I pointed this out in a sibling, but 10% would also include automatically reviewed tweets for misinformation, covid, and any other language filters they have in place.

Also, having worked for a company that did ML sentiment analysis and content analysis against the twitter firehose, the accuracy of ML was closer to 65% than it was 99%. Yes, the company had a huge in-house crew dedicated to checking that 35%, and monitoring twitter manually for anything missed.