In the 2023 Alberta election, we observed a significant amount of what may be inauthentic or fake, artificially generated engagement on Twitter. This included a very small number of users producing a disproportionate amount of abuse and a high number of potential bot accounts. This evidence, combined with other recent Samara Centre research, suggests that significant portions of online engagement in Canadian civic conversations is inauthentic or not representative of real Canadians, or even real people at all.
Our analysis of the Alberta election shows that a very small number of users made up large portions of discussions directed at candidates and party accounts. In fact, we found that 12% of all tweets came from just 200 Twitter users and that 12% of all abusive tweets came from just 50 Twitter users. In the Alberta election, a small proportion of users made up an outsized portion of Twitter engagement, and an even more outsized portion of abusive tweets.
These findings mirror what we observed in Engagement and Abuse on Toronto's Digital Campaign Trail: The 2023 Toronto Mayoral By-election Report, where we found that just 30 Twitter users accounted for 10% of abusive tweets.
These insights also resonate with US research from 2021 that found just 25% of users produced 97% of all posts on Twitter. These high-volume users are sometimes referred to as “power users.” One study found that high-volume Twitter users are less likely to view the civility of online discussions as a major problem compared to low-volume users. This may mean that power users view harsher language, which could cross into abusive territory, as more acceptable than the average user.
Our analysis of the Alberta election seems to show that the most active power users are more likely to publish abusive tweets. We found that abusive tweets were highly concentrated among very few users, and that these users may be having a significantly outsized impact on the overarching nature and tone of our online political discourse. To highlight this link between high-volume users and abuse, we term these users not just “power users” but “power abusers.”
In the chart, each user is represented by a circle. The circles’ sizes are dependent on how many total tweets we tracked from each user.
Only the 12,058 users with five or more monitored tweets are included in the chart above.
The above chart helps us visualize just how drastically different so-called power users’ posting habits are compared to typical users. We can only monitor the tweets sent to the 188 accounts we tracked, meaning that it is likely that each of these power users tweeted at even higher volumes with tweets not directed at our tracked candidates.
In terms of volume, users who tweeted more than 50 times are in the 98th percentile. Yet, many power users far surpassed that total. Given that researchers have defined high-volume users as those who tweet 20 or more times per month, it is striking that high concentrations of Twitter users posted 50-150 tweets over an 18-day period.
The highest volume power user tweeted 567 times to the accounts we monitored during the election period, which amounts to 31.5 tweets per day or more than one tweet per hour. The most prolific power abuser sent 145 tweets to the accounts we monitored during the election period, with more than two-thirds of their tweets categorized as abusive.
What explains these users’ extremely frequent posting behaviours? Some of these power users may be humans or bots attempting to intentionally manipulate civic discussions, and some may be real people who are unintentionally skewing discussion on the platform by being so highly active on Twitter.
These high-volume posting behaviours can also be evidence of astroturfing. This is the practice of hiding the sponsors of a message to make it appear as though it originates from, and is supported by, grassroots participants, in an attempt to manipulate public opinion. On Twitter, astroturfing can look like targeted posting habits on particular topics, inflating user interactions such as likes or retweets, or inflating users’ following counts. When these actions are done with manipulative intent, astroturfing undermines confidence in online platforms and discussion spaces as it influences the content we see, who we interact with, and our collective perceptions of situations at large.
Why does this matter? Power users and power abusers are skewing online political discussions, making them more abusive and less representative of the actual viewpoints of Canadians. They are contributing to affective polarization, shifting public perception of key issues and discouraging healthy and productive online conversations. While Canadians should be able to express themselves online, this does not mean that all types of content should be widely disseminated or promoted by platform algorithms. Algorithmic recommendation systems are amplifying harmful and abusive content, allowing fringe viewpoints to have disproportionate reach and to dominate social media feeds. This is all the more dangerous when these fringe viewpoints may be artificially generated or amplified and not the opinions of actual Canadians.
We also found evidence of what may be bot accounts, or accounts generated automatically or en masse. Bot accounts impact online discussions by masquerading as authentic users, which can distort online discussions. One way to identify potential bot accounts is through their use of numerical usernames (e.g. @abc4567890). In our analysis we classified bot accounts as those with five or more numbers in their username. We took this approach to identifying potential bot accounts because although fake or artificially generated activity in political discussions is on the rise, Twitter does not provide civil society organizations with access to this type of data.
During the Alberta election period, 43,434 Twitter users tweeted at the accounts we monitored. Out of this total, approximately 11% of the 300,861 tweets we monitored came from users with numerical usernames.
These 32,735 tweets came from just 4,751 accounts that were identified as being numerical usernames, which is relatively proportional to the number of tweets these users produced. However, we found that a subset of numerical username users had an outsized impact.
Our data shows that 1,033 of these numerical username accounts made up 80% of the tweets sent from numerical usernames. Each of these users published tweets directed at monitored accounts five or more times during the election period.
The 41 highest volume numerical username users tweeted at monitored accounts between 100 and 347 times each. These 41 users accounted for over 2% of all tweets we monitored during this election period—a remarkable proportion.
These findings illustrate that a very small number of users with numeral usernames took up an outsized space in election discussions. Any very small group having such disproportionate influence makes our collective perception of public opinion less accurate. Since these users have numerical usernames, it is likely that some of them were attempting to conduct astroturfing or other similar forms of inauthentic behaviour.
Although our election data allows us to identify potential bot activity with numerical usernames, there are limits to what we can analyze. The emergence of widely accessible generative AI technologies has allowed people to set up autonomously posting accounts that write in more sophisticated and bespoke ways at scale, and it is far more difficult to detect this kind of bot activity. Content created by these generative AI technologies is called “synthetic content.”
The rise of synthetic content makes attempting to measure inauthentic engagement on social media an increasingly difficult task. One study of synthetic content on Twitter found that it was “indistinguishable from organic content.” Research has also shown that powerful and wide-reaching networks of bot accounts, posting autonomously with help from AI tools like Chat-GPT, exist on Twitter and are not promptly identified or removed from the platform. The problem is likely to only worsen in the future.
Our research suggests that bots are being extensively used in Canadian and international political contexts, but these areas of study are severely under explored, especially as research efforts are stifled or under threat.
Why does this matter? Scholars have cautioned that the rise of generative AI in particular is a threat to democracy. Targeted, bespoke, large-scale misinformation campaigns can hinder the ability of elected officials to accurately gauge their constituents’ opinion on key issues, undermining democratic representation. In turn, these same campaigns can mislead voters, especially since AI-generated propaganda is “nearly as persuasive as propaganda generated by foreign actors.” The net result is weakened trust in our democratic processes.