By analyzing a large amount of tweets, researchers have found a way to improve the detection of these non-human users.
Bots are all over the Internet and especially present in numbers on Twitter. It seems very difficult to measure their real impact, especially in political matters where they are widely used. However, they could play a role in the formation of public opinion by increasing the presence of certain online views at the expense of pluralism.
Social networks are of course hunting these non-human users and millions of accounts are regularly deleted. To be able to recognize these non-human, artificial intelligence can be a very valuable tool. This is precisely the subject of a study published by researchers at the University of Southern California.
These discoveries could help to perfect existing tools
To fine-tune their algorithm, the scientists chose a very specific angle: trying to understand what differentiates the behavior of humans and bots on Twitter. The team worked on a very large sample of 8.4 million tweets from 3,500 accounts managed by normal users and 3.4 million tweets written by 5000 bots.
Several findings emerge from their work. First of all, it appears that humans are much more likely to interact with other users of the platform than robots. They also note that the length of human publications decreases when they spend a lot of time en online,probably because of fatigue.
Nothing like this for bots for which there is no similar evolution. In the latter, there is also a regularity in the time between the publication of tweets. They very often . post every 30 or 60 minutes.
With these teachings in mind, the scientists decided to integrate these parameters into the existing detection AI under the name Botometer. They then noted that the tool was much more likely to detect fake accounts than it was before.
According to the authors, this is therefore an enhancement that could be used by all the detection tools already in place. However, they recognize a key limitation to their work: the tweets analyzed go back three years. It is thus possible . that the bots have in the meantime been improved.