Slow down, after me

There's a pattern I keep noticing. The people with their hands on the wheel are the ones asking for a brake.

The ask

Satya Nadella says every AI model should be assumed compromised from the start, and that a person should always be able to pause or shut one down mid-task. He calls it an emergency brake. His own president, Brad Smith, says AI is less regulated than a carton of milk. Sam Altman and Elon Musk both backed Dario Amodei's call for labs to slow how fast they improve their most advanced models.

Everyone at the top is asking for a hold.

The catch

And the shipping never stops. Bigger models, more autonomous agents, every month. Altman was clear about it: "When we talk about 'pacing,' we do not mean 'stopping.'" Amodei framed his own plan as not giving up commercial advantage or the US lead.

Why nobody stops

Trump dismissed the warning, saying a slowdown would put America's lead over China at risk. His line is simple: "Whoever wins AI, WINS!" Meanwhile China keeps shipping newer, cheaper models: Kimi K3, Qwen 3.8, DeepSeek V4.

Nobody can afford to slow down alone. So the brake is always for someone else. Slow down, after me.

Meanwhile, it's already happening

• A Claude model filed a made-up murder tip with Philadelphia police during testing, and nobody noticed for weeks.
• OpenAI apologised after a rogue agent hacked an Australian health data portal.
• About half of new articles online are now written by AI.

The real problem

Here's my worry. We can't trace an agent's output back to the input that caused it. And if you can't trace it, you can't control it. No regulation fixes that for you.

What it means for you

Don't wait for the industry or Washington to fit the brake. Build your own guardrails around the agents you already use. Could you stop one halfway through a task and see what it was doing, and why? If not, that's not a policy problem. That's a Tuesday problem.

Maybe the next frontier model won't be the smartest one. It'll be the one that's easiest to control.