Existential targets the gradual-disempowerment pathway to AI catastrophe: the scenario where humans progressively hand their reasoning to AI systems, individual and collective judgment atrophies, and meaningful human oversight of increasingly capable AI erodes without any single dramatic failure. Keeping humans epistemically capable and independent is a lever on that pathway, and it is under-resourced relative to alignment and evaluations.
Most consumer AI pushes the wrong way. It delivers finished answers, trains users to defer, and concentrates the underlying cognition inside a few frontier labs that also hold the user's data. Existential is a concrete counter-example at the tool layer. It strengthens the user's own reasoning instead of substituting for it, and it runs entirely on the user's device, so the capability neither depends on nor feeds a frontier model provider.
We are calibrated about scope. This is not alignment research, interpretability, or dangerous-capability evaluation. It is an AI-for-epistemics intervention: building and testing whether a tool can measurably improve how well individuals reason and retain authorship over their own conclusions. The 90-day deliverable produces early evidence on that question, useful to the wider epistemic-security agenda regardless of whether Existential itself scales.