The most persistent mystery in public policy isn’t why problems are hard to solve. It’s why the systems designed to solve them keep growing without getting better.
Budgets increase. Staffing expands. New programs get layered on top of old ones. And yet the underlying problems — chronic homelessness, struggling schools, entrenched poverty — remain remarkably stable year after year. When outcomes disappoint, the explanation is almost always external: not enough funding, not enough political will, too much opposition. What almost never gets questioned is the design of the system itself.
That’s where the real answer tends to live.
Systems that are insulated from failure are also insulated from learning. This isn’t a moral failing or a matter of bad intentions — it’s a structural consequence. When there’s no meaningful way for a system to register that something isn’t working, errors persist, successful approaches can’t be distinguished from unsuccessful ones, and continuation becomes the default definition of success. The system doesn’t improve because nothing in its design requires it to.
Most public oversight is built around compliance rather than outcomes. Audits check whether procedures were followed. Reports document whether activities occurred. What they rarely ask is whether the underlying condition actually improved — whether the problem got smaller, whether people are better off, whether an alternative approach might work better. Legislative oversight often depends on the same departments it’s meant to evaluate, creating incentives toward affirmation rather than honest interrogation. Voters are offered narratives instead of comparisons. The result is that no one is clearly responsible for asking the most basic question: did this work, and what happens if it didn’t?
This dynamic extends well beyond formal government agencies. Much of what government does gets carried out through a network of intermediaries — nonprofits, contractors, community organizations — that receive public funding to address public problems. That’s not inherently a problem, but it tends to diffuse accountability in ways that are hard to see. When responsibility is spread across many organizations, evaluation drifts toward what’s easiest to document rather than what’s most meaningful. Government authorizes and funds. Intermediaries execute. Compliance gets measured. Outcomes become secondary. And persistence — simply continuing to exist and receive funding — becomes the closest thing to success the system knows how to recognize.
When failure becomes structurally impossible, improvement becomes genuinely optional. Expansion becomes the default response to any problem. Evaluation starts to feel destabilizing rather than useful. Over time, job security ties to persistence, funding aligns with narrative coherence, and honest assessment carries real professional risk. Nobody coordinates this — it emerges naturally from the incentives embedded in the system.
Getting out of this pattern doesn’t require a new ideology or a sweeping reform agenda. It requires something more basic: building feedback mechanisms into the design from the start. That means defining what success looks like before a program launches, not assuming it through continuation. It means scheduling real reviews at defined intervals where the options on the table include adjusting or ending a program, not just renewing it. It means ensuring that when evaluation surfaces a problem, someone actually has the authority to act on it. Information without the ability to change anything is just paperwork.
The counterarguments are familiar. Complex human problems can’t always be neatly measured. Accountability will chill innovation and compassion. Oversight already exists. These concerns aren’t entirely wrong — not everything is easily quantifiable, and crude metrics can distort as much as they clarify. But complexity isn’t a reason to abandon evaluation altogether, it’s a reason to think carefully about what to measure. And compassion without feedback is its own kind of neglect — a system that persists regardless of whether it’s reducing harm isn’t compassionate, it’s just persistent.
Systems improve only when they’re allowed to learn. Learning requires feedback. Feedback requires the genuine possibility of correction — including the possibility that something gets ended. Without that, activity substitutes for progress, and the most expensive choice remains the one we make most often: continuing something that isn’t working because we’ve never built in a way to find out.
