Policymakers should encourage this kind of transparency, and advocates should demand it. The full report details how to make this transparency work: cdt.org/insights/tun...

AI companies should also disclose select information about their safety-related training data, which we detail in the report. This information provides crucial context about how the rules contained in a model specification are translated into practice.

Some model specifications are better than others — for instance, OpenAI’s is admirably detailed, while Anthropic’s is more skeletal. (Anthropic’s might also be out of date.) We explain in more detail what model specifications should look like in the full report.

This is why AI companies need to tell us the rules they train their models to follow. A document describing these rules is called a “model specification” — OpenAI published one recently, and @anthropic.com did 2 years ago. Other companies should follow their lead.

Without seeing those rules, scientists can’t accurately diagnose AI incidents. Sometimes models are just following (flawed) orders, but sometimes they’re jumping guardrails. The difference matters: If we can’t tell when AI’s breaking its rules, we can’t make it safer in the future.

But most AI companies keep the rules they train their models to follow private. That forces the public to play whack-a-mole with individual incidents while blind to the underlying rule. It’s like trying to fight a law you’re not allowed to read.

When AI misbehaves, it’s often just following orders — the rules it’s trained to follow have flaws that lead to unintended harm. (Think of Grok promoting antisemitism or “white genocide” theories, for example.)

cdt.org/insights/tun...