Not likely helping with the training- unless I'm underestimating the scale of people visiting the site. Also unlikely because they had to have a well trained model to launch the site and make the YouTube video that everyone has seen. It's possible that they will do fine tuning in certain areas depending on if some generated clips get consistently answered the same way.

More likely is that you're helping them write the results section of their research paper, with a sentence like this, "In a blind trial of 5,000 internet users, over 93% of people were unable to tell the difference between generated audio and real audio at a statistically significant level (P<=0.035)."

Note: if we assume a binomial distribution, you need to get 7 or 8 correct to reach the magical P<=.05 barrier. If you assume that there are a fixed 4 generated clips and 4 real clips, then you need to get them all right (https://en.wikipedia.org/wiki/Fisher%27s_exact_test). I think it's fair to use the binomial distribution because the website does not tell you the number of real and generated clips before you take the test.

I don't think so, at first glance it looked all client-side.