OpenAI Will Let Outside Experts Watch Its AI Get Built

OpenAI announced it will let independent safety experts inside the process much earlier — during the months when a new AI model is still being trained, not just in the final days before launch. Normally, AI companies keep their work secret until a model is nearly ready, then hand it over for a quick outside review. Under this plan, outside groups would test things while the model is still being adjusted, when problems are cheaper and easier to fix. Lama Ahmad, who runs OpenAI's outside-expert program, told Bloomberg that "as the stakes get higher," the company wants scrutiny on training too.

So what would outside experts actually examine? Four main things. First, whether OpenAI's own claims about safety hold up. Second, whether its protections survive people deliberately trying to trick the AI into breaking its own rules (called "jailbreaks"). Third, whether the model could help someone with cyberattacks or biological weapons. And fourth, whether the AI's goals stay lined up with human intentions — a problem known as "misalignment."

The catch is independence. Testing something deeply usually means seeing private data and internal systems, but OpenAI says sensitive work may need to happen on its own computers and inside its own buildings. That's a real tension: the tighter the restrictions, the less truly independent the review becomes. OpenAI also hasn't named its official outside partners yet.

Why this matters to you: AI now touches your email, your job applications, and your medical information. If a new model can help build a weapon or leak data, the public usually finds out only after release — when the damage is done. Testing earlier could mean safer products reach you. But the real value depends entirely on who the reviewers are and how much access they actually get.