[A]lmost no innovative programs work, in the sense of reliably demonstrating benefits in excess of costs in replicated RCTs [randomized controlled trials]. Only about 10 percent of new social programs in fields like education, criminology and social welfare demonstrate statistically significant benefits in RCTs. When presented with an intelligent-sounding program endorsed by experts in the topic, our rational Bayesian prior ought to be “It is very likely that this program would fail to demonstrate improvement versus current practice if I tested it.”

In other words, discovering program improvements that really work is extremely hard. We labor in the dark -- scratching and clawing for tiny scraps of causal insight.

Megan McArdle quoting or paraphrasing Jim Manzi.

[Edited in response to Kaj's comment.]