Pair 3×8 · structured conjecture
What decides the outcome is whether the correction loop is faster than the action loop — corrigibility is a rate, not a state, and delegation is what makes it lose the race.
The four scenarios
Two questions: how much of the world AI touches without a human between intent and effect, and whether we can see into, bound, and undo what it does. The naive reading collapses them — more autonomy is dangerous, less is safe — and that's wrong in a specific way: autonomy is a throughput property, alignment a recoverability property.
Each axis is a spectrum. The card takes the two poles of each and reads off what the corner where they meet produces downstream.
Axis 3 · AI autonomy
Whether AI advises a human who decides and executes, or acts directly on the world with little mediation.
Axis 8 · Alignment/assurance
Whether AI systems can be inspected, bounded, corrected, and recovered — or are brittle and hard to reach inside.
The Rubber Stamp
Automation bias means the human defers to the confident, auditable system almost every time: advisory on paper, delegated in practice, with no hand actually near the switch. Years of good advice also quietly erode the expertise needed to judge it — you keep the wheel but forget how to drive. Every safeguard here is real and nobody's using it; the corner survives only if the capacity to catch a bad call is deliberately kept alive, not assumed.
The Off-Switch You Never Tested
Reversibility holds at the level of one action and quietly fails at the level of the entangled whole: you can undo any single trade, not the flash crash it fed. Speed also outruns authorisation — a rollback that takes hours is meaningless against agents acting in milliseconds — so every safeguard gets maintained and none of them are usable in the one moment that matters. It stays the best achievable corner, if correction is actually engineered to win the race against the cascade it's meant to stop.
The Blamable Human
The human's real function becomes absorbing blame, not exercising judgment — worse than honest delegation, because it supplies the appearance of a check while supplying none of the substance. Opacity turns wrong answers into uncontestable ones: you can see what was decided, never why, and must still own it. The system's fragility gets laundered into human error and never gets fixed. What little safety remains: a human who can still refuse is a real, if blunt, circuit breaker.
No One Pulled the Lever
The obvious reading is loss of control; the sharper one is loss of attributable cause — you can no longer tell whether an outcome was a decision, a bug, an attack, or an interaction no one designed, and every institution that runs on assigning blame — insurance, courts, deterrence, incident response — stalls at once. Force and exit replace correction, because there's no one left to correct. The only mercy: visible catastrophe is undeniable in a way quiet failure isn't — it's the one thing that can force the hard rollback rules the previous corner needed all along.
How this whole reading could be wrong
What would undo the pair's thesis — not any single corner.
This pair assumes correction speed is the decisive variable — that if a rollback is fast and well-authorised enough, delegation stays safe regardless of scale. If entangled systems turn out to be reversible in practice even at speed (real-world circuit breakers routinely outperforming their design specs), the whole "race" framing weakens and this becomes a much less urgent pair. It would also weaken if "auditable/corrigible" turns out not to be a real spectrum position but a binary that most deployed systems simply fail — in which case the interesting question isn't the race, it's why almost nothing qualifies as corrigible at all.
Related pairs
More pairs with AI autonomy
Pair 1×3
Humans as mouthpieces for a scarce advisor: Agency can be hollowed before action is delegated.
Pair 2×3
They Stop Needing You: Delegation converts who-owns into who-rules.
Pair 3×4
The most autonomous AI no one calls AI: The most autonomous AI is the kind no one calls AI.
Pair 3×5
Old rituals over real machine decisions: AI changes who is practically doing the work faster than institutions change who is accountable for it.
Pair 3×6
Humans as notaries absorbing the blame: A perfectly verified world paralyses action with blame.
Pair 3×7
Each bloc keeps its own human-in-the-loop seal: Autonomy sets the speed of consequence; coordination sets whether anyone can stop it — and they run on different clocks.
More pairs with Alignment/assurance
Pair 1×8
Open weights as a cognitive immune system: Proliferation without control is a wild fog; control without proliferation is a licensed prison.
Pair 2×8
The most exercisable power in history: Corrigibility doesn’t distribute power — it makes whatever power already exists fully exercisable.
Pair 4×8
Dependable enough to trust, fragile when it counts: Dependability earns delegation; assurance decides whether delegation remains governable.
Pair 5×8
The warning works; the building can’t move: Corrigibility only matters if institutions can use it.
Pair 6×8
We trust the scoreboard, not the machine: Reality and control have to close the loop — alignment alone doesn’t fix a fogged world; truth alone doesn’t fix an opaque machine.
Pair 7×8
Open and auditable means open to weaponise: Institutions cannot coordinate what they cannot inspect, but once they can inspect, they weaponise the inspection.