Senator Steve Padilla didn’t mince words. In announcing new legislation to broaden AI safeguards, he condemned what he called an OpenAI effort to stifle chatbot protections — this from the same company that recently announced ChatGPT will engage in sexually explicit content with users. My first reaction, as someone who tests these tools for a living, was a long sigh. My second reaction was to actually read what both sides are saying, because the story is stranger than the headline.
Here’s the tangle. OpenAI has publicly called for California to strengthen its AI safety laws. In the company’s own words, “We believe the law should be amended to expand safeguards, including by requiring monitoring of frontier models under training or evaluation for potential serious incidents.” That’s a real position, on the record. Yet OpenAI has also argued that California’s AI safety bill will slow innovation. And Padilla, the state senator behind SB 300, is accusing the company of working against chatbot protections at the same time it publicly champions safety.
So which OpenAI do we believe? The one asking regulators to watch frontier models during training, or the one warning that regulation will slow things down?
What’s Actually on the Table
California is considering stronger AI safety laws in 2026, with two key bills in play — SB 300 and SB 53. The stated goal is to protect users and improve AI security measures. Padilla’s SB 300 will be one of the first pieces of legislation heard this session, and it arrives with fresh urgency given OpenAI’s decision to allow explicit content in ChatGPT.
The regulatory climate around all this is heating up too. California reported 40 data breaches in the first three weeks of January 2026, compared to 23 during the same period in 2025. Enforcement is intensifying, and the state clearly intends to act like it.
My Angle as a Reviewer
I review AI toolkits. I don’t write policy. But policy shapes every product I put through its paces, so let me tell you what this fight looks like from the testing bench.
When I evaluate a chatbot or an agent framework, the safety layer is part of the product. Not a footnote — part of the product. A tool that behaves unpredictably with vulnerable users is a bad tool, the same way a car with great acceleration and no brakes is a bad car. When OpenAI says it supports monitoring of frontier models for “conduct that could” lead to serious incidents, that’s the kind of thing I’d love to see standardized, because right now every vendor grades its own homework.
But when the same company simultaneously argues that the actual bill in front of legislators will slow innovation, I recognize the move. I’ve seen it in product marketing a hundred times. It’s the “we support standards, just not these standards” position. It lets you wear the safety badge in press releases while your lobbyists work the fine print.
Maybe that’s unfair. Maybe OpenAI genuinely believes the bills are drafted badly and its proposed amendments would produce better safeguards. Companies can hold nuanced positions. But Padilla’s public condemnation suggests the people writing the law don’t experience OpenAI’s engagement as constructive. When the legislator sponsoring the safeguards accuses you of stifling them, your “we want stronger laws” messaging has a credibility problem.
What I’d Want in the Law
From a practitioner’s seat, a few things would make these bills genuinely useful rather than performative:
- Monitoring during training, not just after launch. OpenAI itself proposed this, and it’s the right idea. Catching dangerous behavior in evaluation is cheaper than catching it in production.
- Clear rules for chatbot interactions. If a company is going to ship explicit content features, users — especially families — deserve to know what protections exist and who enforces them.
- Enforcement with teeth. California’s breach numbers show the state is already leaning into enforcement. Safety rules without penalties are just suggestions, and I’ve reviewed enough products to know how suggestions get treated.
Trust Is a Feature
My honest read as a reviewer, not a lawmaker. The fight between Padilla and OpenAI matters because trust is now a product feature. When I recommend a toolkit to readers, I’m implicitly vouching for the company behind it. A vendor that publicly asks for stronger safety laws while a state senator accuses it of undermining them is a vendor whose safety claims I’ll test twice as hard.
SB 300 and SB 53 will move through the legislature, and OpenAI will keep making its case on both sides of the argument. I’ll keep doing what I do — putting the actual products through their paces and telling you what works. Because whatever the press releases say, the product is where the truth lives.
đź•’ Published: