Social media bots can distort democratic debate

Bots can influence democracy without spreading false information. A new study from the Swedish Defence University develops a framework for understanding how automated accounts can create a misleading impression of which political actors and ideas enjoy public support.

Social media bots – automated accounts that can mimic human users and, for example, publish or share content – are often associated with disinformation and foreign influence operations. However, they can undermine democratic debate even without spreading false information or being controlled by a foreign power. This is demonstrated by Linus Wahlberg, a doctoral student in political science at the Swedish Defence University, in a new study examining discussions on Twitter, now X, ahead of the 2022 Swedish general election.

"f we only look for false information, we risk overlooking other forms of influence. One problem is that bots can make certain opinions appear far more widespread than they actually are. They can create an exaggerated impression of support for or opposition to a political idea, party or politician. This can distort our perception of political debate", says Linus Wahlberg.

A framework for understanding automated influence

In the study, Linus Wahlberg develops an analytical framework based on the concept of misrepresentation. This describes how automated accounts can distort perceptions of support for or opposition to political actors and ideas.

The framework is based on four principles of democratic debate: proportionality, transparency, pluralism and opportunities for participation. For example, bots can make a particular position appear more widespread than it actually is, obscure who is behind a message or contribute to a climate of debate that discourages people from participating.

The framework complements established concepts such as disinformation and misinformation by focusing on how automated activity can influence our perception of political debate as a whole.

The study demonstrates how the framework can be applied

To test the framework, the study analyses more than 85,000 election-related posts published during the six weeks leading up to the 2022 Swedish general election. Approximately 15 per cent of the active accounts were assessed as likely to be automated. These accounts were responsible for around 22 per cent of all posts in the dataset.

The study identifies clear patterns in the content published by automated accounts. Among the accounts that attempted to influence perceptions of political actors, a large majority focused on criticising the 'red-green' bloc. Their methods included the use of strongly negative language, images and memes.

However, identifying bots involves a degree of uncertainty. Classification is based on probabilities and behavioural patterns. Nor can the study determine who was behind the identified accounts.

High levels of trust do not automatically provide protection

The study also discusses the implications of this type of influence in high-trust societies, such as Sweden and the other Nordic countries.

"High levels of trust are a strength, but we should not assume that they protect us against this type of information influence. If people trust their information environment, there is also a risk that they may not always scrutinise what they encounter as critically", says Linus Wahlberg. He continues:

"If we regard bots solely as a tool for spreading false information, we risk overlooking an important part of the problem. We also need to understand how automated accounts can alter our perception of the people and opinions represented in political debate."

The findings are estimates

Identifying bots is difficult. The study used machine learning tools to analyse activity patterns, metadata and account behaviour, among other factors. The analysis was supplemented by several manual checks. Linus Wahlberg therefore emphasises that figures of this kind should always be regarded as estimates.

"There is no definitive way of determining which accounts are bots. Bot detection is a game of cat and mouse: as bots become more sophisticated, they also become harder to detect. The findings should therefore be regarded as probable indications of automated behaviour, rather than conclusive evidence", he says.