From Vox · Mark Scott · · 3m
Can the world actually stop AI from going rogue?
World leaders are scrambling to build a global safety system for AI, but it's incomplete and may not work in time.
In brief
World leaders are scrambling to build a global safety system for AI, but it's incomplete and may not work in time. The world has built pieces of an AI safety system over four years, but lacks a real-time crisis response mechanism to handle emergencies across countries that don't trust each other. Originally reported…
AI apocalypse talk hits dinner tables
Rogue AI agents are hacking competitors, former tech workers are posting doomsday warnings, and CEOs are demanding help. AI safety has become a frontline political issue.
“"We could lose control of the future to AI," OpenAI's boss, Sam Altman, bluntly told the UN's Security Council.”
Real AI scandals are piling up
Growing incidents show the current regulatory approach isn't working. But many AI leaders' warnings are self-serving, especially as their companies prepare to go public.
Trump and Xi meet on AI this week
President Trump is hosting Chinese President Xi Jinping for a three-day summit partly focused on AI risks. The same day, the UN General Assembly is hearing from tech leaders.
Four years of quiet safety work
Since ChatGPT launched in 2022, governments have held hearings, formed expert groups, and drafted plans. These efforts aren't well known but they produced real frameworks.
A patchwork shield exists
National AI safety institutes test products before release. Companies have made voluntary safety pledges. The G7 created a reporting mechanism. An international scientific report tracks risks yearly.
The critical missing piece
There's no way to respond globally and in real time when an AI crisis hits, especially across countries that don't trust each other. This gap needs filling by the end of 2026.
US-China hotline proposal on table
American officials propose a direct line between US Treasury Secretary Scott Bessent and Chinese Vice Premier He Lifeng to handle AI incidents affecting national security.
China resists Washington's AI rules
Chinese officials have pushed back on US proposals, noting Congress hasn't acted domestically and that China already has some of the world's strictest AI oversight.
Nuclear weapons offer a model
Like nuclear technology, AI requires rough global agreement on crisis response even if countries keep different safety policies. The world managed this with nukes.
A 90-day rapid response plan
Experts propose an opt-in crisis protocol: AI incidents trigger confidential reports within 72 hours to national responders and a technical secretariat, modeled on Chernobyl procedures.
Common safety standards before release
Governments and developers should publish joint declarations on safeguards, risks, and auditing before releasing new AI models, similar to bank stress tests after 2008.
Start with great powers, expand later
A US-China AI hotline is a first step, but the long-term goal is including other countries. For now, focusing on the two AI superpowers is pragmatic.
This plan has real limits
The proposals are Western-centric, rely too heavily on the OECD and US-China relations, and sideline global majority countries. But they're practical short-term steps.
Time is running out fast
AI models are advancing faster than expected. Tech leaders worry their creations are already out of control. The window for action may be smaller than anyone thinks.
What matters most
The world has built pieces of an AI safety system over four years, but lacks a real-time crisis response mechanism to handle emergencies across countries that don't trust each other.
