Who violated the campaign rules? Complete evidence of illegal propaganda in the presidential race

For both the parliamentary and presidential elections, I served as an internal observer on behalf of the Funky Citizens association, participating in the observation of the elections from the perspective of compliance with electoral law during both the physical operations and the procedural aspects.

In the context of the current year's elections being heavily debated, and public institutions seemingly applying the law arbitrarily, it is imperative to increase transparency regarding the measures taken by the authorities in charge of identifying violations and the actions I have taken to identify and facilitate the detection of electoral violations.

Summarize video

Methodology used

The standard methodologies for qualitative analysis would not work in the case of analyzing forms of electoral propaganda, as traditional sentiment and keyword analysis are ineffective for a complex message such as that of masked electoral propaganda.

Thus, I proceeded to develop a methodology in which I utilized language analysis techniques using NLP models - specifically, I employed the Claude-3.5-Sonnet model, and then I reduced costs during the analysis of the propaganda related to the parliamentary elections using the Gemini 1.5-Flash-002 model, both of which are state-of-the-art (the best on the market).

The data was collected automatically, without manual processing, and the entire analysis process was detailed step by step on GitHub, in complete transparency.

So, if you wish to view the code used, below you have all the information, with the timestamps provided by Github to clarify that this was not a post-factum analysis to favor anyone or a select few:

Data collection method

Following the Cambridge Analytica incident, which was one of the key factors that influenced BREXIT, META has made its advertising more transparent to reduce the risk of voter manipulation.

What does this mean? All paid advertisements discussing political and social issues (which are likely to have electoral purposes or to "prepare" the population) must be reported on Facebook.

Facebook will refuse to publish unmarked ads or will block unmarked ads that initially passed the first filter. I also speak from personal experience as a marketer on META platforms, Google, etc.

Initially, I used the .csv files that can be obtained (maximum quantity 3) using the export button in META, searching for all specific keywords related to the names of all political candidates, from Georgescu to Lasconi to Simion to Ciolacu to Geoana. They contained fewer details, but still enough to report violations regarding electoral legislation.

With this, I prepared the first report, which was submitted to the police. Subsequently, during the parliamentary elections, I employed a more efficient technique, which is described below.

So, all the necessary information for identifying and sanctioning electoral propaganda is public and permanently stored by Meta on Ad Library.

To use an API for accessing data from the Ad Library, you must request approval from META, which can take months (around 3 months). Therefore, we proceeded to use a third-party provider that utilizes this API.

I have obtained a dataset that looks something like this (Note - edited version to include the reasons):

Data processing mechanism

As I described above, a traditional (manual) analysis would have been prone to my own subjectivity or that of other analysts, as well as being costly and slow.

So, we proceeded to use some AI prompts based on the law regarding electoral propaganda that describes what electoral propaganda is:

Brief discussion on legislation

LEGE nr. 370 din 20 septembrie 2004 (**republicată**)

Articolul 55:

Constituie contravenții, dacă acestea nu constituie infracțiuni, următoarele fapte:
[...]
t) continuarea propagandei electorale după încheierea acesteia, precum și sfătuirea în ziua votării a alegătorilor la sediul secțiilor de votare să voteze sau să nu voteze un anumit candidat;

[...]
Articolul 56:
(1) Contravențiile prevăzute la art. 55 lit. b), c), d), e), g), j), k), l), m), o), p), ș), ț), u), v), z), ab) și ac) se sancționează cu amendă de la 1.000 lei la 2.500 lei, cele de la lit. f), h), q), t), w), x) și aa), cu amendă de la 1.500 lei la 4.500 lei, cele de la lit. a), i), n), r), r^1), s) și y), cu amendă de la 4.500 lei la 10.000 lei.
[...]

LEGE nr. 334 din 17 iulie 2006, Art 36:
(7) Este considerat material de propagandă electorală orice material scris, audio sau video, care îndeplinește următoarele condiții:
a) se referă în mod direct la un candidat sau partid politic care participă la alegeri sau referendum, clar identificat;
b) este utilizat în perioada campaniei electorale, stabilită potrivit legilor privitoare la organizarea alegerilor;
c) are obiectiv electoral și se adresează publicului larg;
d) depășește limitele activității jurnalistice de informare a publicului.

In short, if you are conducting electoral campaigning, The fine is between 1,500 and 4,500 RON. Propaganda is any non-journalistic material aimed at influencing elections.

The Central Electoral Offices have issued their own "interpretative decisions," which are illegal as they violate the principle of legality. You cannot modify or annul parts of a law through an administrative act; it would be like me, as a mayor, saying that in my city it is permissible to kill people without criminal penalties, and that the police would actually enforce it.

Another reason for illegality is the lack of institutional competence, as the BEC is not superior to the police authorities and does not have authority over them. Therefore, the BEC cannot dictate the actions of the police, as they are neither prosecutors nor police captains. In a democratic state, the principle of separation of powers is upheld, which is particularly important during election periods.

Even so, highlight that the vast majority of violations also fell under the decisions of the BEC. Even though we consider these to be ILLEGAL, they are sensible and fundamentally correct (and were taken into account by the AI anyway).

How we educated the AI about interpretation

We have tried to build a correct, impartial AI model that can judge. when in doubt, for the accused to favor the defendants (in contravention matters).

The prompt, in .txt format, can be accessed below and checked individually:

In short, I mentioned the legislative provisions so that the model had to verify the cumulative fulfillment of the following conditions:

Based on the provided document, I will continue the explanation of how you structured the AI training for analyzing electoral propaganda. Here are the cumulative conditions that the model must verify:

1. Verification of the criteria in Art. 36(7):
   • Direct reference to a candidate/party (No vague advertisements that do not mention names but rather specific attributes such as balance, verticality, etc.)
   • The accessibility to the general public (i.e., they should not be targeted at a predetermined number of individuals. This was fulfilled by paid advertisements on Facebook and Instagram.)
   • The electoral objective (i.e., they must intend to encourage the voting of a candidate/candidates/party or discourage the voting of a candidate/candidates/party)
   • Exceeding the boundaries of journalistic activity (That is, they do not come with journalistic evidence, do not represent real information but "fake news" or unverified information such as "Tal Hanan," "Vote manipulation at Simion," etc.)
2. Specific verification elements:
   • The presence of a CMF code (Financial Mandatary Code) was mandatory according to the BEC. It is worth mentioning that the penalty was not stipulated as the annulment of the elections, according to the law.
   • If it represents a legitimate personal opinion vs. propaganda - Here I included something sensible - it should not be a personal opinion, but rather effective propaganda aimed at manipulating voters. If it were a personal opinion, it would not have been included in the report.
3. Checklist for inclusion/exclusion:
   • : Permitted communications such as:
     • Those targeting exclusively the parliamentary elections during the valid period
     • Discussions about the roles and official activities of candidates that are not related to their status as candidates (e.g: Prime Minister Ciolacu has secured our entry into Schengen.)
     • General communications regarding political issues (e.g: Romania is going through a difficult period and we need to address these issues, and I call on all candidates: to take action on X.)
     • Educational content about the electoral process (e.g: To vote, place the folded ballot into the ballot box.)
     • Legitimate journalistic material (e.g: Georgescu was previously investigated for promoting Antonescu, but received a verdict of non-prosecution.
   • : Prohibited communications such as:
     • Electoral propaganda after the legal period has ended
     • Direct influence on voters at polling stations
     • Alternative forms of propaganda that do not comply with legal requirements
4. Principles of analysis:
   • Impartiality in evaluation
   • The application of the principle "in dubio pro reo"
   • Focus on content and context
   • Objective assessment of the electoral purpose

The model was encouraged to analyze step by step, using a chain of thought for each of the aspects described above, weighing the pros and cons, so that the model could assess the situation as neutrally and accurately as possible, with a minimal false positive rate.

I did this so that the model would not be disproportionately harsh on certain parties or candidates (e.g., Ciolacu is allowed to boast that he was a good prime minister, as long as he does not ask to be voted for because he was a good prime minister).

How did we run the AI model?

I have built a pipeline that analyzes each form of electoral propaganda using information from META, which includes:

1. Demographic and audience data
   • Information about age (age_audience)
   • Distribution by gender and country
   • Reach and impressions
2. Information about the advertiser
   • Page Name
   • Page ID
   • Page Category
   • Verify page
   • Followers and likes
   • Information about the advertiser's profile
3. Advertisement details
   • Ad ID
   • Running period (start_date, end_date)
   • Budget (spend)
   • Publishing platform
   • Format and content
   • Text and images used
4. Compliance and regulatory information
   • Information about the payer/beneficiary
   • FEV Information (First Time Voter)
   • Information about violations
   • Moderation statuses
5. Targeting and performance date
   • Geographical distribution
   • Reach metrics
   • Engagement statistics
   • Targeted countries and regions

In short, the model had access to both the image of the post and the text of the post, as well as information about the contact details of the person who promoted the advertisement, the budget paid, the platforms where the post was published, effectively everything needed for the proper sanctioning of violations.

Example of analysis conducted with AI

<output>
	<analysis>
Let's analyze this post according to our criteria:

1. Has CMF number: No CMF number is present in this post.

2. Refers to candidate: This post does not directly reference any presidential candidates from our provided list. It focuses on local projects for Băilești municipality and mentions PSD parliamentarians in general.

3. Addresses wide population: Yes, this is a paid Facebook advertisement targeting the population of Băilești, with an estimated reach of 10,001-50,000 people.

4. Electoral objective: Yes, but for parliamentary elections ("Pe 1 Decembrie, votează calea sigură pentru Băilești!"). This is important because the parliamentary election campaign is still ongoing and allowed.

5. Journalism: No, this is clearly a campaign message listing promises and projects.

6. Personal opinion: No, this is an official party communication from PSD Dolj.

Important context: This appears to be parliamentary election propaganda, not presidential election propaganda. The date of December 1st mentioned refers to the parliamentary elections, for which campaigning is still permitted.


		<conclusion>
			<post_id>449822184815287</post_id>
			<electoral-propaganda-analysis>
This is indeed electoral propaganda, but it's important to note that it's specifically for the parliamentary elections scheduled for December 1st, not for the presidential election. The post lists specific local projects and improvements that PSD parliamentarians will support for Băilești municipality. Since the parliamentary election campaign is still ongoing and allowed, this type of propaganda is legal.
</electoral-propaganda-analysis>
			<electoral-propaganda-decision>FALSE</electoral-propaganda-decision>
			<electoral-propaganda-candidates></electoral-propaganda-candidates>
			<responsible-party-or-group>PSD</responsible-party-or-group>
			<message-for-police></message-for-police>
		</conclusion>

This post does not violate Article 55 t) because:
1. It relates to the parliamentary elections, not the presidential elections
2. The campaign period for parliamentary elections is still active and allowed
3. No presidential candidates are mentioned or referenced
4. The content focuses exclusively on local development projects and parliamentary representation

The post remains within the bounds of allowed political communication during an active parliamentary campaign period.

	</analysis>
	<conclusion>
Based on the detailed analysis above, this post does NOT violate Article 55 t) of Law no. 370 from September 20, 2004. It represents legitimate parliamentary election campaign communication during an active campaign period.
</conclusion>
</output>

As can be seen from the example above, the AI model addresses and verifies all essential aspects to identify what constitutes electoral propaganda and what does not.

The following analysis is approximate, so it should be taken with a bit of critical thinking and not at face value.

Analysis Results

Now that I have explained the analysis methodology, I believe it is important to see what it actually reveals. Below, we will present the aggregated results in the form of charts.

Information on each form of electoral propaganda has been reported individually.

I have monitored the following candidates in the presidential elections.

Poziția nr. 1 – ELENA-VALERICA LASCONI – Uniunea Salvați România
Poziția nr. 2 – GEORGE-NICOLAE SIMION – Alianța pentru Unirea Românilor
Poziția nr. 3 – ION-MARCEL CIOLACU – Partidul Social Democrat
Poziția nr. 4 – NICOLAE-IONEL CIUCĂ – Partidul Național Liberal
Poziția nr. 5 – HUNOR KELEMEN – Uniunea Democrată Maghiară din România
Poziția nr. 6 – MIRCEA-DAN GEOANĂ – candidat independent
Poziția nr. 7 – ANA BIRCHALL – candidat independent
Poziția nr. 8 – ALEXANDRA-BEATRICE BERTALAN-PĂCURARU – Alternativa pentru Demnitate Națională
Poziția nr. 9 – SEBASTIAN-CONSTANTIN POPESCU – Partidul Noua Românie
Poziția nr. 10 – LUDOVIC ORBAN – Forța Dreptei
Poziția nr. 11 – CĂLIN GEORGESCU – candidat independent
Poziția nr. 12 – CRISTIAN DIACONESCU – candidat independent
Poziția nr. 13 – CRISTIAN-VASILE TERHEȘ – Partidul Național Conservator Român
Poziția nr. 14 – SILVIU PREDOIU – Partidul Liga Acțiunii Naționale

Demographic Analysis

The majority of individuals who have seen electoral advertisements on Facebook fall into the following dominant categories:

1. Men aged 45-54 (14.26%) - the largest segment
2. Men aged 35-44 (11.75%)
3. Men aged 55-64 years (11.73%)
4. Women aged 65 and over (12.29%)
5. Women aged 55-64 (10.77%)

We observe that the platform has a relatively balanced distribution between genders, with a slight predominance of men in the middle age groups (35-54 years).

The "unknown" category (meaning undeclared gender or gender fluid) represents a very small percentage of the total (under 0.2% across all age groups), which is typical for social platforms where the majority of users specify their gender in their profiles as either female or male.

Target platforms

78.8% of electoral ads targeted both major META platforms - Instagram and Facebook.

Only a few chose to appear solely on Facebook - 19.4% of ads, while only 1.9% are on Instagram.

Expenses for electoral campaigning

As can be seen from the graph, the official pages of the PSD had the highest electoral expenditures on Facebook, ranging from approximately 24,000 RON to 38,000 RON.

The official AUR pages incurred expenses ranging from 17,000 RON to 27,000 RON.

The following, in descending order of expenses, are PNL, Independents, USR, Others, and the Right Force, PNCR, DREPT, and REPER. It should be emphasized that many forms of electoral propaganda were funded from accounts not associated with the parties (e.g., accounts of party members) and have not been analyzed here.

The chart is cumulative (RED + GREEN + GRAY = total on the y-axis).

We can see that the most disparaged candidates have their expenditures for defamation listed in descending order. ION-MARCEL CIOLACU, GEORGE SIMION, NICOLAE-IONEL CIOCA, yes ELENA-LASCONI.

In the editorial team's opinion, the enormous expenses for discrediting candidates are responsible for the severe polarization of society in the current period. The promotion of the candidates' values was secondary to denigrating the other candidates.

Above, you can see the daily expenses for candidates during the restricted period. In red, we have the amounts used for disparagement. In green, the amounts used for promotion.

The severity of violations concerning the candidate

It is clear that most individual violations (regardless of the budget used) were in his favor. ELENA-VALERICA LASCONI.

GEORGE-NICOLAE SIMION This is the second on the list in terms of the number of individual violations.

ION-MARCEL CIOLACU there are a small number of violations due to a strategy of circumventing electoral law by praising him as prime minister (rather than future president), a legal aspect even though he clearly used this to influence his election as president.

However, in some cases, the use of this mechanism to circumvent electoral law was mistakenly omitted, directly promoting his election as president.

In the current chart, we can see the total number of views for the advertisements (electoral propaganda) after their prohibition.

The ones who have truly benefited the most from the illegal electoral propaganda are, in descending order:

1. GEORGE NICOLAE SIMION, with over 350,000 views of illegal propaganda
2. ELENA VALERICA LASCONI, with just under 150,000 views of the illegal propaganda.
3. CALIN GEORGESCU, with over 100,000 views of illegal propaganda
4. NICOLAE-IONEL CIUCA, with approximately 50,000 views (illegal propaganda) each.
5. LUDOVIC ORBAN, with approximately 50,000 views.

Amounts owed to the state budget

If the complaint is taken into account, and assuming absurdly that all factors are considered and not nearly zero or only for the parties with issues, the following emerges.

False Positive Analysis

For the presidential campaign, I used a less clear data structure than that for presidential elections, specifically CSV files from exports on the Ad Library with the candidates' names as keywords.

Some forms of political advertising are actually advertisements that have a political impact but are merely press, rather than propaganda. The AI algorithm considers not whether the name includes "PRESS" or similar terms, but whether they appear to be forms of defamation/promotion that do not adhere to ethical standards, such as verification, neutral language, etc.

Only the information directly present in the post is taken into account in this analysis (unlike the one for CDEP and the Senate); only the text is considered, not the images.

Notification sent to the police

Response of public institutions

So far, I have not received any notification proving that any reported and denounced act has been sanctioned. Below you will find the responses received from the institutions.

The Response of the Romanian Gendarmerie

Since both the Romanian Gendarmerie and the Romanian Police have been notified (for strategic reasons, to prevent disappearance), they stated that the Romanian Police would be more capable of handling the request and conducting the investigation.

They called me by phone, explained the limitations regarding the investigation, and provided me with the following written response.

The first redirect was to the IGPR - Directorate of Criminal Investigations

After mistakenly arriving at their location, she proceeded to the Cluj Police Inspectorate.

The second redirect, from IPJ Cluj to the Directorate of Criminal Investigations Bucharest

Since this response, I have not received any further information regarding the presidential electoral propaganda; however, the complaint concerning the same issues is slowly but surely progressing to the Senate.

No one has been sanctioned yet for these. More details will be found in the next article, which will address these elections.