Google Wants Powerful AI Safety-Checked Before You Use It

Google and its AI lab DeepMind started a new group called the DeepMind Institute on Wednesday. Its job: get smart people talking about AGI, or artificial general intelligence — AI that could match humans at basically any task. The group's leaders include DeepMind co-founder Shane Legg, Google executive James Manyika, and DeepMind chair Demis Hassabis. Their first release is four essays on how to handle super-smart AI responsibly. The group openly admits its members won't always agree and will likely change their minds as new information arrives.

The headline idea comes from Hassabis himself. He proposes a US-led standards body that would test the most advanced AI models before companies release them. At first, companies would volunteer to submit models up to 30 days early. If the system works well, passing those tests could become mandatory. Crucially, the body would eventually design secret tests — so AI labs couldn't train their models to game the known ones. Hassabis says the rules could be "ratcheted up" if things get serious, including a coordinated slowdown across AI companies.

A second essay tackles a quieter but scarier problem: we're losing the ability to see how AI thinks. Today, you can often read an AI model's step-by-step reasoning and check its work. But newer, more powerful designs hide that process. Safety researchers Rohin Shah and Anca Dragan argue this isn't unavoidable. They suggest limiting certain kinds of hidden computation or requiring companies to prove less transparent systems are still checkable.

This all signals a real shift. The AI safety conversation is moving from vague worry to concrete rules about disclosure, outside scrutiny, and possible slowdowns. That matters to you because the AI tools heading into your job, your doctor's office, and your kid's classroom may soon face a checkpoint before they reach you.