On 6 December 2024, the Constitutional Court of Romania took an unprecedented step: it annulled the first round of the country’s presidential election. Not over ballot-box fraud, nor over irregularities in the count, but because one candidate, the previously unknown Călin Georgescu, had emerged in first place through a coordinated influence operation on TikTok, with indications of foreign state involvement.
For the first time in the history of the European Union, an electoral process was annulled due to social media manipulation. The Romanian case is not an isolated incident. It is the most emblematic example of a new reality: electoral processes have become a domain of hybrid threats, in which artificial intelligence and social networks act as force multipliers for anyone seeking to influence democratic outcomes.
What has changed since 2016
When the Cambridge Analytica scandal broke in 2018, the world learned that data drawn from 87 million Facebook profiles had been used to build psychographic voter profiles for the 2016 U.S. presidential election and the Brexit referendum. As sophisticated as the microtargeting of that era appeared at the time, it was manual and analog compared to what is available today.
From 2023 onward, three factors have fundamentally reshaped the landscape:
Democratization of generative AI. Tools such as ChatGPT, Midjourney, and open-source voice-synthesis models have made the creation of convincing text, images, audio, and video content accessible to anyone with no technical expertise required.
Engagement-driven algorithms. Platforms such as TikTok, whose recommendation systems are based purely on user behavior rather than the user’s social graph, can propel an obscure profile onto millions of screens within a matter of days.
Professionalization of influence networks. State actors have integrated social media into the arsenal of psychological operations and hybrid warfare, in what the EU now formally terms FIMI (Foreign Information Manipulation and Interference).
2024 was a landmark year: 3.7 billion citizens in 72 countries were called to the polls. Never before has so large a share of humanity voted in a single electoral cycle. And never before have the tools of manipulation been so cheap, so accessible, and so effective.
The election-manipulation arsenal
1. AI-powered microtargeting
The analysis of data drawn from social-media interactions, likes, shares, comments, and time spent on each post enables the construction of detailed psychographic profiles. With the aid of machine learning, a political campaign can segment the electorate into dozens or even hundreds of categories and deliver to each a distinct message, not merely different in tone but at times contradictory in content. The problem is not only one of privacy. It is that the very notion of public political debate is dissolved: when every voter sees a different candidate, there is no longer a common reference point against which to compare programs and positions.
2. Deepfakes and synthetic media
Deepfakes — AI-generated videos, images, or audio files depicting real individuals saying or doing things they never said or did have already surfaced in a wide range of electoral contests:
- United States, January 2024: a robocall using a synthetic voice of Joe Biden urged Democratic voters in New Hampshire not to turn out for the primary.
- Slovakia, 2023: days before the election, a fabricated audio clip circulated on Facebook in which liberal candidate Michal Šimečka allegedly discussed rigging ballots.
- Argentina, 2023: both leading candidates, Javier Milei and Sergio Massa, deployed deepfakes to ridicule one another.
- Ireland, October 2025: a deepfake video showed a presidential candidate purportedly withdrawing from the race on the eve of the vote.
According to an IE University survey conducted in October 2024, 40% of Europeans are concerned about the misuse of AI in elections, while 31% believe AI has already influenced their vote.
3. Bots, inauthentic networks and algorithmic amplification
In Romania, the country’s intelligence services declassified documents demonstrating the coordinated use of thousands of TikTok accounts to artificially amplify Georgescu’s profile, through coordinated posting, fake “organic” interactions and paid promotion that was not declared as political advertising.
Research by Global Witness found that the platform’s algorithm recommended content favorable to the candidate between 4.6 and 14 times more frequently than content favoring his opponent. No state actor is even required. Today, a small number of operators, assisted by LLMs, can sustain thousands of convincing fake accounts that post, comment and interact across multiple languages, without the tell-tale errors that betrayed the previous generation of bots.
4. Poisoning the AI models themselves
A more recent threat is data poisoning. Groups such as the Russian influence operation Storm-1516 have begun “contaminating” the online environment with fabricated stories designed not for human consumption but to be ingested by the training data of future AI models and by recommendation systems. When a citizen subsequently asks a chatbot about an event, they receive disinformation, lent credibility by the perceived “neutrality” of the machine.
Why this is a cybersecurity matter
Traditionally, electoral security has been treated as a question of protecting infrastructure: electronic counting systems, digital voter rolls, and results-transmission networks. That approach remains necessary, but it is no longer sufficient.
The new battlefield is the information ecosystem. Just as a SCADA network can be compromised by a targeted attack exploiting known vulnerabilities, so too can the public sphere be compromised by operations that exploit known cognitive biases, psychological vulnerabilities, and algorithmic security gaps.
Viewed from this angle, an election influence operation does not differ structurally from an APT attack:
- Reconnaissance: collection of voter data, mapping of social networks, identification of “entry points” (polarised communities).
- Initial access: creation of inauthentic accounts or hijacking of existing ones.
- Lateral movement: dissemination through algorithmic amplification and influencers.
- Payload: the disinformation content itself.
- Persistence: sustaining the narrative through recurring campaigns.
- Exfiltration: the “loot” is not data, it is electoral behavior.
This analogy is not merely figurative. A growing number of organizations, from government agencies to media outlets and political parties, are now confronting FIMI campaigns with the same tools and frameworks they use to detect cyberattacks: SIEMs, threat intelligence, indicators of compromise, and post-incident analysis.
The European Regulatory Framework
Today, the European Union is the world’s most stringent regulator in this field. Three key instruments shape the landscape:
Digital Services Act (DSA). Fully in force since February 2024, it requires Very Large Online Platforms (VLOPs) to assess and mitigate systemic risks, including risks to electoral processes. On the basis of the DSA, the European Commission opened formal proceedings against TikTok over the Romanian case.
AI Act (Regulation 2024/1689). In force since August 2024, with full application from August 2026. It imposes transparency obligations: deepfakes must be labelled as such, and AI-generated text informing the public on matters of public interest must be flagged. Penalties reach €15 million or 3% of worldwide annual turnover, depending on the infringement.
Code of Practice on Disinformation and Code of Practice on Transparency of AI-Generated Content. Voluntary frameworks that serve as a bridge until full activation of the AI Act, comprising platform commitments to label synthetic content, deploy watermarks, cooperate with fact-checkers, and share data with researchers.
In parallel, pan-European political parties signed a code of conduct in April 2024 pledging not to use deceptive AI-generated content in the European elections. Norwegian parties signed an equivalent commitment in November 2024.
How We Defend Ourselves
The response is not the exclusive responsibility of any single actor. It demands coordinated action at four levels.
At the citizen level. Digital and information literacy is the first line of defense. Initiatives such as EUvsDisinfo and national fact-checking programs must be integrated into schools’ and professional development curricula. The basic principle: the more intense the emotional reaction a piece of content provokes, the more skeptically it should be treated.
At the organizational level. Political parties, media organizations, electoral authorities, and other critical institutions must treat FIMI risks as ordinary cybersecurity risks. In practice, this means dedicated social-media policies, executive training on deepfakes, synthetic-content detection tools, and rapid-response procedures in the event of an eve-of-election attack, the so-called “October surprise” can now unfold digitally.
At the platform level. Transparency in recommendation algorithms, rigorous labeling of political advertising, robust identity verification for accounts that achieve mass reach during election periods, and watermarking of all AI-generated content.
At the state level. Cross-agency cooperation between electoral authorities, national CERTs, intelligence services, and regulators. Regular tabletop exercises simulating manipulation scenarios. Clear allocation of roles and responsibilities, which proved costly in the Romanian case.
Conclusion
Artificial intelligence and social networks are not, in themselves, a threat to democracy. Like any powerful technology, they are dual-use tools. They can increase participation, strengthen transparency, and give a voice to those who had none. Yet the same technology that enables personalization can be used for manipulation. The same algorithmic amplification that brings valuable content to light can propel disinformation. The same generative AI that lowers the cost of creativity simultaneously brings the cost of propaganda to near zero.
Romania is not an exotic case; it is a warning. No European democracy, Greek or Cypriot included, is immune.
Protecting electoral integrity now requires the same tools, skills, and maturity we apply to safeguarding critical infrastructure. Democracy is not protected only at the ballot box. It is protected every time an organization, an institution, or a citizen chooses to recognize manipulation rather than amplify it.
CEO & Founder @ Audax Cybersecurity | Offensive Security Specialist
The opinions expressed in this post belong to the individual contributors and do not necessarily reflect the views of Information Security Buzz.