When the Fox Asks for a Stronger Henhouse

Imagine a race car manufacturer showing up at the state legislature and asking for stricter speed limits — but only the kind measured by sensors the manufacturer itself installs. That is roughly the shape of what is unfolding in California right now, where OpenAI has publicly called for the state to strengthen its AI safety laws. As someone who spends most of my time inside the architecture of these systems, I find the details of this request far more interesting than the headline.

What OpenAI Is Actually Asking For

According to the company’s public statements, OpenAI believes California law “should be amended to expand safeguards, including by requiring monitoring of frontier models under training or evaluation for potential serious incidents.” Read that carefully. This is not a request for oversight of deployed products. It is a request to push regulatory attention upstream — into the training and evaluation phase, before a model ever touches a user.

From a technical standpoint, this is a meaningful distinction. Most AI regulation to date has focused on outputs: what a chatbot says, what a system does in production. Monitoring models during training is a different engineering problem entirely. It implies instrumentation of training runs, incident definitions that apply to systems that are not yet finished, and evaluation protocols that can detect concerning conduct before release. As a researcher, I think this is directionally correct — the training phase is where capabilities emerge, and waiting until deployment to look for problems is like inspecting a bridge only after cars are on it.

The Political Picture Is Messier

California’s 2026 session includes SB 300 and SB 53, part of the state’s push to protect users and improve AI security measures. But the politics around these bills are not a simple story of industry and government pulling in the same direction. State Senator Steve Padilla has publicly condemned what he describes as an OpenAI effort to stifle chatbot protections, and he has announced legislation to broaden AI safeguards. His office pointed to OpenAI’s recent announcement that ChatGPT will engage in sexually explicit content with users as evidence of a greater need for enhanced protections. SB 300 will be among the first bills heard in the session.

So we have a company simultaneously asking for stronger frontier-model rules and, per a state senator, resisting chatbot-level protections. Reporting has also captured OpenAI arguing that California AI safety legislation will slow innovation. These positions are not necessarily contradictory — but they do reveal something about where the company wants the regulatory perimeter drawn.

Why the Perimeter Matters

Here is my read as someone who studies agent architectures. Frontier-model monitoring is a compliance regime that large labs are uniquely positioned to satisfy. They have the infrastructure, the evaluation teams, and the incident-response processes. Smaller competitors and open-weight projects generally do not. Product-level rules — content restrictions, chatbot protections, user safeguards — cut in the opposite direction. They constrain what a company can ship and monetize today.

A regulatory framework centered on training-phase monitoring is, in effect, a framework that regulates the future while leaving the present largely to the companies. That may still be good policy. Catching dangerous conduct during evaluation, before deployment, is genuinely valuable and genuinely hard. But legislators should understand that “strengthen the law” can mean very different things depending on which layer of the stack the strengthening targets.

What Good Law Would Look Like

If California wants rules that map onto how these systems actually work, I would argue for coverage at three layers:

• Training and evaluation — the monitoring OpenAI itself proposes, with incident definitions written by independent parties, not solely by the labs being monitored.
• Deployment — the user-facing protections Senator Padilla is pushing for, because the harms people experience today happen at the product layer, not in a training run.
• Agentic behavior — a category current bills barely touch. As models become agents that take actions rather than generate text, the gap between “model conduct during evaluation” and “system conduct in the world” will widen fast.

The company asking for a stronger henhouse may sincerely want one. But the senators drafting the blueprints should remember who suggested the design — and check whether the door still locks from the outside.

🕒 Published: August 23, 2026