> The total numbers of Twitter followers they analyzed during the election campaign was 838,026, which, assuming the claimed “social bot” prevalence of 9.9%, would correspond to roughly 83k “social bots”. However, the paper does not provide any examples of “social bot” accounts and the raw data was not shared. When we contacted the authors, they were unable to provide us with a single credible example of a “social bot”.
> To summarize the result, not a single one of the accounts in the list was a “bot” in any meaningful sense of the word, certainly not a “social bot”. In other words, every single “bot” in our sample was a false alarm; the precision (positive predictive value) of Botometer on this task is 0%. Assuming that Botometer does not perform worse than a random number generator (as we did not check the accounts rated as human), we can also conclude that the best guess for the number of “social bots” following the political parties in Germany is not 83,000, as claimed by Keller and Klinger, but zero.