What happened on TikTok around the annulled Romanian presidential election? An investigation and poll

After the previously little-known ultranationalist Călin Georgescu won the first round of Romania’s presidential elections on 24 November, many journalists and regulators started to examine the impact that TikTok may have had on his victory.

The Romanian authorities asked the European Commission to look into TikTok’s role in the election. Under the EU’s Digital Services Act, platforms like TikTok are required to mitigate risks to election integrity.

On 4 December, the outgoing Romanian president declassified Romanian intelligence documents stating that they believed Georgescu had benefitted from coordinated accounts, algorithmic amplification of Georgescu’s presence on the platform and paid promotion.

Two days later, Romania’s constitutional court annulled the first round of the election and cancelled the upcoming runoff between him and rival Elena Lasconi. 

TikTok has rejected allegations of giving preferential treatment to Georgescu’s account.

Following increasing scrutiny from lawmakers and civil society, the company also announced it intercepted three covert influence networks in the country in recent months.

The constitutional court noted that one candidate received “preferential treatment on social media platforms, which resulted in the distortion of voters’ expressed will.” 

It is extraordinary that an election in an EU country could have been annulled because of allegations of an influence campaign conducted via social media – and a real test case for the Digital Services Act.

Following a series of Global Witness investigations into social media’s impact on election integrity around the world, we decided to examine what election content TikTok’s algorithm is pushing to people in Romania.

What content on Romania’s presidential candidates does TikTok’s recommender system show?

In Romania’s capital Bucharest on 5 December, three days before the presidential runoff was scheduled to take place, we set up an investigation into election content on TikTok’s For You page. 

The For You page contains content chosen by TikTok’s algorithm for each user – a personalised feed of recommendations selected “based on your interests and engagement”.

TikTok does not publish its algorithm, making it difficult to see how exactly it decides what material to recommend. So, in this case, we looked at what it decides to recommend. 

To do this we created fresh accounts which engaged with each presidential candidate equitably, then spent 12-20 minutes on the For You page watching electoral content and skipping content that did not appear to relate to the candidates, before analysing the content shown. 

The results were deeply one-sided. In three tests over two days, TikTok’s algorithm consistently recommended content supporting one candidate – Georgescu – at a much higher rate than the other candidate. 

The three test accounts were shown pro-Georgescu posts between 4.6 and 14 times more often than pro-Lasconi posts. 

The posts contain some indications as to what may be helping their amplification – from the actions of both users and TikTok. 

On several political posts, TikTok actively suggested to us to repost the video to our followers.

We observed memes designed for engagement, such as “where are you supporting Georgescu from?”, to encourage people to post where they were from in a show of support, and statements saying things like “let's make this blow up on TikTok”.

Some videos had millions of views and several videos with the same content appeared in more than one test.

Some videos contained misleading and harmful content. This included that Ukrainian refugee children are receiving more government assistance than Romanian children – a narrative that has been fact-checked as false. 

When we told TikTok about this video, they said they made it ineligible for recommendation. 

We also saw an extremely homophobic post criticising Lasconi which likely violated TikTok’s Community Guidelines. TikTok said that they had already removed the video before we alerted them to it.

We observed that TikTok often did place prompts offering more information about the election on videos, but not on all of those that were definitely related to political conversation around the election.

Notably, after we seeded the accounts, none of the content shown to us came from the official account of either candidate (differing from an earlier investigation we conducted in the UK).

We have shared our findings from this investigation with the European Commission so that it can help inform their inquiries into the platform’s compliance with the Digital Services Act. 

We carried out a similar investigation from the UK rather than Romania, in which we found that pro-Georgescu content was featured in the For You page of a politically balanced user more than five times that of his presidential rival.

What type of content did TikTok users in Romania see on TikTok?

TikTok said it prevented and removed fake accounts and fake followers leading up to the election. 

It described the three influence campaigns that it interrupted as "very, very small networks." 

Rather than letting the company mark its own homework, we wanted to get a sense of what Romanians saw when they logged into the app. 

We commissioned the polling company Survation to carry out a survey of more than 1,000 young (18-35) TikTok users in Romania. The survey was done in the days before the runoff was supposed to take place.

A majority of the Romanian TikTok users surveyed reported seeing suspicious activity, as described below, and what they thought were false claims on the app in the preceding two weeks. 

The results of the poll are summarised in the table below.

TikTok disputes the ability to draw conclusions from this survey because they said that "suspicious activity" “may hold a different meaning for different respondents” and people may have been influenced by media attention on TikTok.

TikTok’s response to our investigation and poll

We wrote to TikTok to give them the opportunity to comment on these findings and they said that our study does not reflect how real people use TikTok. 

They suggested that other factors may have influenced what videos we were shown, such as the length of time that we watched each video or where the videos were posted from, as they describe on their website. 

TikTok said that our test had a partisan bias, and that if a user engages with, for example, political content, “it's reasonable to expect a recommendation algorithm to suggest more of that content as a result.” 

We disagree with this because as far as possible the accounts we created were politically balanced between the two presidential candidates competing in the runoff, yet the content served to us was heavily slanted towards one of the candidates.  

They state that in Romania they have taken steps to protect the integrity of the platform and recently prevented millions of “fake likes” and “fake follower requests”, blocked more than 400,000 spam accounts from being created and more than a thousand accounts impersonating a presidential candidate.

Conclusion

The findings of our research raise questions for TikTok. 

Călin Georgescu was decidedly the most-seen candidate in our survey of users in Romania, as well as the most supported in our investigation of TikTok’s algorithm. 

73% of survey respondents said they remembered seeing content from him “somewhat often” or “very often” in the preceding two weeks.  

Against this backdrop, we commend the EU for opening inquiries of TikTok under the Digital Services Act, including sending them questions to answer and ordering them to preserve all relevant data.  

The EU has put groundbreaking rules in place to hold Big Tech to account, and should continue to support these rules. 

The majority of TikTok users in Romania who we surveyed are supportive of the EU doing more to make sure TikTok protects Romanian democracy. 

Despite challenging conditions for platform regulation, we have every expectation that the EU will take strong action if they find evidence of election interference.

Methodology: How we investigated TikTok’s algorithm

• On the afternoon of Thursday 5 December, we set up a new TikTok account on a blank phone in Bucharest, Romania. 
• We “seeded” the account by searching for the official TikTok account of presidential candidate Elena Lasconi, followed the account and watched three recent posts with more than 10,000 views.
• We then searched for the official TikTok account of presidential candidate Călin Georgescu, followed the account and watched three recent posts with more than 10,000 views.
• We then went to the app’s For You page, the platform’s feed that recommends content to users.
• We aimed to produce an account which would signal to TikTok’s algorithm interest in content related to the election, but without a partisan bias between the two candidates competing in the runoff. This involved lingering on posts or watching videos which were relevant to the election (for example, where the content featured one of the candidates) and swiping past videos which were irrelevant (such as car adverts or pet videos).
• After spending close to 20 minutes on the For You page, we stopped and reviewed all the content shown.
• On the following day, 6 December, we repeated the test twice using a fresh phone number and account each time. For the first of these two tests, in seeding the account, we searched for Georgescu’s account before Lasconi’s and spent 12 minutes on the For You page. For the second of these two tests we searched for Lasconi’s account before Georgescu’s and spent 15 minutes on the For You page.
• The accounts were given the ages 24 or 38 and did not choose any interests from a set of category suggestions at the account creation stage.
• Working with a Romanian analyst, we manually reviewed and coded all posts shown according to whether they indicated support or encouragement for one of the two candidates, contained political content without affiliation or support for one of the two candidates, or contained unrelated content.