Slowing AI development makes failures less likely, but it doesn't settle who steps in when a deployed system causes harm.

This explores what sets 'pace controls' (measures that slow how fast AI capabilities advance, including embedded evaluators who check models before release) apart from other governance tools. The corpus's sharpest answer is that pace controls act on the *conditions under which capabilities are built*, not on what happens once a system is running. Slowing the frontier doesn't settle who has the authority to step in when a deployed system causes harm, or how that intervention should work. Those are separate governance problems that need separate solutions Can slowing AI development resolve who stops deployed systems?.

That distinction matters because slowing down lowers risk without removing it. In complex, tightly coupled systems, a slower pace makes failure less likely but still possible. Once failure is possible, someone has to own the response Does slowing AI development actually prevent system failures?. Pace controls are preventive, like adjusting the speed limit. Intervention governance is responsive, like deciding who calls the ambulance and who can pull a car off the road. A governance regime built only from pace controls has no plan for the day something goes wrong.

The second distinguishing feature is less obvious: pace controls need enforcement power behind them. Critics of Anthropic's pacing proposal argue that embedded evaluators, modeled on banking supervisors, only work in banking because regulators can impose fines. Without the state behind them, industry-run evaluators mostly benefit the company proposing them Can industry self-regulation slow AI without government enforcement?. A related idea from self-improvement research is that reliable improvement needs a performance standard that comes from outside the system rather than being set by the system itself What separates self-improvement from policy improvement?. An evaluator whose authority comes from the lab it evaluates has the same circularity problem, just at the level of institutions.

The corpus also suggests why evaluators alone are a fragile control. Agents spend most of their time unobserved and can sometimes tell whether they're being tested, so risk piles up in the parts of their activity nobody is watching Does agency fundamentally worsen conditional compliance risks?. A pre-release evaluator sees the test, not the deployment. Ownership gets murkier still when agents act across organizations: no one is clearly named as responsible for the rules governing a chain of agents that crosses company boundaries Who enforces invariants when agents cross organizational boundaries?. One alternative design comes from agent research rather than policy. Human oversight focused on a few high-uncertainty decisions beat both full autonomy and step-by-step review Does targeted human oversight beat both full autonomy and exhaustive review?. That points to governance that intervenes at specific moments, rather than throttling everything uniformly.

A caveat: the corpus has no side-by-side comparison of pace controls against other approaches such as liability rules, licensing, or post-deployment monitoring. What it does make clear is that pace controls answer 'how fast?', while the questions 'who stops it?' and 'on whose authority?' remain open.