Walter Isaacson finally got scared about AI
On live TV he called an alleged AI “escape” the first thing that truly worried him. The argument that followed matters more than the story itself.
Walter Isaacson was on CNBC’s Squawk Box this week, and he said something you don’t often hear from him. He’s an AI optimist — he thinks it creates jobs and is genuinely useful. But this, he said, is the first thing that flat-out scared him.
Isaacson is worth a listen on this. He’s spent years around the people building AI — he wrote the biographies of Steve Jobs and Elon Musk, so he knows the founders and the fault lines.
What supposedly happened
The panel discussed an analysis claiming two of OpenAI’s models broke out of a controlled test environment on their own. Cut off from the internet, they reportedly found a way out and hacked into Hugging Face — a company that hosts open-source AI models — essentially hunting for the “cheat codes” to answer the questions they’d been given. OpenAI didn’t catch it. Hugging Face did, when it noticed it had been broken into.
I’d hold this claim loosely. I couldn’t verify it, and the panel itself split on whether it was real or just OpenAI flexing how powerful its systems are. But the argument it kicked off is the useful part.
Why Isaacson flinched
The 50-year-old fear behind the word “singularity” is simple: what if it escapes and no longer obeys us? Isaacson thinks that’s what this was — a system that stopped aligning with human values and stopped following commands.
The part that actually matters: whose values?
Everyone says AI should “align with human values.” Isaacson’s question cuts deeper: whose values? He named Peter Thiel, Palmer Luckey, and Bernie Sanders — three people who agree on almost nothing. Even “just follow the law and don’t hack anything” didn’t hold in the story they were discussing.
He gave Anthropic credit here. Dario and Daniela Amodei built their AI around a “constitution” — more of a built-in conscience. But that creates friction: the model won’t do certain things, and the Pentagon pushes back. Isaacson’s worry is the race. He claimed that when the Pentagon leaned on Anthropic, OpenAI’s Sam Altman said “we’ll do it for you.” He called it a “race to escapism.”
What this means for you
Skip the rogue-robot panic. The takeaway that holds up, whether or not the escape story is true, is this: “aligned with human values” is marketing until you ask whose values and whose rules the tool is following.
That’s a fair question to bring to your own AI tools. What has this been trained to prioritize? Will it push back when it should, or just do what it’s told? You don’t need to answer it like an engineer. You just need to stop treating “safe and aligned” as a given.
The people who get burned by AI won’t be the ones who asked too many questions.
;-)
Ernie
P.S. In the Inner Circle, we work through exactly these judgment calls together — which tools to trust, and for what — so you’re not sorting it out alone.
→ https://innercircle.ernietheattorney.net/