AI Control: usefulness-safety frontier (Greenblatt et al. 2024)
usefulness (useful outputs / total) ->
safety (1 - backdoor ship rate) ->
baseline (no protocol)
upfront audit (H)
trusted editing
untrusted monitoring
protocol roles
U = untrusted strong
T = trusted weak
H = trusted human
baseline : ship U
upfront : audit all w/ H
t-edit : T gate, H on low conf
u-monitor: U on U, H on flag
trusted editing Pareto-dominates untrusted monitoring on the Greenblatt 2024 code-backdoor task.
collusion risk: two U instances can coordinate via steganography or shared training priors.
limits: H does not scale; T-to-U capability gap caps editing signal; control is not alignment.