Useful proof of work sounds good on paper, where does it usually break?

I get why people like the idea. redirecting energy from pure hashing toward something society can point at and say that had value.

But in practice, the weak points keep recurring. Verification is costly or subjective, incentives drift toward gaming metrics rather than producing real value, and quality is hard to measure without centralized judgment.

Over time, systems either reintroduce trust, collapse into low-value outputs, or become economically uncompetitive versus simpler alternatives.

For people who've looked at this category longer, where does useful proof of work usually fail in practice?