It depends on your goal. So, you should ask yourself first which variables you're studying, what you're doing with them, and which kind of questions you want to ask afterwards.
In general, CMIP6 data is found to be less biased than CMIP5, although there are still enough biases to warrant bias-adjusted datasets. However, both bias adjustment and downscaling can contribute to uncertainty in CMIP6, as has also been discussed for other climate model generations and set-ups. As a sidenote: in my opinion, the book by Maraun and Widmann (2018) still has one of the best discussion on uncertainty increase in climate projections. If you don't have access to the book, I've mentioned this discussion in my PhD thesis, see p.31.
Thus, what should you do? I'd suggest you ask yourself the following questions, and look up the literature for some more information. In essence, do whatever is necessary, but only that. But note that the literature is not always conclusive.
- Do you need information at the same grid level as the projections?
- No: downscaling is necessary, but the best downscaling method for you issue is another question altogether. Afterwards, check the biases according to the steps given here (note that these steps may also be in the opposite order).
- Yes: look at the biases
- Are there, based on literature of a quick data analysis, biases present for the relevant variables?
- No: great, you should be able to use the data as-is.
- Yes: go to question 3
- Are these biases relevant for your application? You should check literature on this, it always depends.
- No: great, you can use the data as-is.
- Yes: apply bias adjustment. Think - again - from the perspective of your application, and choose the adjustment method accordingly.