.svg)
The Samara Centre for Democracy monitored activity on Twitter during the 2023 Alberta election as part of our SAMbot project, a multi-year machine learning initiative that measures abusive content received by Canadian political candidates on social media.
The Samara Centre for Democracy monitored activity on Twitter¹ during the 2023 Alberta election as part of our SAMbot project, a multi-year machine learning initiative that measures abusive content received by Canadian political candidates on social media. We found that online abuse and potential inauthentic behaviour—including likely fake and artificially generated engagement—were prevalent during this election. These negative, harmful, and manipulative aspects of online civic engagement are democratic threats. They increasingly make politics less accessible to ordinary citizens and frustrate attempts to foster productive and pro-democratic dialogue.
This report provides a brief background on the Alberta political context and data about the candidate and party accounts we monitored. We then analyze candidate experiences on the digital campaign trail, correlating spikes in engagement and abuse to real-world events. Key areas of analysis include: how inauthentic engagement is shaping our practice of politics; how social media platforms are amplifying anti-LGBTQ+ rhetoric; and how first-time candidates are facing online abuse right out of the gate. We conclude by offering recommendations about how Canadian institutions and digital platforms can best respond to this ever-changing online landscape.
¹ The social media platform formerly known as “Twitter” has rebranded as “X” since we monitored this election. The platform was still called “Twitter” when we collected this data, so we refer to the platform as “Twitter” throughout this report.
We tracked 188 accounts including seven official party accounts and 181 candidate accounts. Candidates from nine political parties were represented, as well as four independent candidates.
The tracking period consisted of 18 days from 12:00 am MT on May 12, 2023 to 11:59 pm MT on May 29, 2023 (the beginning of the day after the end of the candidate nomination period, to the end of the day on election day).
During the tracking period we monitored a total of 300,861 tweets, of which 12,502 were abusive.
The tracked candidates were mentioned 387,792 times.² Of those mentions, 15,376 contained abusive content.
²Tweets can mention multiple candidates simultaneously. Thus, there are more mentions than total tweets. To learn more read X’s “Mention overview.”
This small group is skewing online political conversations and making them more abusive and less representative of the views of actual Canadians. These high-volume users are often referred to as “power users.” To highlight our finding that some power users are likely to circulate online abuse, we use the term “power abusers.”
We found evidence of potentially widespread inauthentic engagement in the form of possible bot accounts. However, as AI-generated content becomes more sophisticated and easily accessible, it is increasingly difficult to distinguish between authentic and synthetic engagement.
Discussions regarding LGBTQ+ rights garnered heightened online engagement and abuse, a worrying finding we have observed in other Canadian elections.
It is not only high-profile candidates that receive vitriol online. Political newcomers face abuse too. We observed that first-time candidates received high levels of abusive content relative to incumbent candidates. This indicates that candidates can expect abuse as a condition of work from the very beginning of their political careers.
Defined by researchers as “a trend where citizens develop a strong affective connection toward their own political side, while increasingly disliking and feeling animosity toward people with opposing political allegiances.”
The practice of hiding the sponsors of a message to make it appear as though the message originates from, and is supported by, grassroots participants, in an attempt to manipulate public opinion.
Accounts operated automatically or en masse, often with the intention of skewing online discussion to particular ends.
Online activity that is the product of inauthentic use, such as posts by bot accounts, fake user engagement, or artificial amplification. Inauthentic engagement may be coordinated in attempts by either domestic or foreign actors to influence Canadian political processes or, more generally, to sow discord and confusion online. However, as AI-generated content becomes more sophisticated and easily accessible, it is increasingly difficult to distinguish between authentic and artificially generated or synthetic content.
To highlight the link between high-volume posting behaviours and online abuse, we use the term “power abusers” to suggest that high-volume social media users are also very likely to post high volumes of abuse.
Users who post frequently on social media. While some power users may be real users who simply post frequently, others may be bot accounts posting at rates not humanly possible in an attempt to skew online discussions.
Online content, in the form of text, image, or audio, that is fully or partially artificially altered or generated.
By all accounts, the 2023 Alberta provincial general election was a close one. Of the 87 seats up for grabs, 49 went to the UCP and 38 went to the NDP, with incumbent UCP Premier Danielle Smith re-elected. Looking at the popular vote, 53% of Albertans voted for UCP candidates while 44% voted for NDP candidates, leaving just 3% of votes going to other parties. Calgary proved to be a key battleground, with several ridings won by very narrow margins.
Despite the UCP’s win, several UCP cabinet ministers lost their seats and the party’s vote share decreased in more ridings than it increased, with a loss of 11 seats. Conversely, the NDP added 15 seats to become the strongest opposition party in Alberta’s history.
The race, as some commentators noted, was high on drama and low on issues, with voters deciding between two high-profile and established candidates: incumbent Premier Smith and former premier Notley. Given this context, the election was ultimately regarded as a two-party race, with smaller parties and independents failing to make their mark. In fact, smaller parties saw their share of the vote fall by 74%, which meant that for the second election in a row, only the UCP and the NDP had members elected to the legislature.³
The election saw a number of notable firsts. It marked the first time that a Black woman, Rhiannon Hoyle, and a First Nations woman, Jodi Calahoo Stonehouse, were elected to the legislature. The percentage of women Members of the Legislative Assembly (MLA) also increased from 30% to 38%.
³ Jennifer Johnson is included as an elected UCP member; however, she was expelled from the UCP caucus due to her comments about transgender children in 2022. She sat as an independent MLA until October 2024, when she was welcomed back into the UCP caucus.
.jpg)
We limited our monitoring to candidates that had public and active Twitter accounts as of the end of the candidate nomination period on May 11, 2023. This means that we did not monitor any candidates from the Advantage Party of Alberta, the Communist Party – Alberta, the Reform Party of Alberta, the Buffalo Party of Alberta, and the Pro-Life Alberta Political Association, despite the fact that they did have candidates running in the election. We monitored at least one candidate from every riding in the province, except for Chestermere-Strathmore where no candidates had a public or active Twitter account by the end of the candidate nomination period.
We also monitored political party accounts for the UCP, the Alberta NDP, the Green Party of Alberta, the Alberta Party, Independence Party of Alberta, the Alberta Liberal Party, and the Wildrose Loyalty Coalition. Other parties did not have a public or active Twitter account when our tracking commenced.
It is clear that NDP- and UCP-affiliated accounts dominated the online conversation. Just 4% of all mentions (17,408) were directed at non-NDP and non-UCP candidates.
Over half of these tweets (9,904) were directed at Artur Pawlowski, the party leader of the Solidarity Movement of Alberta, and the only Solidarity Movement of Alberta candidate monitored during our tracking period. The remaining 7,504 mentions (2%) were directed at 50 accounts that weren’t affiliated with the UCP or NDP.
In the 2023 Alberta election, we observed a significant amount of what may be inauthentic or fake, artificially generated engagement on Twitter.
Read MoreParty leaders Rachel Notley and Danielle Smith’s Twitter accounts had the most engagement of all tracked accounts.
Read MoreSince 2021, we have deployed our SAMbot project in 12 elections across federal, provincial, and municipal elections, and in one municipal by-election, producing five reports. We have analyzed over 4.1 million tweets and tracked over 1,500 different candidates across Canada.
Based on our analysis of the data collected during the 2023 Alberta election, combined with findings from our past SAMbot deployments, we make the following recommendations:
Altering how digital platform operators incentivize, empower, or permit certain forms of conduct on their platforms will reduce online abuse and the effects of other information threats at scale.
In order to combat democratic backsliding and reduce online harms more broadly, Canada should follow “design code”-style legislative efforts popularized by the UK, EU, Australia, and the US (and separately the State of California) who have either adopted or are working on design code-style legislation. The Canadian federal government’s Online Harms Act, proposed earlier this year, is a design code-style bill.
With a design code, digital platforms have duties to user safety as well as consumer protection and care, while government regulators are responsible for identifying non-compliance with the code and holding platforms accountable. When non-compliant, platforms are responsible for amending their platform design to become compliant again. Platforms that refuse to comply and defy requests from regulators are subject to large fines, usually based on a percentage of global revenue.
One possible design solution to limiting inauthentic behaviour and the outsized impact of power users and power abusers is through altering rate limit policies on users’ accounts. Rate limits restrict how many times a user, device, or application can repeat an action during a certain timeframe. Traditionally, this is used to protect networks from being overwhelmed by influxes of requests, but rate limiting how users can post could also be used to encourage healthier social media use. It has been recommended by technology experts as one way to combat election manipulation.
Ensuring that Canadian researchers receive access to digital platforms’ data is paramount for them to thoroughly investigate and quantify the online harms and democratic threats that plague our online spaces. This access is of utmost importance in order to understand technology’s influence on our democracy.
Legislation that facilitates this access would allow Canadian researchers to further investigate cases of inauthentic engagement, astroturfing, and phenomena surrounding power users and power abusers, in addition to other digital information threats. At a macro level, Canadian researchers could access data that would help advocate for and justify specific platform design changes, such as increased rate limiting.
Supporting data accessibility and increased transparency is vital, as social media researchers are facing increasing suppression. Recent restrictions in application programming interface (API) access on X (formerly Twitter), Reddit, and TikTok are reducing the volume and quality of data available to researchers, and corporations (particularly Twitter/X) frequently use litigation as an intimidation tactic against researchers.
Abuse isn’t an acceptable condition of work. We need candidates from a wide range of backgrounds to ensure our democracy is truly representative of all Candaians, but abusive and dangerous working conditions will make it difficult to attract or retain them. As political newcomers, first-time candidates are the most vulnerable to abuse as they have the least support and experience to help them navigate online (and offline) abuse.
Negative experiences, furthermore, may deter first-time candidates from running for office again. Political parties can play a key role in providing first-time candidates with support, as can non-partisan mentorship schemes. Adding protections and expectations of conduct within our public institutions (such as workplace safety legislation) could also help limit both on- and offline abuse. We must foster a healthy civic culture that can help reverse the hostility and toxicity that have unfortunately become the norm.
Anti-LGBTQ+ hate has significantly increased in Canadian civic discourse in recent years. In 2022, anti-LGBTQ+ candidates organized across Canada to run in school board elections. Municipalities and school boards have placed bans on pride flags on their property and rainbow crosswalks in their communities. Protests at drag events have become commonplace, and LGBTQ+ people have been targeted by hate crimes and violence. This year, CSIS has warned Canadians that Pride celebrations could be targeted by “lone actors” inspired by the growing anti-LGBTQ+ movements and rhetoric online.
Technology-facilitated gender-based violence is an epidemic that is threatening the safety of women and gender-diverse people. This poses a democratic threat as women and gender-diverse people subsequently avoid politics and public life both on- and offline because they do not feel safe.
Democracy can only function in tandem with the upholding of human rights, allowing people of all identities to participate in society safely and earnestly. It becomes increasingly difficult for LGBTQ+ Canadians to participate in public life as anti-LGBTQ+ hate is emboldened.
Foreign interference is a national threat that has catapulted into public consciousness as of late, with diasporic Canadian communities at particular risk of being targeted. We know that foreign—and domestic—actors are actively using online discussion spaces to skew Canadians’ perceptions.
Online foreign influence can include targeted influence operations which aim to shift or embolden public sentiment about particular policies, parties, and candidates, or social issues. It can also just aim to destabilize democratic discussions and promote abusive, harmful, or anti-democratic rhetoric in any way possible, in order to increase affective polarization and undermine democracy.
Increasingly, Canadians need to be aware of how and where they may be seeing online foreign interference or inauthentic behaviour more broadly. More accessible and powerful generative AI technologies are making the creation and spread of convincing disinformation faster and cheaper for bad actors.
Outside of digital platform regulation, Canadians can best be inoculated against these information threats by improving their understanding of digital platforms’ content recommendation systems, inauthentic behaviour, foreign interference, and artificial intelligence. These literacy efforts would need to reach Canadians across demographics by using a variety of mediums and methods.
Our research directs attention to big questions about the democratic threats that Albertans, and Canadians more broadly, face in digital environments. These information threats include astroturfing, the rise of accessible synthetic media generation, and the worrying normalization of hateful rhetoric, digital violence, and affective polarization. In some cases, our online civic discussions are being distorted, as a small number of users are having an outsized effect on the conversation, which could result in many Canadians deciding to disengage from political conversations altogether. While our research looked at a single social media platform during a single provincial election, our findings help us to understand the broader digital democracy ecosystem.
This report also illuminates the abuse that political candidates face when running for office. Increasingly, this abuse impacts not only candidates with long and established careers in politics, but first-time candidates as well. When abuse and toxicity are conditions of work, our democracy suffers as more and more people—particularly municipal politicians, women, and politicians from racialized or marginalized backgrounds—opt out of politics altogether. No one wins in such a system.
Canadians need to act collectively to change our democratic culture to be more collaborative and productive, while combating affective polarization (which is increasing because of social media conduct). We need significant changes to our information diets and political culture to ensure we not only reach our goal of preserving Canadian democracy, but also push forward and actively improve our democratic systems.
There is a silver lining to our findings, however. If online abuse in Canadian political conversations is caused by a small proportion of users, the situation may not be as dire as it often seems. With the right guardrails and regulatory systems in place, coupled with digital literacy campaigns that teach Canadians about new and evolving online threats, there is hope that our online discussions can become more productive, representative, and pro-democratic.
We monitored the 2023 Alberta provincial election for 18 days, from May 12, 2023 at 00:00 MT to May 29, 2023 at 23:59 MT.
Twitter accounts were monitored if candidates and political parties had public and active Twitter accounts as of the end of the candidate nomination period, May 11, 2023.
We monitored 188 Twitter accounts in total. Seven of these accounts were political parties’ official Twitter accounts: the United Conservative Party, Alberta NDP, Alberta Green Party, Alberta Party, Alberta Liberal Party, Wildrose Loyalty Coalition, and the Independence Party of Alberta. The remaining 181 accounts belonged to candidates across nine political parties, plus independents. We monitored at least one candidate in every riding in Alberta, except in Chestermere-Strathmore.
Although the following parties ran candidates in this election, we did not monitor any candidates from the Advantage Party of Alberta, the Communist Party – Alberta, the Reform Party of Alberta, the Buffalo Party of Alberta, or the Pro-Life Alberta Political Association, as zero candidates from these parties had public and active Twitter accounts as of the end of the candidate nomination period.
Our SAMbot project has tracked abuse in federal, provincial, and municipal races since 2021. However, since data from each election is collected during different time periods, with different lengths, and with different totals of tracked candidates, it is not useful, nor advised, to compare SAMbot data across elections.
We do not collect data on retweets, as counting the same tweet more than once can distort the analysis. We only evaluate text within a tweet; content such as images, audio, or videos that may spread abuse cannot be evaluated by the machine learning tools that we use.
In our SAMbot project, we use machine learning models to assess abusive language. These models are ever-evolving which means that during each deployment, our data is more accurate and informed.
Please note that compared to our previous SAMbot deployments, we have evaluated abusive tweets significantly differently in this election.
Our machine learning model makes a confidence prediction to assess whether a tweet should be considered “abusive.” When measuring abuse, our model gives each tweet a score from 0% to 100% for each category, based on how confident it is that the tweet is abusive in nature.
Previous SAMbot deployments used a 51% confidence prediction to evaluate abuse; we now use a 70% confidence prediction. This change means that our analysis will be more accurate, and that cumulative results will paint a better picture of how abuse is distributed across the entire election and across all candidates.
Simultaneously, this greater threshold also means that some nuanced and subtle forms of abusive language will be missed by our machine learning model, and will at first make it appear that there is comparatively less abusive content, which is not necessarily the case. Machine learning models will never be able to monitor all abusive language across an election, as the subjective nature of what constitutes “abuse” does not permit the possibility of 100% accuracy. This methodological change allows us to more accurately represent how abuse is distributed overall.
We have made this change consistent with our constant effort to strive for more accurate and ethical methodological practices within the field of social media and machine learning research. Using confidence intervals in this way is in line with recommendations for social science research.
This change makes abuse volumes look considerably lower than in previous elections we have tracked (although SAMbot data should never be compared across elections regardless). Please consider these significant methodological changes while interpreting the data in this report.
SAMbot deployments use machine learning tools — software applications that run automated tasks. Using machine learning allows us to analyze tweets at a massive scale. Through our SAMbot project, we can evaluate millions of tweets for how likely abusive they are. We track all English and French tweets sent to candidates. Each tweet tracked, whether a reply, quote tweet, or mention, was analyzed against five abuse categories using a machine learning tool called Perspective API:
Perspective API provides a confidence prediction to assess whether a tweet meets an abuse category. When a tweet is evaluated, it is given a score from 0% to 100% for each category, based on how certain the machine learning model is that the tweet meets that abuse category. If the tweet is assessed as >=70% likely to meet an abuse category, we determine that the tweet has met the criteria. If a tweet meets at least one of the five abuse categories at the >=70% confidence interval, it is counted as an abusive tweet. The abusive tweet category serves to aggregate all tweets that meet at least one abuse category.
We thank the team at Areto Labs for their technical expertise, input, and collaboration. We also acknowledge the contributions of Jessie Lawrence for their research assistance.
Alex MacIsaac
Andrea Mariko Grant
Sabreena Delhon
Beatrice Wayne
For access to the data included in this report, please refer to the SAMbot 2023 Alberta Election spreadsheet. If you have any other questions about our data or the SAMbot project, please reach out to us at hello@samaracentre.ca.
The Samara Centre for Democracy, Astroturfing and Abuse: The 2023 SAMbot Alberta General Election Report (Toronto: The Samara Centre for Democracy, 2025).
Exploring engagement, toxicity, and civic conversations across online and offline spaces in the Toronto by-election.
