Admittedly, not a very high bar.
We’re building a think tank that intends to follow its advice. That takes more than asking a chatbot for a political opinion. It takes a research process that investigates the evidence, challenges its own arguments, and explains its conclusions.
Our premise is that we can build an advanced AI research system reliable enough to make its policy judgments our default. If we can meet that standard, we intend to follow its recommendations, even when they conflict with our own politics.
A convincing response from ChatGPT or Claude can still be wrong. A policy recommendation deserves deeper investigation: checking evidence, considering alternatives, and taking serious objections seriously.
We aren’t committed to producing conservative, liberal, socialist, or libertarian answers. Policy also involves judgments about fairness, freedom, and whose interests count. Those judgments should be explicit.
The system’s conclusions may change, especially while we’re still building it. When they do, we’ll revise our positions with it. We hope to see convergence as the research develops, but agreement alone does not establish that an answer is right. Every institution should be capable of revision. Human institutions change their minds too, whether they admit it or not. We intend to do so openly.
Our research system is still in development. These samples are provisional, and their recommendations may change substantially.
What should the federal government do to better support children and families?
Read the researchWe’re building an AI research harness: software that gives AI the structure to carry out sustained policy research. Our goal is work that is significantly deeper, more insightful, and more rigorous than any chatbot response.
It should also be more epistemically aware: explicit about the evidence behind a claim, the assumptions it depends on, and the uncertainty that remains. We want recommendations that explain both their reasoning and what could change the answer.