USIAIS was created to do three specific things. To publish research on AI safety that any operator can use. To advise organizations through the practical work of deploying AI without ignoring its risks. And to issue independent assessments of AI products and the organizations that build them, against published standards anyone can read.
Each of those activities is a discrete program, but they share a single discipline: every conclusion we publish, every assessment we issue, and every piece of advice we offer is anchored to a methodology that exists in writing before the work begins. We do this because the field is moving faster than its institutions, and because the institutions that should exist will only become trustworthy by behaving as if they already are.
What this is, and what it isn't
USIAIS is not a research lab. It is not a foundation. It is not a non-profit. It is a for-profit research and advisory organization that generates revenue from advisory engagements, consulting projects, certification fees, and corporate membership. The choice to be for-profit is deliberate. Independent work, done seriously and durably, requires being paid for it. The discipline that prevents the people who pay us from compromising our research is policy, not preference. That policy is published in full and binds us.
USIAIS is also not in the business of declaring AI dangerous or safe in the abstract. Those are not useful conversations. We are in the business of asking specific questions about specific systems in specific contexts and giving honest answers. A product may be safe in one deployment and dangerous in another. An organization may operate one model responsibly and another carelessly. The work is contextual, which is why the methodology exists.
The work ahead
We are starting with three programs because they are the three places where careful, independent work is currently most undersupplied.
Research is undersupplied because the institutions that should be doing it either do not exist or are constrained by funding structures that compromise their independence. Advisory is undersupplied because the people who know how to do this work are concentrated inside the AI labs themselves. Certification is undersupplied because there is no widely-accepted independent standard against which AI products or organizations can be assessed.
This will not be solved by USIAIS alone, and we should not be the only institution doing it. But we will do it as if the work matters, because it does.
Charlie Marucco Director, USIAIS