Ownership concentration × Alignment/assurance

Every model update becomes something like an election — broad ownership puts a crowd's hands on a single steering surface, and every consequential correction has to run the gauntlet of contested governance. That is productive but slow, and the frontier migrates to wherever someone can just decide. Audit capability also doesn't distribute automatically with equity: if only a few can interpret the interpretability output, a priesthood recentralises what ownership spread. Still, it's the only corner where correction serves the corrected, and where failures become cumulative public learning rather than private scandal — a saddle worth defending, not a basin to relax into.

Safety research becomes the R&D arm of control — the better corrigibility gets, the more total the concentrated owners' grip, and it wears the white hat the whole time. An opaque system resists even its owner; a corrigible one obeys perfectly, so mature alignment technology in a few hands produces not safety but the cleanest available tyranny. The audit trail legitimises the cage: every audit passes, the system demonstrably keeps its promises, and all accountability runs to the owner. One thin opening: because the system is corrigible, the entire political fight collapses to a single legible question — who owns the switch.

The ownership structure of the 2008 financial crisis, applied to the systems running everything: no one is at the wheel, and risk is socialised across a diffuse public with no ability to prevent the failure it will pay for. De facto control flows to whoever interfaces with the machine day to day — operators, maintainers — while formal ownership stays broad: you can democratise the rails and still have zero collective ability to steer them. The one perverse upside — diffuse opacity blinds everyone equally, so no single owner captures the exercisable power the corrigible-and-concentrated corner hands its owner.

Opacity blocks the powerful first — and hiring their way out doesn't help. Concentrated owners can buy operators, but an uninterpretable system means they can't tell whether those operators actually understand it or are faking control to secure their position; the owner ends up captive to an unfireable technical priesthood while holding all the catastrophic liability. Extraction beats maintenance, so fragility accumulates precisely where the money is, and when it fails — correlated, monoculture, ungoverned — there's a legible villain who can neither explain nor repair it. The redeeming instability: a collapse big enough is the single likeliest event to break feudal capture entirely.