Researchers from the United States have presented an artificial intelligence (AI) model that can independently find messages from Internet trolls. Its accuracy is 91%.
Researchers have developed an automated machine learning system, a type of artificial intelligence (AI) that can find messages from internet trolls based on their content. It does not require human intervention to work.
The team notes that their approach differs from simple bot detection – the operation is much more complex. The model takes into account the volume of posts, their content, and keywords. Over time, she builds up a sample troll template and finds publications more efficiently.
In Science Advances, the scientists noted that they tested their model on four social networks. They analyzed publications from December 2015 to December 2018, as well as from July 2015 to December 2016 – periods when “trolls tried to influence the results of the US elections”.
After training the system on a subset of the data, the team noticed that AI can not only find troll posts but also determine their hierarchy, as well as calculate accounts on other social networks. However, in some cases where the trolls are not using a single script, the model was unable to find them.
The team also found differences in performance by language – for example, the activity of Chinese trolls is easier to spot than trolls from Russia. The model was able to identify the Venezuelan trolls with an accuracy of 99%.
The average AI score was 90-91%. However, the researchers are confident they will improve this result in the coming months.