February 25, 2015 • Staffan
The following analysis was done in collaboration with Sky News.
The recent discussions about ISIS’s activity on Twitter and the US countermeasures led us to look into some statistics from Recorded Future’s index. Looking at the number of references to ISIS in Twitter versus other media since January 1, 2014 shows a staggering increase in volume. Even recent growth is big: There are an average 250% more ISIS-related tweets per day in February 2015 compared to December 2014.
Note that general media coverage of ISIS has actually gone down since the August-October period, whereas Twitter volume is booming (e.g. with more than 1.2 million references on February 3). How are we to interpret this increase, and what characterizes the tweets being sent?
First, let us split the tweets by language; as can be seen by this chart, there has been a shift starting in September 2014 from English being the dominant language discussing ISIS to Arabic having the largest volume. The big growth of total volume in 2015 is almost entirely due to Arabic tweets:
This shift becomes even more obvious if we chart the percentage of the total discussion per language:
The discussion is thus today clearly dominated by Arabic — but please note this in itself does not tell us whether the tweeters are for or against ISIS, just the conversation has clearly shifted to Arabic from English. As we can see in this chart, the two major spikes in February are closely related to ISIS activity in Jordan and Libya:
Another way to characterize the discussions is to look at sentiment. Recorded Future computes several sentiment metrics, including positive, negative, and violence. We have previously suggested using “positive AND violence” as a way of trying to identify pro-ISIS Twitter accounts. This metric is quite crude, especially since tweets are very short and hard to compute sentiment for, and also since it does not take into account, for example, irony and jokes.
Furthermore, this metric could actually be measuring both a positive attitude towards ISIS’ use of violence and a positive attitude towards violence against ISIS. Our preliminary analysis suggests indeed the biggest volume increase here derives from accounts that are against ISIS.
However, we believe by using the same metric on very large number of tweets and comparing the same metric over time, the ratio of “positive AND violence” tweets to total volume is a relevant metric showing some shift in attitude. The following chart shows how this metric has been fairly stable from June 1, 2014 and up till the end of January 2015, when it increased significantly (close to a factor of six on a 14 day moving average):
This trend is even more obvious when looking just at tweets in Arabic with positive and violence sentiment:
In English tweets, there is no such change in attitude over time:
One more aspect worth looking into is to what extent the huge volume of tweets is driven by bots. Differentiating a bot from a human tweeter is not trivial, but looking at volume of tweets per day and the frequency of tweeting by a certain Twitter account is a fairly good indicator. Based on this, we could identify several bots, however no single one contributing more than 5,000 tweets per day. Manual inspection of some of these accounts also showed they represent both sides of the conflict. Several of the accounts we classified as high volume and pro-ISIS have been suspended by Twitter.
Looking at an average number of tweets per account per day during 2015 shows most of the time this number is around three, probably indicating the bots are averaged out by (probably) human low volume tweeters. We can see two spikes in January where the increase in total volume appears to be triggered by a smaller number of accounts sending large volumes, however the two large volume spikes on February 3 and 16 were accompanied by an increase in the number of active accounts, and thus shows no signs of high volume bots driving the traffic alone:
Above: Number of unique active accounts (blue curve) and number of tweets (green curve) on left y-axis and average number of tweets per author (red curve) on right y-axis.
Traffic could of course be driven by large numbers of low volume bots. We found one example of a tweet “meme,” containing the text “الله يعين ويحفظ رجال الامن والعسكري كل مٍْـٍْ♥ـنٍْ سهر علي حفاظ البلد بعد الله #اعتقال #داعش #الدولة_الإسلامية” (“God help and preserve the security and military forces who stand vigil for the preservation of the State after God #Arrest #Daash #Islamic State”) plus some additional texts and images of non-ISIS soldiers; this occurred in more than 470 thousand tweets (example below) and is in support of Saudi security and military forces.
We decided to investigate this campaign in detail, and found it started on January 28. Here is the number of tweets per day – note the significantly lower number of tweets on Fridays:
We also looked at the number of active accounts per day for this campaign, and the average number of daily tweets per account:
In total, since the campaign started, 2,463 accounts have been active, and on no day has any account tweeted more than 104 tweets. This, and the low activity on Fridays, gives this campaign the clear appearance of a manual campaign engaging a fair number of accounts to create an impact.
Given the big increase in volume of Arabic tweets from January to February (up 60% per day, on average), it is interesting to see where, geographically, these tweets originate. Unfortunately, extremely few (about one in 3,000) of all tweets are geotagged by where they are actually tweeted from, but of the January tweets, 10% are tagged with the author’s home location (and 19% of the February tweets are tagged that way).
Given this information, we can see the biggest contributing country is Saudi Arabia, whereas the biggest increase (in percent) is to be found in Egypt, where we observe five times as many tweets in February. Maybe this is reflecting an increasing interest in ISIS in new countries as their operations spread throughout the region?