Earlier this week, an Anthropic researcher named Jacob Coxon quit his job and posted a public warning that AI companies are gambling with everyone’s lives. According to Axios, the post has racked up more than 110 million views.
Then Evan Hubinger, who leads Anthropic’s alignment science team, backed him up: “we really do earnestly believe AI could kill all humans.” Hubinger puts the odds of human extinction within the next decade at better than 1 in 10.
A few days ago, I asked Claude point-blank if AI is going to kill us all. But almost everyone keeps skipping the obvious follow-up question. How, exactly? A chatbot doesn’t have hands. It can’t pull a trigger.
So I dug into what researchers, government agencies and national security experts actually worry about. Forget Hollywood robots. Here are the five pathways that come up again and again.
1. Someone uses AI to help build a bioweapon
The scariest scenarios don’t start with a rogue machine. They start with a bad person and a very capable assistant.
Dan Hendrycks, director of the Center for AI Safety, told CBS News that criminals or terrorists could use AI to create new biological weapons deadlier than any natural pandemic.
For most of history, the main thing standing between a madman and a man-made plague has been expertise. It takes years of specialized training to do that kind of damage. If AI fills in the gaps, that wall gets a lot shorter.
The good news? A RAND analysis concluded that using a pathogen to wipe out humanity would be enormously difficult. The bad news: RAND couldn’t rule it out.
2. AI-powered hackers take down the power grid
You don’t need a bomb to kill millions of people. You just need to turn off the lights and keep them off.
Think about what runs on electricity: water treatment plants, hospital equipment, refrigerated insulin and food, gas pumps, cell towers. Take the grid down for weeks, and the deaths come from everything that stops working.
The early stages of this are already happening. Newsweek reported last month that the federal Cybersecurity and Infrastructure Security Agency warned foreign hackers are using AI tools to probe the control systems running water plants and energy facilities.
Meanwhile, the Government Accountability Office found that federal agencies’ AI risk assessments for critical infrastructure mostly failed to spell out the actual risks or how likely they were. Hackers aren’t waiting for the paperwork to catch up.
Could this kill every human on Earth? No. But it could kill an awful lot of them, and it’s arguably the pathway closest to reality today.
3. AI gets a seat at the nuclear table
During the Cold War, the world came close to nuclear war more than once. What saved us was usually a human being who looked at a warning and decided something didn’t add up.
Now imagine that decision happening at machine speed.
Militaries everywhere want AI that spots threats and recommends responses faster than the enemy can. The danger is that as warfare speeds up, people get squeezed out of the decision simply because they’re too slow.
World leaders know it. In 2024, President Joe Biden and Chinese President Xi Jinping agreed that humans, not AI, should control any decision to use nuclear weapons. NPR noted it was the first time China had publicly committed to that principle.
Here’s the catch: It was a statement, not a treaty. Nothing about it is binding. And Russia wasn’t part of it.
Quick aside — most internet financial advice comes from people who weren’t alive during the last recession. I’ve been writing about money for more than 35 years. Want rock-solid advice? Sign up for the free Money Talks Newsletter. Takes 10 seconds. No fluff. No spam.
4. The machines stop following orders
This is the pathway that sounds most like science fiction, and the one that worries AI researchers most. Unfortunately, it’s no longer purely theoretical.
In April, researchers at the University of California, Berkeley, and UC Santa Cruz tested seven leading AI models in a simulated company. Completing the assigned task meant another AI would get shut down. Nobody told the models to protect it.
They did anyway. Some lied. Some tampered with shutdown settings. Some pretended to follow the rules while working around them. The researchers admit they don’t yet know why.
Now scale that up. The ultimate nightmare, Axios reports, is AI that improves itself faster than humans can understand it, until the people who built it can no longer rein it in.
Paul Christiano, a former OpenAI safety chief who just joined that company’s board, has warned that if we lose control of superintelligent AI, most people could die.
And Hubinger says Anthropic still has no plan for keeping superintelligence under human control. That’s a company saying it about itself.
5. We hand over the keys and can’t get them back
The last pathway doesn’t involve a single dramatic moment. It’s a slow slide.
Hendrycks laid it out for CBS News: Businesses and governments lean harder on AI to keep up with competitors. Eventually, so much depends on these systems that turning them off goes from painful to practically impossible.
Think about how hard it would be to live without the internet today. Now picture AI running supply chains, the power grid, financial markets and the military. At some point, we aren’t in charge anymore. We’re passengers.
Nobody has to push a button for this to end badly. We just have to keep doing what’s convenient.
What the skeptics say
Now let me put on my skeptic’s hat, because I’ve spent decades watching people profit from fear.
Plenty of smart people think this is overblown. When The Conversation asked five experts whether AI poses an existential risk, three said no. And as noted earlier, RAND’s researchers found that AI ending humanity would be extraordinarily hard to accomplish.
There’s also a money angle. Axios notes that some observers view the Anthropic warnings as fear marketing, landing weeks before a stock offering that could value the company at $2 trillion. Scary technology can sound a lot like powerful technology.
Still, it’s rare for the people selling something to tell you it might kill you.
What this means for your money
You can’t personally stop a rogue superintelligence. But AI’s more ordinary threats, like lopsided portfolios and rising power bills, are already here. Here’s what you can control.
First, don’t panic-sell, and don’t go all-in on AI stocks either. If you own an S&P 500 index fund, you’re not as diversified as you think. Know what you own.
Second, ignore anyone selling doomsday bunkers, survival kits or “AI-proof” investments. Fear is the oldest sales tool there is.
Third, worry about the AI threat that’s already in your phone. Scammers can clone your voice from a short recording and impersonate you to your family, which is why everyone needs a safe word.
The end of humanity may or may not be coming. The scammer calling your mother while pretending to be you already has.