A new study from researchers at The University of Manchester investigates the sophisticated network of agents on Twitter who work to distribute fake news during election campaigns.

The paper – “Political Astroturfing on Twitter: How to Coordinate a Disinformation Campaign” – was published in the journal Political Communication. ‘Astroturfing’ comes from the way in which agents appear to be part of a genuine grassroots movement, when in fact, they are part of an orchestrated and centrally managed campaign.

Using court records from a case in South Korea, where the National Information Service (NIS) was caught trying to influence the 2012 presidential election, the researchers were able to identify 1,008 Twitter accounts controlled by NIS agents. They then examined the patterns of interaction between these known ‘astroturfers’, discovering that they showed clear traces of coordination.

This coordination emerged from what University researcher Dr David Schoch calls “the principal agent problem”. He says; “The campaign organizer (or "principal") wants things done a certain way. For astroturfing campaigns, this means that the agents should try to appear as if they are part of a legitimate grassroots campaign. The "agents", however, may lack the motivation to do so and try to cut corners to please the organizers.”

Their desire to speed up the process leads to the agents copy and pasting messages across several accounts, which allowed the researchers to look for tweets with exactly the same content, which were posted in a short time window.

“In summary,” says Dr Schoch, “the coordination patterns we looked for are two accounts posting the same tweet within a short time window, and two accounts retweeting the same tweet within a short time window.”

A key point which makes this paper notable is that it focussed on networks of real, human agents, rather than the automated ‘bots’ which we usually think of when we talk about fake news on Twitter. In fact, the results of the study show that networks of human actors actually display more coordinated patterns than bots.

The authors also had the NIS data to use as a “ground truth” dataset, says Dr Schoch. Otherwise, they would “have to rely on detection algorithms being 100% accurate, which of course is wishful thinking!”

Applying this information to the Twitter data obtained from the court case, the researchers were able to detect a further 921 suspect accounts likely to be involved in the NIS campaign.

They are also investigating previous campaigns, Dr Schoch adds; “We are currently studying around 10 more recent campaigns around the globe to see whether these coordination patterns can also be observed. Preliminary results suggest that this is indeed the case.”