Superintelligence is only a worthwhile ambition if it serves humanity. We are interested in how increasingly capable systems can remain accountable to people, respect meaningful boundaries, and support human flourishing.
Questions we’re interested in
- Which evaluations reveal failures that ordinary demonstrations miss?
- How can human oversight remain meaningful as systems become more capable?
- How should we evaluate who benefits, who bears risk, and whether safeguards work?
Our starting point
Our intended approach combines capability evaluation, alignment research, misuse analysis, adversarial testing and explicit deployment limits. We want human agency to be a concrete property of a system’s behavior, not just a promise about its purpose.
What this means today
This is a research direction and statement of intent. It is not a completed safety evaluation, certification or evidence of a deployed frontier-model safeguard.
Our safety approach