At the core of my work is a simple question: when does any intelligent system, whether human, machine, or any collective combination of agents, truly have enough knowledge to act, and when is it safer to hold back?
I call this the epistemology of justified action: the study of what counts as “knowing enough” to move forward versus when abstaining is the responsible choice.
And I pair it with pluralistic alignment: the idea that safe systems must protect diversity in strategies, perspectives, and values until there is principled convergence, instead of collapsing everything into one fragile answer too early. This holds whether the system is AI-AI, AI-human, human-human, or even human-animal. The architecture does not matter; the epistemic question does.
The most dangerous failures of intelligent systems are not only technical errors. They are failures of epistemic discipline: acting confidently when the grounds for action are not actually there. In safety-critical contexts such as medicine, law, governance, or driving, unjustified action does not just create mistakes. It magnifies uncertainty, silences dissenting perspectives, and locks in errors that no amount of later optimization can undo.
Traditional approaches lean on heavy planning: rolling out futures, simulating partners, or solving complete models. But in open and multi-agent settings, that is infeasible. The space of possibilities is too large, behaviors too unpredictable, and wrong assumptions too costly.
So my work develops planner-free epistemic checks. Instead of simulating everything, agents test whether their current abstractions preserve distinctions that matter, and whether their internal perspectives have stabilized enough to justify action. Planner-free here doesn’t mean reckless. It means acting only when there is genuine epistemic sufficiency.
Across these cases, abstention isn’t about inaction — it’s about preserving the option of safe coordination until genuine convergence emerges.