The ultimate goal of statistics is to test the hypotheses. As you rightly noted, non-parametric tests are largely used when the data is not normally distributed. This means that either the skewness and kurtosis are beyond the acceptable norms. Under such conditions, the use of parametric tests such as t-test and ANOVA are not very powerful or accurate. Therefore, non-parametric alternatives such as the Kruskal-Wallis test are used.
I am a researcher in the field of social sciences and I have published two papers in scopus indexed and web of science indexed journals, one of which is by SAGE and the other by Routledge: Taylor and Francis, where I have used non-parametric tests and I have provided my justification for using them as well.
Article Wellbeing of Palliative Care Workers During Covid-19 Pandemi...
Article Online learning experiences of social work students in India
The knee-jerk reaction of switching to non-parametric tests when the assumption of normal distribution for the variable of interest is obviously unfitting is a common, old-fashioned, and a statistically rather illiterate reaction. It would be more appropriate to go for a better understanding of the data-generating process and to find a more appropriate statistical model allowing to test the relevant hypothesis. You wrote:
"non-parametric tests are usually performed to test the relevant hypotheses"
This may be the intention of the researcher, but this is not what is usually provided by the non-parametric tests. Most of the non-parametric procedures are based on the analysis of ranks, what translates to a comparison of distributional shapes, at best with a relatively high sensitivity to detect location shifts, besides detecting other differences. It seems to be largely ignored that tests like the Wilcoxon-Mann-Whitney U-test is testing the hypothesis of stochastic equivalence, which implies a location shift only under strict additional assumptions which are very clearly unreasonably in almost all practical situations (namely that the distributions are identical in all groups except for the location, so they must all have the same variance and the same kurtosis and all higher moments must also be identical).
Thus, applying an U-test instead of a t-test means to change the hypothesis being tested, demonstrating stochastic non-equivalence may just not be useful for the problem at hand, where typically a conclusion should be drawn about which of the populations has the higher (or lower) expected value.
This is different for non-parametric tests based on a resampling/bootstrapping procedure to infer the sampling distribution under the appropriate null hypothesis. These procedures are rarely used and they have only very little power for small sample sizes. They are also more prone to project any bias in the sample selection affecting other characteristics but the mean into the result.
So just to answer the question in your title: yes, unfortunately!