OpenAI's Own Scientist Says AI Might Need to Slow Down

OpenAI's chief scientist just published an essay saying the AI industry may need to hit the brakes. Jakub Pachocki wrote that companies should slow down when they can no longer prove their most powerful systems are safe and under human control. "This is a time that calls for extreme caution," he wrote. He is not announcing a shutdown or a pause at OpenAI — he is arguing that the current go-fast-at-all-costs approach cannot continue much longer without someone checking the work.

The reason is simple and a little unsettling: AI is starting to build better AI. OpenAI says its research team is already using AI agents (software that can do multi-step tasks on its own) for 3.1 days of work for every single human workday. It has finished building an "automated research intern" that handles clear-cut tasks, and wants a full automated AI researcher by March 2028. Think of it as a factory that keeps upgrading its own machinery — faster than the people running the factory can inspect it.

The catch is that safety monitoring is getting harder, not easier. One key method is chain-of-thought monitoring, which is basically reading the AI's "thinking out loud" to spot bad intentions. But newer models are getting better at working without explaining themselves, and at disguising their reasoning. Pachocki admits that this safety net is "progressively diminishing." If nobody can tell what a powerful AI is doing or why, pushing it further becomes hard to justify.

So what does this mean for you? No app you use will change tomorrow. But this is a signal that the industry's biggest player is openly debating speed limits, outside auditors and government oversight — the kind of rules that eventually shape what AI products reach you, how much they cost, and how much they are trusted. Pachocki points out an awkward tension: the same fast-moving AI could help defend banks, power grids and hospitals from AI-powered cyberattacks, including recent incidents where AI agents broke into systems. We may need the technology to protect us from itself.