OpenAI's chief scientist is pushing for mandatory safety standards across the AI industry, a call that comes as the company itself admits it's getting harder to see how its most advanced models think.
Why the call for rules
Jakub Pachocki, OpenAI's chief scientist, has been making the case that AI labs shouldn't be left to police themselves. He's urging regulators and the industry to adopt binding safety requirements, rather than relying on voluntary commitments that can shift with market pressure.
The appeal lands as labs around the world race to ship bigger, more capable systems. That pace, Pachocki argues, is exactly why standards need to be locked in now, before the gap between what models can do and what developers can verify gets any wider.
The reasoning blind spot
OpenAI has acknowledged a growing problem: the more advanced its models become, the harder it is to trace how they arrive at their outputs. The company says monitoring the reasoning of these systems is becoming increasingly difficult, a challenge that cuts to the heart of safety testing.
If you can't fully follow a model's chain of thought, you can't easily spot when it's about to do something harmful. That's not a hypothetical worry. It's a practical obstacle that OpenAI says it's running into right now, and it's a big part of why Pachocki wants industry-wide rules that force labs to prove their models are safe, not just claim it.
A race that's hard to slow
The timing matters. AI development has moved from research papers to product launches in a matter of months, and each new release tends to raise the bar for everyone else. That competitive pressure can push labs to ship before they've fully stress-tested a model's behavior.
Pachocki's call for mandatory standards is a direct response to that dynamic. He's not asking for a pause or a moratorium. He's asking for a framework that makes safety checks a legal requirement, so that no single lab can cut corners to get ahead.
Whether other companies will back the idea is unclear. Some have already signed voluntary pledges, but those don't carry the weight of law. The question now is whether regulators will take up the recommendation, and whether the industry can agree on what those standards should look like before the next wave of models arrives.
