Every few months, someone inside the AI industry says something that makes the rest of us stop scrolling. This week, it was Evan Hubinger — the alignment science lead at Anthropic, one of the world’s leading AI safety labs — who put a number on humanity’s odds of survival.
His estimate? More than a 10% chance that AI could cause human extinction within the next decade.
That’s not a headline dreamed up by a tabloid or a doom-focused YouTuber. It came directly from someone whose entire job is to make AI systems safer.
What Exactly Did Evan Hubinger Say?
Hubinger, who leads alignment science at Anthropic, shared his concerns publicly, saying he genuinely believes AI has a real chance of killing everyone — and that he personally puts that probability above 10% within the next ten years. He also acknowledged something striking: Anthropic, despite its safety-first branding, does not yet have a fully worked-out plan to keep a future superintelligent AI under control.
That’s an unusually candid admission from inside a company that has built its reputation on being the “responsible” AI lab.
Why Is This Coming Up Now?
Hubinger’s comments didn’t appear out of nowhere. They followed the resignation of Anthropic researcher Jacob Coxon, who reportedly stepped down over unease about how fast AI companies are racing to build more powerful, self-improving systems — while safety research struggles to keep pace.
In other words, this isn’t just outside critics worrying about AI. It’s people who build these systems for a living, from the inside, raising the alarm.
The Real Danger Isn’t Today’s Chatbots
It’s important to be clear about what Hubinger is not saying. He isn’t claiming that ChatGPT, Claude, or any AI tool available today is about to end civilization. Current AI models, however impressive, are nowhere near that level of capability.
His concern is about what comes next: a future superintelligent AI system — one that surpasses human intelligence across the board and, crucially, can improve itself without human help.
What Is “Recursive Self-Improvement” — And Why Does It Scare Experts?
This is the concept at the heart of the warning: recursive self-improvement.
In simple terms, it means an AI system reaching a point where it can design a better, smarter version of itself, which can then design an even better version, and so on — without needing human engineers to guide each step.
If that loop kicks in, an AI’s capabilities could accelerate far faster than anyone anticipates, potentially outrunning our ability to understand it, control it, or shut it down if something goes wrong. That’s the scenario safety researchers like Hubinger are trying to prevent — and admitting they haven’t fully solved yet.
The Bigger Question: Is Safety Keeping Up With Capability?
AI labs across the industry — not just Anthropic — are locked in a race to build more powerful models. Faster releases, bigger context windows, more autonomous agents. It’s an exciting time to watch the technology evolve.
But Hubinger’s warning points to an uncomfortable question: is AI safety research advancing at the same speed as AI capability? Right now, many experts inside these companies seem to think the honest answer is “not yet.”
That 10% figure isn’t meant to spark panic. It’s meant to be a wake-up call — a reminder that the people closest to this technology are still uncertain about how to keep it safely under human control as it grows more powerful.
The Debate Is No Longer Just Outside AI Labs
For years, warnings about AI risk mostly came from outside commentators, philosophers, and a handful of vocal researchers. What makes this moment different is where the concern is coming from: directly inside one of the industry’s most prominent labs.
When a company’s own alignment lead says the problem isn’t solved — and a colleague resigns over the same worry — it signals that this conversation has moved from academic speculation to an active, unresolved debate happening in real time, inside the rooms where these AI systems are being built.
This article is based on recent public statements from an Anthropic researcher regarding long-term AI safety concerns.
