Access to capability × Alignment/assurance

Safety here isn't a property of the system so much as a credential for market entry — every downstream tool has to prove it belongs inside the licensed perimeter before it's allowed to exist, which turns assurance into the language of gatekeeping rather than the practice of it. Because correction and rollback power concentrate with the same few actors who hold access, the public may never see which behaviours got quietly edited or redirected — safer in the narrow technical sense, while accountability for what "safer" means stays invisible. Strong official corrigibility doesn't eliminate demand for the uncorrected version, it creates a black market for it: banned, embarrassing, or merely non-standard uses migrate to whatever runs outside the assured channel.

The standard fear about open access is that bad actors strip the guardrails; the likelier opposite is this: if correction and inspection are genuinely available to everyone, proliferation becomes the mechanism by which control scales rather than the thing that defeats it. A decentralized patch culture takes over the job a single regulator can't do fast enough — community-issued patches against newly discovered exploits, local audit tools that catch stripped-guardrail models before they spread. The catch nobody advertises: maintenance burden and liability democratize along with the control, so small organizations that used to say "only the vendor could have known" now own that duty too, ready or not.

A few actors owning power they can't fully inspect produces governance that looks like nuclear containment, not software deployment — layers of insurance, paperwork, and sandboxing around an engine nobody can completely explain. Because internals stay opaque, safety knowledge stops being interpretability and becomes folklore: stress-test specialists and red-team rituals function as an empirical priesthood the gatekeepers depend on without ever really understanding what they're managing. The one real mercy is blast-radius containment — concentrating access limits how many places an undebuggable failure can happen, a thin argument for the corner but not a worthless one.

Widely available systems that work often enough to get adopted but fail in ways nobody can explain don't produce visible chaos so much as reliability theatre — a market of wrappers, reputational badges, and "AI-safe" stickers that certify usage patterns instead of understanding. The failures that matter hide inside apparent diversity: thousands of ostensibly independent apps share the same opaque training data or base behaviours, so the ecosystem looks plural right up until one blind spot propagates through all of them at once. What survives as a safety valve is expensive and regressive — those with money retain human experts and paper trails, while everyone else stays inside the brittle automation because it's what they can afford.