Yes. The simple corollary to Bayes theorem: all observations, even biased ones, improve your knowledge of the world.
But it doesn't say anything about efficiency. A single photodiode mechanically scanned behind a pinhole can capture the same image as an image sensor behind a lens, but it takes millions of times longer. A self-reporting survey will say something, but it will be very noisy. How many self-report surveys would it take to match an actual double-blinded study with double digit participants? What if the answer is "hundreds"? Should we wait a century for all these surveys to be done before even attempting data analysis? (After all, surveying the same set of people over and over won't give you additional data. You need to wait for some of the survey respondents to die) What about all the people in the preceding 99 years who, perhaps reasonably, look at a stack of fifty self-reported surveys that all say the same thing, and come to an incorrect conclusion? Science is right eventually, not right right now.
You say that "millions" of participants should be useful, almost automatically. Why should that be the case? If you're off by an order of magnitude somewhere in your stack of assumptions, the data could be noisy enough that you need to be sample from a population of a hundred billion. If the data is bad enough or biased enough then even a large survey wouldn't be powered enough without a world population ten times bigger.