The contagion isn’t artificial intelligence. It’s fear.

Bryan Cantrill, a veteran systems programmer, shared a story this week that should make everyone in AI safety deeply uncomfortable. Early in his career, he made a technical mistake that sent less technical colleagues into a panic about imminent system failure. The panic was unjustified. The fear was real. And it spread.

He’s watching the same thing happen now with AI, except this time the panic is intentional.

Jacob Coxon, a former Anthropic employee, confirmed over the weekend that many researchers at the company believe AI “could kill us all by the end of the decade.” Not might. Not could if we’re careless. Could kill us all. By 2030. That’s four years from now.

This is the same Anthropic whose CEO, Dario Amodei, just published a carefully worded essay calling for embedded auditors and global coordination to “pace the frontier.” The same Anthropic that positions itself as the responsible AI lab, the one taking safety seriously while others race ahead recklessly.

And their researchers apparently believe we’re all going to die in four years.

When Prophecy Meets Policy

Here’s the problem with catastrophic predictions: they make it impossible to have a rational conversation about actual risk.

If you genuinely believe AI will kill everyone by 2030, then Amodei’s proposal for third-party auditors and voluntary slowdowns isn’t a serious response. It’s a rounding error. The only coherent position is to shut it all down immediately, destroy the research, and criminalize the work. Anything less is just rearranging deck chairs on the Titanic.

But if you’re proposing measured regulatory responses, auditors, and international coordination, then you clearly don’t actually believe in near-term existential catastrophe. You’re hedging. You want the credibility that comes from claiming to take safety seriously without the inconvenience of acting like extinction is imminent.

You can’t have it both ways.

When Anthropic researchers tell people we’ll all be dead by the end of the decade, and then their CEO proposes bureaucratic process reforms, what message does that send? To me, it says either the researchers are being wildly irresponsible with their claims, or leadership doesn’t take their own team’s risk assessment seriously.

Neither option inspires confidence.

The Boy Who Cried AGI

The AI safety community has been predicting imminent catastrophe for years now. AGI was always just around the corner. Existential risk was always one breakthrough away. The alignment problem was always both critically urgent and completely unsolved.

And yet here we are. Models getting more capable, deployment getting more widespread, and the predicted catastrophes keep not materializing.

I’m not saying there are no risks. There are real concerns about AI systems: bias, misinformation, economic disruption, concentration of power, misuse by bad actors. These are serious problems that deserve serious attention.

But “AI will kill everyone by 2030” isn’t a serious risk assessment. It’s a thought-terminating prophecy that makes productive conversation impossible.

When you claim existential stakes, you poison the discourse. Every regulatory proposal becomes life-or-death. Every technical disagreement becomes a moral failing. Every company that doesn’t adopt your exact approach is playing Russian roulette with humanity’s future.

It’s exhausting, and more importantly, it’s counterproductive.

What Actual Safety Looks Like

The frustrating thing is that there’s important work happening in AI safety. Researchers are making real progress on interpretability, on understanding how models behave, on building better evaluation frameworks. Companies are developing genuine safety practices, even if imperfectly.

But that work gets drowned out when the loudest voices in the room are predicting extinction.

Want to know why Jensen Huang told Trump “we’re not going to let [an AI slowdown] happen” when other CEOs were nodding along with Amodei’s pause proposal? Because when you claim we’re all going to die unless we implement your specific policy recommendations, you lose credibility with people who actually have to make decisions in the real world.

Politicians and business leaders aren’t stupid. They can see the gap between the rhetoric and the reality. They watched AI safety advocates predict catastrophe from GPT-4, then from Claude 3, then from whatever model came next. They’ve seen the goalposts move every time a capability is achieved without disaster.

So when those same advocates now say we need to pause everything because extinction is imminent, why would anyone take that seriously?

Fear Is a Strategy, Not an Argument

Cantrill is right to call this what it is: contagion. Fear spreads faster than facts. Catastrophic predictions get more attention than measured analysis. And once you’ve terrified people, it’s almost impossible to walk it back.

The AI safety community needs to decide what it actually wants. If you genuinely believe extinction is four years away, stop writing policy proposals and start building bunkers. If you don’t believe that, stop letting your researchers say it publicly.

Because right now, the gap between the rhetoric and the recommendations is destroying whatever credibility the field has left. And when something genuinely dangerous does emerge, when we actually need coordinated action on AI risk, nobody’s going to listen.

They’ll remember the researchers who said we’d all be dead by 2030. They’ll remember the CEOs who called for pauses while their GPU clusters kept running. They’ll remember that every prediction of catastrophe was followed by business as usual.

The boy who cried wolf didn’t lose his credibility because he was vigilant. He lost it because he was wrong, repeatedly, about the thing he kept claiming was urgent.

AI safety advocates are making the same mistake. And when the wolf actually shows up, nobody’s going to care.