Figure 6

Twitter’s tampered samples. (A) Sampling artifact created by a time-triggered frequency bot. For several hours, we sent a Tweet exactly every 42 seconds. The bar-graph shows the distribution of milliseconds of the Tweets’s ID. Most Tweets are in a very narrow range, none is in the 1% Sample API time window. (B) Schematic representation of Tweet selection process for time-triggered accounts (x-axis). Accounts send Tweets (circles) over time (y-axis). Statistically sound sampling would give every Tweet from every account the same chance of being in the sample. Instead, Twitter’s sampling mechanism selects a large amount of Tweets from about 1% of time-triggered frequency accounts for any time period