The scientific method is much more then just sampling data and correcting the model afterward. As I explained before you need a model that includes all the actions based on the predictions the model provided before processing said feedback information. If you don't you can reinforce false predictions based on initial wrong assumptions to become true over time! And for this to be effective you have to split identified groups. Double blind and act differently on each of the previously grouped subgroups (to create combination's and get enough data and eliminate noise).

This all is of course is impossible in human to human interactions like these. But if we leave that fact out and assume a company to do the tests scientifically you run into other issues. People need to have the idea that they are actually doing real work, so simulated companies don't work! To allow for meaningful data you need a large real company and huge sums of money that acts like a profit motivated entity but at the same time is willing to make a substantial loss during the scientific tests. It can backfire after all, or they can't downsize during hard times without compromising the quality of the data.

I very much doubt that Google can incorporate its own decisions back into the model along with the feedback. I even doubt that in human to human interactions it is possible to classify every important interaction in such a way you can use the data reliably (maybe in a lab with hidden observers and cameras, but not a regular work floor).

See a pattern to the science argument here?

Numbers end up given a meaning based on to little and much too tainted data with initial assumptions on top. And then a human is going to use these number for further processing..it is insanity IMHO.