There've been a number of criticisms of Dunning-Kruger in recent years, many focusing on the sample-bias problem that Animats brings up (which is also common to a lot of attempts to reproduce it), and some also mentioning the possibility of floor effects (i.e. if you have a natural statistical variance around some quantity that is close to the bottom of what your test can measure - as if you're unskilled - then your average will be skewed upwards because for most people "there's nowhere to go but up"). Here's a relatively recent literature overview:

https://pdfs.semanticscholar.org/a387/04facf73e00523bff8182d...

tl;dr: It's complicated. Science is hard. Effects similar to Dunning-Kruger's result have been observed in multiple different domains, but because competence usually has a power-law rather than normal distribution, it's hard to say whether observed effects are due to real metacognitive skills or sampling biases. Controlled experiments are also very difficult because "training people up" inevitably results in some fraction of the sample dropping out, which also introduces sample bias.

Not scare quotes. The impression I got was, they’re not competent so not real researchers.