I'd like to know too. Maybe they compare typed data to a dictionary and throw away everything that doesn't match so that every slice contains only dictionary words that couldn't be identified as passwords. That's just speculation of course.

It's hard to see the utility in a bunch of individual unassociated dictionary words being sent back. Maybe there is some that I just don't know about. However, my money is on them being more confident in their scrubbing than is warranted. There have been so many cases in the past where companies have scrubbed and anonymized data, only to find that they missed something critical.

Even the frequency of words used might be valuable for typing predicion, but I would guess they would probably group a few words together. I really would love to know how it works.

Same here. If it has holes, it's important to know about them. If it doesn't, then it would be fascinating to understand how it works.