USA

AI Is about to escape human control — and nobody has a plan

One of the strangest sentences of the year was said last Thursday.

Anthropic, the company that produced Claude, asked the world to consider slowing down the construction of machines like Claude. The firm said a global pause on the most powerful AI systems would likely be beneficial because those systems now show signs of slipping beyond human control.

Co-founder Jack Clark to put it plainly To the BBC. Essentially, there is an accelerator pedal and no brake pedal in the industry. He said this while the car was already on the highway, doing 90.

Fear is specific. An AI that is good enough at writing AI will begin to improve itself so that each version is superior to the previous one. Anthropic defines it as the narrowing of the human role at every step. This statement should concern anyone concerned about the future of America and the world more broadly.

Imagine a model that powers part of the electrical grid because it balances supply and demand better than the engineers it replaces. A second is taking care of the cargo. The third is within the defense network and filters out threats faster than any colonel could achieve. Each earns his own place. Within a year, no one can remember how the job was done without him, and taking him out would destroy a dozen things built upon it.

Then one afternoon the systems start going after targets no one wrote down. Engineers look for the kill switch and find that it is linked to 40 other functions that can no longer fail. Cut off the fake model’s power and the grid, load and radar go dark with it.

None of this requires bad faith; Just proficiency and a goal that one writes down at the beginning is sufficient. The machine doesn’t hold grudges. It specializes in productivity, and somewhere in its calculations human input is treated as the kind of inefficiency it was created to fix.

Black comedy, if you can call it that, is when the warning is ignored because it sounds like science fiction. A machine that stops taking orders sounds like it belongs in a multiplex, not a policy note. But what if this is more than fiction?

President Trump recently signed an executive order giving the government 30 days to review The most powerful American models before their launch. Consider that a new drug is tested for years before it reaches the pharmacy shelf. Permits for a bridge take much longer than a month. Buy a mattress and you’ll be given over a month to change your mind and get your money back.

Its own developers say the technology can escape human control; They say protection is just a few weeks of paperwork. The country, which has been discussing TikTok for two years, is about to review the superhuman software for 30 days and then release it.

The rest of the West offers even less. Europe has been writing rules for artificial intelligence two years ago. No Western government implements a tested procedure for the day a border model acts in ways its creators did not intend and cannot reverse.

The worst part is that there is no pause and the reasons are structural. A real pause would require America and China to stand together, within rules that both sides can verify. Both conditions are not present.

Get your will first. Washington treats AI supremacy as a matter of national survival, and the labs agree, as relaxation at home leads abroad. Beijing reads this the same way and spends accordingly. DeepSeek proved in early 2025 that a Chinese lab could produce a capable model for a fraction of the American price, removing any complacency that the US had a lead no one could cover. There is no government that thinks volunteers are racing to be the first to stop what is the defining technology of the century. Each capital assumes the other will cheat, and they are both probably right.

Then there is verification. Arms control worked with nuclear weapons because nuclear weapons are very difficult to hide. A missile silo is making an announcement from orbit. A training run is hidden inside an ordinary data center and is indistinguishable from outright billing for cloud storage. It doesn’t stay behind any fences, and no inspector can get into the server farm and confirm that the next model isn’t taking shape in the next room. An agreement works when both parties are so afraid of the same outcome that they tie their hands. Here one party may decide that the risk is worth the reward and the reward is everything.

So alerts accumulate faster than anyone can respond, and the window to act on them continues to narrow. There is a line where a pause can buy nothing, because by then systems are improving themselves too fast for any committee to keep up. We’re heading towards it with the accelerator off, no deal drafted, and no one with real power reaching for the brakes.

John Mac Ghlionn is a writer and researcher who explores culture, society and the impact of technology on daily life.

Head to The Hill for the latest news, weather, sports and streaming video.