Let’s talk about the law that finally made the machine say where it came from.
There is this moment every writer and photographer and musician started having a few years ago. You find out your work was in the training data. Your sentences, your photographs, your voice — ground up into a model that now competes with you. And when you ask how it got there, who licensed it, who decided, the answer is the same everywhere: nobody will tell you. The model is the only product in the world that arrives without a birth certificate.
That changed. In one place, on paper, with penalties attached.
California passed the AI Training Data Transparency Act — SB 243 — in September last year, and the law is now in its first reporting cycle. From this year, developers of frontier-scale models — the ones trained on roughly a hundred million dollars of compute or more — have to document what went into the model and publish a statement saying so. Publicly. On their websites. Every year.
Not the data itself. They don’t have to hand over the dataset, and trade secrets stay secret. That was the price of getting the bill through. But they have to say what category the data falls into. Whether it contains personal information. Whether it contains copyrighted material. And whether any of it was generated by another model.
Read that last one again, because it’s the line nobody is talking about.
The machine now has to admit when it was trained on itself.
Here’s why that matters. For three years the industry has been quietly feeding its own output back into the next generation of models. It’s cheaper, it’s infinite, and it’s documented to make models go a little more hollow with every pass — less diverse, less accurate, less honest about their own limits. The researchers call it model collapse. The companies call it, well, nothing. They mostly don’t call it anything.
That’s the gap this law reaches into. If a developer has to disclose, in a public document, that a significant slice of the training data was synthetic — machine-made, not human-made — then the conversation stops being a rumor and becomes a fact you can quote.
Now, the part everyone will miss, because the headlines will be about the drama and not the documents.
Start with what the statement actually is. It’s a categorical disclosure, not a real audit. A company can say “our data includes synthetic content” and “our data includes personal information” without telling you how much, from where, or what the model actually learned from it. It’s a label on a jar, not the ingredients list. That’s still worth something — a label is worth something when there was no label before. But don’t confuse the summary with the substance.
Then there’s the enforcement question. The real lever sits in the Attorney General’s office, not in the public statement. Under the law, the AG can demand the full documentation privately — the version with the details the public never sees. So the public gets the summary, and the state gets the receipts. Which is fine, until you ask how many regulators are actually big enough to read them.
And the line that will age badly for the companies is the synthetic one, because they know exactly what it costs them. A model trained substantially on other models’ output is a copy of a copy. Its relationship to human reality is indirect. The whole public pitch of frontier AI is that it learned from the sum of human knowledge. If the fine print says a meaningful share of that sum was machine-generated, the pitch changes. This law is the first time any government has forced that admission onto paper.
So what’s the operator lesson?
If you build, buy, or rely on AI, start asking for provenance the way you ask for references. Which data went in. Where it came from. How much of it was synthetic. If the vendor can’t answer, that’s an answer. And if you’re a creator whose work may have been scraped — watch this law, because it’s the template. California proved a state can make the machine account for its origins. Everything that follows will be built on this one’s bones.
The principle underneath it is simple. Every product should be able to account for itself. A tool that shapes your credit, your work, your medical advice, your children’s homework — that tool owes you its history. Not a marketing story about its history. The history.
Because here’s the thing nobody wants to say out loud: the reason the labs fought this so hard is not that disclosure is expensive. It’s that the answer to “where did the data come from” is the one question the industry has never wanted to answer in public.
The law just made them print it.
You’ve got this. Next time a model gives you an answer, ask it where it came from. The question finally has legal teeth.