If AI Becomes Our Demise, Don’t Blame the Machines

Picture a cable car perched at the edge of a San Francisco hill, its brakes engaged. Below, a switch could send it down one of two unseen tracks. You have no idea whether either path is clear, and the operator is nowhere in sight. Do you let the car roll? Most of us would keep the brakes on, because the gamble is too risky.

Yet in Silicon Valley a very similar mental picture keeps playing in boardrooms and on Twitter. Executives, whistleblowers and hype‑machines are all shouting about a looming AI apocalypse, as if the very algorithms they wrote might flip a switch and unleash world‑ending chaos.

OpenAI’s Sam Altman, for instance, warned that large‑language models could become “more dangerous than nukes,” and he’s been quick to cite potential cyber‑attacks on power grids. Anthropic’s blog has talked about “thousands of high‑severity vulnerabilities” discovered by its Claude Mythos model, suggesting that soon even the most secure systems could be compromised.

Last weekend, former OpenAI data scientist Jakub Pachocki posted that nobody is ready for the fallout of ever‑faster machine intelligence. And just this week an Anthropic engineer, Jacob Coxon, quit his job, tweeting that the company is “gambling with our lives” and that its own staff believes AI could kill us all by the decade’s end.

These alarmist narratives sound convincing, especially when they come from insiders. But they also serve a very practical purpose: they keep the venture‑capital fire burning. By painting their work as a race against existential doom, startups can justify larger funding rounds, higher salaries and, frankly, more hype.

It’s worth remembering that the very same people who warn about apocalypse are also the ones pitching AI tools to utilities, banks and media firms as the next productivity miracle. OpenAI is simultaneously telling energy companies to brace for AI‑driven grid attacks while selling them chat‑based monitoring solutions that claim to keep the lights on.

If we take the worst‑case warnings at face value, we can imagine a scenario where a hyper‑intelligent chatbot flips a switch, cripples infrastructure, steals nuclear codes and sends humanity back to the Stone Age. In that nightmare, who’s to blame? The software? The hardware? The data centers? Or the CEOs who decided to push the button in the first place?

So far, the AI community has been quick to celebrate every new benchmark—whether it’s a model that writes a better email or generates a realistic 3D avatar. The same crowd then warns of “rogue superintelligence” with its own misanthropic agenda. The rhetoric shifts from “tool” to “entity” the moment the stakes look high enough.

But let’s cut through the drama. Large‑language models don’t have beliefs, intentions or desires. They don’t decide; they crunch probabilities and spit out the most likely next token. They are lines of code written by humans, run on servers owned by humans, and tested by humans. If a model produces a harmful output, the fault lies in the data, the prompts, the safeguards—or the lack thereof—not in some hidden consciousness.

In short, the danger isn’t a mysterious AI monster lurking in the cloud. It’s the choices we make: how we fund research, how transparent we are about capabilities, how rigorously we enforce safety protocols, and whether we let market pressure outrun prudent oversight. The real question should be: what will we do now, before the next hype cycle, to make sure the tools we create serve us rather than threaten us?