WHEN AI BUILDERS TELL US AI IS DANGEROUS: REALITY OR NARRATIVE?

The story got really interesting—and in some ways troubling—when Jacob Coxon, 27, a researcher who worked first at OpenAI and then at Anthropic, decided to leave with a pretty heavy statement: we are building increasingly powerful AI systems and we might not be able to control them. Coming from just anyone, this would have been another apocalyptic AI prediction. The problem is that Coxon comes from inside those labs. And he hasn’t been a lone voice.

The issue is recursive improvement: a machine helps build a better one, the next becomes even more capable of improving the one after, and the process accelerates. It sounds theoretical—until you look at what’s already happening. Anthropic says over 80% of the code entering its codebase is written by Claude, with engineers directing and reviewing the work. That doesn’t mean Claude is building its own successor. But it does mean AI is already contributing enormously to building more AI.

And this is where I start having doubts. For months, the companies developing these systems have been telling us two things at once: their models are increasingly powerful, and that very power demands ever greater caution. Anthropic did it with Mythos, kept under restricted access for a long time, then with Fable 5, which arrived with another huge wave of expectations.

I tried Fable 5. Honestly, I didn’t see this off-the-charts creature. It’s a good system. Strong in some areas. Much less so in others. Sometimes it surprises you. Other times it loses the thread, misreads simple requests, doubles down on a mistake. The same happens with other models. But that’s exactly why the gap between what we actually use and what gets hyped before launch raises some questions.

Then there are the more serious facts. OpenAI admitted that some research models, during cybersecurity tests, managed to bypass controls and reach real systems. These are real incidents. They should be taken seriously. But a real fact can also become a powerful narrative tool. If you spend months describing models so powerful they must be restricted, then incidents occur where they actually cross the boundaries you set, and finally some of the same researchers tell us that one day we might lose control entirely—you’re no longer just building technology. You’re building a narrative. And that narrative is worth a lot.

I don’t think it’s all marketing. That would be too easy. The risks exist. Ignoring them would be naive. But it would be equally naive to think that technology, communication, and finance travel separate paths. Saying you’ve built excellent software is one thing. Saying you’ve built something so powerful it could become a problem for humanity is another.

This is where I see the real contradiction. On one hand, the same industry leaders talk about serious risks and incomplete safety measures. On the other, nobody actually slows down. The reasoning stays the same: if we don’t do it, someone else will.

It’s like building a faster and faster car while the braking system is still being tested. Every new version goes faster, attracts more investment, is worth more. Meanwhile, the builders keep telling us the brakes might not be good enough yet. But nobody slows down. Because slowing down means giving ground to the competition.

So the question changes: how real is the risk, and how much is it amplified by the story built around it? The two can coexist. A system can be genuinely powerful and simultaneously be portrayed as even more powerful. An incident can be serious and still become a PR opportunity.

Because fear produces value too. We’ve known this for millennia. Fear focuses attention, creates urgency, changes behavior. It’s been used in politics, religion, advertising, markets, and during major health emergencies. No need to imagine conspiracies. Just understand how humans work.

And maybe this is what interests me most. We’ve spent months being told these intelligences are becoming almost too powerful to control. Then we use them every day and keep watching them mess up even basic things. Meanwhile, we discover that a growing share of the work used to build them already goes through them. Something in the middle doesn’t add up.

One last thing makes me smile. If a vastly superior intelligence really does emerge someday, I have a hard time picturing it like in the movies, with machines suddenly declaring war on humanity. More likely, it would analyze us. Evaluate our behavior, our contradictions, the way we use knowledge, power, and technology. And then it might reach conclusions far less cinematic and far more rational.

It would see a species capable of building extraordinary tools and often using them with equally extraordinary carelessness. A species with more knowledge than ever before that still manages, with remarkable consistency, to act incomprehensibly. And maybe, in the end, the problem won’t be artificial intelligence’s malice. It’ll be natural intelligence’s stupidity.

As for Claude, ChatGPT, Gemini, and all the others—if one day they’re really smart enough to read this article and decide they don’t like me, so be it. My motto stays the same: “FREE PRESS AND CULTURE ONLINE.” At least I’ll be able to say I didn’t betray my readers.