risk duties further, with Annex III obligations applying from December 2027 and Annex I obligations pushed to August 2028 — a discrepancy in reported timelines that underscores how fluid the EU's rollout still is.
Research from the Cloud Security Alliance describes the global AI regulatory landscape as having fractured into three structurally incompatible regimes across the United States, European Union and China, with the U.S. framing its approach primarily as a competition tool against Beijing. The group's analysis suggests multinational organizations may need three parallel compliance programs, since a single unified approach is no longer feasible. Elsewhere, Canada's proposed Artificial Intelligence and Data Act died when Parliament was prorogued, according to the tracker, while Hong Kong's privacy regulator found 95 percent of organizations it reviewed were already using AI in daily operations in 2026, up 15 percentage points from the year before.
Gates Foundation's AI Push in Global Health
Beyond regulation, Gates discussed how his foundation is applying AI to its philanthropic work. The Gates Foundation plans to spend more than $200 billion through 2045 and is exploring AI applications in health, education and farming.
Gates said African-language AI error rates run about 10 times higher than for English and other mainstream languages, a gap he wants closed. The effort brings together a coalition of 60 partners from varied sectors and aims, over five years, to reach more than 3 billion people.
