Executive Summary

We monitored the Romanian online space between June 29 and July 13, 2026 — two weeks marked by the governance crisis and the DIICOT investigation into the care centres administered by Viorel Pașca in Bihor County. We were interested in what the public was discussing, how information circulated, and how much of its visibility was organic.

We collected and analysed:

• 696.202 posts in Romanian from the Facebook platform, covering the dominant topics of the period and dissemination patterns;
• 4.857 posts dedicated to the Bihor care homes case, analysed from a sentiment perspective;
• 7.000 Facebook comments from the most popular posts about the case, analysed from the perspective of hostility: how frequent it is, at whom it is directed, and what form it takes;
• synchronised dissemination events — accounts that publish identical posts within intervals of a few minutes — in order to separate authentic reactions from coordinated amplification.

Key Findings

1. Hostility is concentrated on institutions, regardless of the post’s own sentiment. The public is not divided into two comparable camps. It is divided only in its assessment of Viorel Pașca; but regarding institutions, hostility is almost unanimous and appears regardless of the orientation of the post being commented on. On the post by the Pașca family's lawyer, 97.5% of hostile comments attack targets other than the author: the state, other persons involved, the press, the justice system. On the posts by the journalist who documented the case, the direction reverses: 90.3% and 84.6% respectively of hostile comments target her, dominated by personal attacks. On Marcel Ciolacu's post, 88.3% of hostile comments target the author. Even on a neutral, informative post (Oradea24.ro), the majority of comments are negative — directed at the press and the state, not at Pașca. The real axis of conflict is not pro/anti-Pașca; it is citizens versus institutions.

2. On average, approximately 13% of the public exposure of an online narrative (in the studied dataset) comes from the coordinated republication of the same messages — and on political and institutional topics, the share rises toward one fifth. 395 accounts generated 173 synchronised dissemination events — identical posts published within intervals of up to 5 minutes — on the Pașca case alone. More than half of these coordination links were concentrated within a 3-hour and 25-minute window on July 1.

3. Artificial amplification targets precisely the topics that divide. The strongest amplification appears in security and defence (21.92%), governance (21.17%), war and geopolitics (21.12%), EU and NATO (20.60%), and justice and corruption (19.65%) — these percentages represent the share of a topic's exposure obtained through verbatim-distributed messages — even though the dominant content of the period was social and emotional: family, tragedies, health. The pattern is consistent in our data: the greater a topic's potential to divide public opinion, the more coordinated republication/amplification it attracts.

4. The case was reported predominantly informatively, but commented on with anger. Of the 4,857 posts about the case, 73.52% had a neutral tone and only 13.47% a negative one. In the comments, the ratio reverses: nearly two thirds (61.94%) are negative. The posts generated, in total, 1,919,395 reactions, 385,978 comments, and 342,159 shares.

5. Who is attacked depends on the perceived camp; how they are attacked differs by gender. Authors perceived as favourable to Pașca are almost bypassed by direct hostility: the family's lawyer is targeted by only 2.5% of the hostile comments on his own post, the MP who contested the reporting — by 8.4%. Authors perceived as hostile become targets themselves: the journalist — 90.3% and 84.6%, the former prime minister — 88.3%. The highest rate of strictly personal attacks appears at a male politician (75.9%); but gender stereotypes, sexualised insults, and dehumanising formulas appear exclusively directed at the journalist. 

Why Does This Matter?

The analysed period caught Romania in the midst of a governance crisis, with a caretaker Executive and without a clearly formed parliamentary majority, while the August 31 deadline for completing PNRR reforms was rapidly approaching. Against this backdrop, the Dumbrava Case became the convergence point of two crises (the political and the trust crisis), uniting the emotional dimension (vulnerable persons, deaths) with the institutional one (the responsibility of a state that directed people toward unauthorised centres for almost 20 years). In a context where distrust in institutions is already generalised, amplification networks do not need to convince the public of anything new: it is sufficient to sustain and accelerate a pre-existing hostility. The stake of this analysis is the separation of layers — what part of the pressure on institutions is an organic reaction and what part is constructed — a necessary step for correctly understanding not just this case, but any sensitive topic that will appear in the Romanian public space in the period ahead.

The data reveal an asymmetric distrust, not a classic case of polarisation. The discussion is not conducted between a pro-Pașca and an anti-Pașca camp of comparable size. The public is divided only in its evaluation of Viorel Pașca; regarding institutions — the prosecutor's office, the ministry, the press — hostility is almost unanimous and appears regardless of the orientation of the post being commented on. The same audience that supports an author perceived as close to Pașca attacks the state and journalists; the same audience transforms into a target any author perceived as hostile to Pașca.

The real axis of conflict is citizens versus institutions

This convergence of hostility matters all the more given that an artificial amplification mechanism operates on top of it. The 395 accounts that synchronously disseminated content about the case, in 173 events concentrated in windows of a few minutes, show that part of the visibility of the narratives does not come from the authentic interest of users, but from coordinated republication, adding up to 13% additional exposure for the targeted topics.

Main Online Topics

The content disseminated during this reporting period is dominated by social themes with strong emotional charge. Family conflicts, separations, and financial difficulties intersect with numerous accounts of accidents, deaths, domestic violence, and other serious incidents. In parallel, the personal lives of celebrities, health issues, and messages about resilience, faith, and emotional support occupy significant space in public discourse and in the circulation of content on social networks.

Distribution of dominant topics within the corpus

On the political and economic front, the online environment is marked by confrontations between leaders and parties, government formation negotiations, disputes over reforms and fiscal measures, and concerns about the cost of living, incomes, infrastructure, and the property market. The justice sphere is dominated by investigations and controversies regarding the actions of the authorities, including the Viorel Pașca case and the care centres in Bihor, against a backdrop of visible distrust in institutions.

The comparison between the total exposure of posts and the estimated level in the absence of republications shows how much additional attention is generated by copies of the same message. On average, these republications increase public exposure by approximately 13%, but the effect is stronger in political, institutional, and geopolitical topics. The greatest amplification appears in the security and defence category, at 21.92%, followed by governance at 21.17%, war and geopolitics at 21.12%, EU and NATO at 20.60%, and justice and corruption at 19.65%. The data indicate that the repetition of the same messages contributes significantly to the visibility of these subjects.

The most identically and synchronously disseminated posts are concentrated in the Family and Relationships, Tragedies and Accidents, and Crime and Violence categories. Although republishing the same content through numerous channels increases exposure, the data indicate a clear saturation effect: tripling the volume of verbatim messages relative to the number of original messages generates only approximately 17% additional engagement. In other words, beyond a certain threshold, the repeated multiplication of the same text produces increasingly diminishing gains in public attention, a phenomenon observable across other thematic categories as well.

Publishing the same text from multiple accounts within a very short interval can help the message exceed an initial attention threshold, but can also produce faster saturation, as copies frequently reach overlapping audiences and compete for the same reactions. As a result, the engagement curve may rise steeply at first but flattens quickly, with each additional copy generating fewer and fewer new interactions.
 

The ratio of "unic/total" shows the percentage of audience that was obtained through original posts, out of the total.

The difference between the Total Proxy and the Unique Proxy indicates an important contribution of duplicated content to the level of attention generated, but with significant variations between topics. The ranking above is by percentage; in absolute values (the raw volume of exposure brought by duplicates) the hierarchy differs, because large topics produce more duplicate volume even at smaller shares. The largest absolute values appear in Domestic Politics, where the difference reaches 18.91% of the Total Proxy, and in Justice and Corruption, at 19.65%. Economy and Family and Relationships follow, with shares of 16.67% and 17.03% respectively, as well as Religion and Spirituality, at 11.25%. In Crime and Violence and Tragedies and Accidents, the difference represents approximately 15% of total attention, while Censorship and Freedom of Expression reaches 10.02%. The lowest shares in this group are recorded by Entertainment and Celebrities at 8.01% and Motivation and Personal Development at 7.97%, suggesting that, among dominant themes, the effect of multiplication through repeated content is most pronounced specifically in political and judicial areas and in social subjects with strong public charge.

Truly interesting, however, is the case of the second general category by audience reception, Justice and Corruption. Here, a large part of the audience's attention is concentrated on the scandals regarding the illegal care homes in Bihor and the Viorel Pașca case, including abuse accusations, investigations, searches, and interventions by the authorities. A high level of interest also appears around the accusations targeting Ilie Bolojan — a narrative that, as we show in the coordination section, was synchronously disseminated by 21 accounts on July 3, within a window of under five minutes — and an alleged involvement in acts of corruption and institutional incompetence. The graph above also shows that some subjects were intensely replicated, especially in the case of narratives related to topics sensitive for public trust, particularly on political topics, justice, and corruption.

The Dumbrava Care Centres Case

At the end of June 2026, DIICOT conducted searches in Bihor County at a network of care centres administered by Viorel Pașca for almost 20 years, under the cover of the "Dumbrava" Association. Hundreds of vulnerable persons, most of them with mental disabilities or without family, were housed without authorisation in 17–18 houses, some of them having been directed toward these centres by state institutions: hospitals, prisons, city halls, and DGASPC offices. According to investigators, the persons housed were in a state of dependence on the centre administrators, without access to adequate medical care; on this basis, prosecutors opened an investigation for human trafficking. Viorel Pașca and members of his family were detained, then placed under judicial supervision, and the court subsequently decided that they should be investigated in a state of liberty, a decision contested by DIICOT.

The case quickly became one of the most discussed topics on Romanian Facebook, because it did not remain merely a criminal investigation. It transformed into a public debate about where responsibility lies: with Viorel Pașca, dubbed by some the "good Samaritan" for his activities, or with the state, which sent people to Dumbrava for years without requiring the centres' authorisation or offering alternatives. From the total corpus of 696,202 posts, we looked at 4,857 posts to better understand what people discussed online in reference to the words "Dumbrava," "azil" (care home), and "Viorel Pașca."

How the Coordination Works

The data behind the interactive 3D graph below show the synchronised dissemination of identical messages about the Pașca case and tell a clear story: in the majority of cases, we are not dealing with a network that posts organically and at random, but with accounts that act in tandem, seconds apart from one another. Within a span of just 3 hours and 25 minutes, between July 1 at 15:45 and 19:10, more than half of all coordination links that artificially amplified various narratives about the Pașca case were disseminated.

Two further periods of intense activity followed:

• From July 2^nd to 8^th, there was an increase of over 17% in the number of identical messages disseminated within 5-minute intervals.
• From July 8^th at 08:30 UTC (11:30, Bucharest time) to July 12^th at 15:40 UTC (18:40, Bucharest time), there was an increase of nearly another 30%.

3D Graph: the synchronised dissemination of identical messages concerning the Pașca case

This pattern, including periods of quiet, interrupted by short and intense bursts of activity, is a typical signal of coordinated behaviour, not of spontaneous and independent posts by ordinary users.

The graph above suggests that coordination links initially preceded the accumulation of engagement: around July 2, more than half of the total coordination links had already been recorded, while reactions and engagement had accumulated only approximately 35–40% of their final values. This gap is compatible with a "seeding" phase, in which the synchronised publication of the same content rapidly produces visibility and the impression of wide interest. Subsequently, between July 2 and 4, engagement grows much faster than coordination, reaching approximately 75%, which may indicate the uptake of messages by users or channels beyond the initial core. After this point, a saturation effect becomes visible: although the number of coordination links continues to grow substantially, especially around July 9–10 and 12–13, engagement is already close to its ceiling and advances very little. In other words, coordination appears to have been effective in accelerating initial exposure.

Main coordination events

The largest coordination event in the graph took place on July 1, 2026, between 19:10:16 and 19:12:44 UTC (between 22:10:16 and 22:12:44, Bucharest time), meaning that only 2 minutes and 28 seconds elapsed between the first and last post. During this interval, 134 accounts disseminated the same text. Among these accounts are Observatorul, Iubire & Trădare, Autovehicule Românești, Terapie prin zâmbet, România Noastră, Românii din Străinătate, Mănăstirea Vlădiceni, Fabricat în România, Bancuri BETON, Acasă, and România ta.

On July 2, 2026, between 11:30:55 and 11:33:41 UTC (between 14:30:55 and 14:33:41, Bucharest time), during the second coordination wave described above, 20 different pages and profiles (including DIVIN, DUMNEZEU Există, Din România, Flori Si Peisaje, Românașul, România, URĂRI, ȘTIRI, and ȘTIRI Plus) published the same material related to the Pașca family scandal. The difference between the first and last post was 2 minutes and 46 seconds.

On July 3, 2026, between 16:05:59 and 16:10:43 UTC (between 19:05:59 and 19:10:43, Bucharest time), another group — this time 21 accounts, mostly Facebook pages such as Curiozitatile momentului, Dumnezeu e bun, Exploratorul Minții, FapteDiverse, Filme Vechi, Muzica Eurodance, Paradisul florilor, Remedii Naturiste, Suflet Românesc, Umorul, and Ziua Bună — distributed, minutes apart from one another, material suggesting the existence of a link between Ilie Bolojan and the Pașca scandal. Total coordination window: 4 minutes and 44 seconds.

Another coordination wave was recorded on July 12, 2026, beginning at 15:40 UTC (18:40, Bucharest time), when 41 accounts published the same material simultaneously, on the accounts Inima mea, Eşti tot ce am., and Terapie prin zâmbet, with the last post appearing at 15:42:50 UTC (18:42:50, Bucharest time), on the account Acasă. The difference between the first and last publication was 2 minutes and 32 seconds. Among the participating accounts were also Observatorul, România Noastră, Te ıubesc, Românii din Străinătate, România mea, DRAGOSTEA MEA, Dulce Romanie, Bancuri BETON, Ora De Ras, Fabricat în România, and MAMA. The text displayed in the graph, in normalised and abbreviated form, was: "Viorel Pașca, the head of the Bihor care home, explains where the money found by pro[secutors] came from. See more."

Account names are reproduced exactly as they appear on the platform, including non-standard characters (e.g. «ı» without a dot) — a frequent practice for evading duplicate content detection.

The concentration of more than half of the coordination links within an interval of approximately three hours, followed by the repetition of similar patterns in the following days, indicates a synchronised amplification of content associated with the Pașca scandal, rather than a succession of isolated occurrences. The cumulative evolution suggests that this coordination preceded the growth of engagement and contributed to the acceleration of initial visibility, after which the messages were picked up beyond the initial core of accounts. The distributed content included polarising narratives regarding Romanian institutions and social tensions.

Overall, the data indicate a well-synchronised distribution mechanism that we have documented before, activated promptly upon the occurrence of political crises or security incidents. These networks can be activated at any moment of vulnerability in our society and can be used to split communities, to create and maintain a state of uncertainty or confusion, to undermine Romanian institutions, or for any other type of influence operation.

Interpretation of Sentiment Categories

Positive sentiment

The positive category comprises 220 posts, representing 4.53% of the corpus. Messages that express predominantly support, admiration, solidarity, or appreciation are categorised here.

One direction consists of direct support and public mobilisation: "Very interesting and lucid, Viorel Pașca! He explains everything very clearly!", "Over 2,500 people attended the support rally for Viorel Pașca," "Wave of solidarity for Viorel Pașca," "Moment of prayer and solidarity," or "The participants [...] concluded the demonstration by reciting prayers in unison for the Pașca family."

A second direction presents him favourably by recounting his activity: "Viorel Pașca and his family tried to be a support for people in difficult situations" or "tried to help numerous people in difficulty."

A third direction defends the conditions in the centres, as in the assertion "The conditions are better than at my own home!"

In all these cases, positive sentiment is determined by the expressed public support, the appreciation of the activities carried out, and the presentation of Viorel Pașca as a well-intentioned person.

Around the positive posts, 190,790 likes, 11,239 "love" reactions, 5,951 "angry" reactions, 7,562 "sad" reactions, 3,982 "haha" reactions, and 1,405 "wow" reactions accumulated. Cumulatively, these represent 220,929 reactions. The posts also generated 35,617 comments and 35,806 shares. The high number of likes and "love" reactions is compatible with messages of support, solidarity, and appreciation, such as "Massive support for Viorel Pașca" or "Wave of solidarity." However, the presence of "angry" and "sad" reactions shows that users could support the person presented while reacting negatively to the intervention of the authorities or to the general context of the case. 

Posts with the highest interaction

The post ranked first by likes, total reactions (angry + care + haha + like + love + sad + wow), comments, and shares is signed by lawyer Răzvan Doseanu and was published on June 30. In it, the lawyer announces legal assistance for the Pașca family and solicits public testimony about their activity.

Negative sentiment

The negative category comprises 655 posts, i.e. 13.49% of the corpus. The texts express accusations, indignation, fear, anger, or disapproval. The target of the sentiment differs: some messages criticise Viorel Pașca, others criticise the authorities, DIICOT, the press, or the way in which persons from the centres were relocated.

Criticism directed at Viorel Pașca appears in formulations such as "A new shocking scandal hits Romania," "exploited innocent victims," and "'The camp of death,' in God's name." Terms such as "horror care homes," "innocent victims," "inhumane conditions," or "shocking scandal" indicate a strongly negative assessment of the centres' activities. In the same direction fall alarmist headlines such as "Devastating blow for Viorel Pașca" and "ASTONISHING scandal."

Criticism directed at the authorities appears in formulations such as "GENOCIDE by the ROMANIAN STATE," "This is not justice, it is a crime," "CHRISTIANITY BANNED BY THE GENERAL PROSECUTOR'S OFFICE AND DIICOT!," and "A transferred patient has died, and the Minister of Labour is hiding it..."

Both directions use direct accusations, capital letters, exclamation marks, and terms of high emotional intensity, which distinguishes them stylistically from informative texts.

The negative posts recorded 381,291 likes, 12,009 "love" reactions, 8,882 "angry" reactions, 28,721 "sad" reactions, 3,716 "haha" reactions, and 4,171 "wow" reactions. Cumulatively, these represent 438,790 reactions. Around the negative posts accumulated 60,220 comments and 87,380 shares. The "sad" and "angry" reactions are more numerous here than in the positive category and correspond to texts using expressions such as "shocking scandal," "inhumane conditions," "grotesque action," or "despicable manoeuvre." The high number of shares suggests that critical messages, accusations, and alarmist information had significant circulation capacity in the online space.

Posts with the highest interaction

Ranked first by likes, total reactions (angry + care + haha + like + love + sad + wow), and shares is the post by MP Emanuel Ungureanu from July 5, which contests the reporting about how deceased persons at Dumbrava were buried. Ranked first by number of comments is the post by journalist Carla Cristina Tanasie from July 4, about DIICOT prosecutors' accusations regarding funeral allowances.

Neutral sentiment

The neutral category is the largest and comprises 3,571 posts, representing 73.52% of the corpus. It includes mainly news, informative headlines, judicial decisions, statements, and updates on the evolution of the case.

Clear examples are: "The Ministry of Health evaluated 411 persons from Viorel Pașca's care homes," "The request for preventive arrest of Viorel Pașca was rejected," and "Dragoș Pîslaru will meet with ITM representatives to request information about the 1 million fine." These texts communicate a fact or an action, without the author directly expressing approval or disapproval.

Also in this category fall posts such as "Viorel Pașca's first reaction after the DIICOT investigation: he published photographs taken during the searches," "The network of illegal care homes in Bihor [...] was dismantled by DIICOT," "Why the court decided that Viorel Pașca and his family should not be preventively arrested," "What prohibitions the judge imposed on Viorel Pașca," "Viorel Pașca, his wife, and the centres' coordinator were detained by DIICOT," and "Forensic doctors established the cause of death."

Even if the subject is controversial, the formulations are centred on who stated something, what the court decided, how many persons were evaluated, or what measures were taken. The neutral classification reflects the predominantly informative character of the post, not the gravity of the event reported.

The neutral posts concentrated the most interactions in absolute values: 697,865 likes, 30,683 "love" reactions, 41,927 "angry" reactions, 47,568 "sad" reactions, 29,962 "haha" reactions, and 8,916 "wow" reactions. They generated 225,864 comments and 140,406 shares. These high values are explained primarily by the much larger number of neutral posts in the corpus. Although the texts are predominantly informative — of the type "The Bucharest Tribunal has ruled..." or "Forensic doctors established the cause of death" — the public could react emotionally to the event reported. Therefore, neutrality describes the formulation of the post, not necessarily the users' reaction to the subject.

Posts with the highest interaction

The post with the most likes and total reactions (angry + care + haha + like + love + sad + wow) is signed by Oradea24.ro, from July 1, and reports on the support demonstration organised at the Dumbrava cemetery. The most commented post in this category belongs to Marcel Ciolacu, from July 3, comparing the political reactions of certain officials in connection with the case. The most shared post, also by Oradea24.ro, from July 2, invites the public to listen to the testimonies of the persons cared for in the centres before forming an opinion.

Mixed sentiment

The mixed category comprises 411 posts, i.e. 8.46% of the corpus. These texts contain both support or appreciation and criticism, reservations, or significant unfavourable information.

A first example: "For 20 years, the state sent abandoned people to Viorel Pașca. Today it accuses him of crimes. If criminal acts exist, they must be proven and sanctioned. But Romanians deserve clear explanations." The text acknowledges the possibility of crimes, but simultaneously criticises the state's behaviour.

A similar structure appears in the statement "I believe I did good," followed by the clarification that Viorel Pașca "admits that it is possible he made mistakes." The favourable assessment of intent is accompanied by the acknowledgement of possible errors.

Another example combines the negative label "the owner of the 'care homes of horror'" with the information that Viorel Pașca had been awarded and described as "a remarkable man." In the same vein, the formulation "supports Viorel Pașca's good faith," followed by "Justice is done in court, not on social media," expresses support but maintains a reservation regarding the establishment of guilt.

Also in the mixed category fall texts that appreciate the fact that abandoned persons received food and shelter, but simultaneously mention the lack of authorisations, possible administrative errors, or the accusations formulated by the prosecutors.

The mixed posts accumulated 347,570 likes, 12,667 "love" reactions, 10,370 "angry" reactions, 21,661 "sad" reactions, 7,669 "haha" reactions, and 2,818 "wow" reactions. They generated 64,277 comments and 78,567 shares. The combination of favourable and negative reactions corresponds to the ambivalent character of the messages, which may appreciate the intention to help vulnerable persons while simultaneously criticising the lack of authorisations, possible illegalities, or the intervention of the authorities. For example, a text may state that "Viorel Pașca did good," but admit at the same time that "it is possible he made mistakes." The high volume of comments and shares indicates that these nuanced posts stimulated debate and the confrontation between different perspectives.

Posts with the highest interaction

The post with the most likes, total reactions (angry + care + haha + like + love + sad + wow), and shares in this category is signed by the page "Adi Rusu - Detectivul de Presă ȘOC," from July 3, and reports the conditions in detention as described by the Pașca family's lawyer. The most commented post belongs to journalist Carla Cristina Tanasie, from July 1, and presents how the case was initially documented.

Reactions across the entire corpus

In total, the analysed posts accumulated 1,617,516 likes, 66,598 "love" reactions, 67,130 "angry" reactions, 105,512 "sad" reactions, 45,329 "haha" reactions, and 17,310 "wow" reactions. The corpus generated 385,978 comments and 342,159 shares. Likes represent the most frequent form of reaction, but they should not be automatically interpreted as agreement with the message, as users may like a post for the relevance of the information or for the visibility of the subject. Similarly, emotional reactions describe the public's response to content, while the sentiment category describes the tone of the published text.

Comparative Table: Comment Analysis by Post

The comment analysis presented below is based on a sample of 1,000 comments per post, extracted for each of the seven posts with the highest interaction identified earlier within the four sentiment categories (positive, negative, neutral, mixed). The selection was not arbitrary: for each category, the posts ranked first either by likes and total reactions (angry + care + haha + like + love + sad + wow) or by number of comments were retained, considered representative of how the public reacted to the dominant content of the respective category.

Distribution of sentiment within the comments, based on the 7 analysed posts.

The comparison between the seven posts shows that the "negative" label does not automatically equate to hostility toward Viorel Pașca. On the posts signed by Răzvan Doseanu, Emanuel Ungureanu, Oradea24.ro, and Adi Rusu, negative comments predominantly target the authorities, DIICOT, or the press — not the author of the post or Pașca himself. The main constant of the conversation is the shifting of the target of hostility according to the author's identity: posts favourable to Pașca attract criticism of the state and journalists, while posts signed by journalists (Carla Cristina Tanasie) or political actors (Marcel Ciolacu) themselves become the target of hostility. Marcel Ciolacu's post is the most negative and most polarised (92.3% negative), with direct criticism of him and of PSD. Adi Rusu's post is the most balanced, with only 5 percentage points between negative and positive. In all cases, support for Viorel Pașca frequently coexists with criticism of the authorities, the press, or the conduct of the investigation, without the two positions being mutually exclusive.

Hostility Analysis

For this section we looked more closely at how hostility manifests in comments: whether it is directed directly at the post's author, whether it concerns content, or whether it transforms into a personal attack. We also tracked situations in which hostile comments do not attack the author but other persons, institutions, or groups mentioned in the discussion.

Of the seven posts initially analysed, we selected five that allow a clearer comparison: the two posts by Carla Cristina Tanasie and the posts by Răzvan Doseanu, Emanuel Ungureanu, and Marcel Ciolacu. All have a clear individual author and 1,000 comments each, making it possible to compare how hostility is personalised.

Distribution and target of hostility, based on the 5 compared posts.

Comment analysis by post

In the case of Carla Cristina Tanasie's first post, 693 of the 1,000 comments are hostile, and 626 of these — 90.3% — directly or secondarily target the author. Of the comments addressed to her, 372/626 (59.4%) are exclusively personal attacks, and 200/626 (31.9%) combine contesting the material with disparaging the person. Exclusively content-based criticisms represent only 49/626 (7.8%). Personalisation takes the form of direct insults, such as "Pathetic!," "You're stupid.," "Imbecile," or "Hyena, not a journalist!," as well as general accusations regarding character and integrity: "You're a liar!" or "Can you keep lying?" Some attacks contest intelligence through a gender stereotype — for example "You can tell you're a blonde" — while others seek to silence the author, such as "Stop already!!!!" The 67 hostile comments that do not target her are directed primarily at the press, the authorities, and the justice institutions.

In the case of Carla Cristina Tanasie's second post, 811/1,000 (81.1%) of comments are hostile, of which 686/811 (84.6%) target the author or the journalists associated with the material. Exclusively personal attacks represent 275/686 (40.1%), and mixed attacks 255/686 (37.2%). Exclusively content-based criticisms are more numerous than in the first post, 154/686 (22.4%), but personal attacks remain present in over three quarters of the hostile comments addressed to the author if mixed cases are included. The language is dominated by challenges to intelligence and competence through formulas such as "Two idiots.," "Village idiots," "Morons!," "Blondes gone wrong!!!" or "Blondes...and that's all." Forms of dehumanisation and moral degradation also appear, such as "Two lizards," "TWO NITWITS," and "Disgusting liars," as well as sexualised insults. The hostility not targeting the author — 125/811 (15.4%) — is directed primarily at state authorities.

The situation is different in the case of lawyer Răzvan Doseanu. Of the 280 hostile comments, only 7/280 (2.5%) effectively target him, while 273/280 (97.5%) direct their hostility toward other targets: 107/273 (39.2%) toward state authorities and institutions, 93/273 (34.1%) toward persons or groups involved in the case but insufficiently identified in the text, 31/273 (11.4%) toward political actors, 16/273 (5.9%) toward the press and journalists, 13/273 (4.8%) toward DIICOT, prosecutors, or the justice system, 9/273 (3.3%) toward Viorel Pașca or his family, and 4/273 (1.5%) toward other commenters. Of the seven comments that target the lawyer, four are exclusively personal attacks, two are mixed attacks, and one expresses hostility without substantial criticism. Examples include professional delegitimisation — "Doseanu, another crook lawyer..." —, accusations of financial interest — "You sensed there was money to be made" — and other examples such as "Get lost" or "Building your audience like that...get out."

Also in the case of MP Emanuel Ungureanu, the majority of hostility is not directed at the author. Of 632 hostile comments, only 53/632 (8.4%) directly or secondarily target him, while 579/632 (91.6%) attack primarily the press, state authorities, political actors, or the Pașca family. Of the 53 comments addressed to the MP, 16/53 (30.2%) are exclusively personal attacks, and 28/53 (52.8%) are mixed attacks. The attacks include political and moral insults, such as "Stinking USR thief," "You wretch, how much money did you get from his millions," or "A scoundrel and you're the same, filming people," challenges to integrity — "Did they buy you too???" or "How much are you getting for this report!" — and ridicule of a personal characteristic: "You stutterer, ask normal questions." Other comments combine criticism of the post with orders to be silent, for example "Ungureanu you're rubbish, get out of here!"

In the case of Marcel Ciolacu, 946/1,000 (94.6%) of comments are hostile, and 835/946 (88.3%) target the author. Of these, 634/835 (75.9%) are exclusively personal attacks — the highest rate of strict personalisation in the corpus — and another 136/835 (16.3%) are mixed attacks. Personalisation is dominated by political ridicule, general insults, and attacks on intelligence or integrity. References such as "Pretzel man! How much is the pretzel?," "Marcelica, where's the money?," and "Marcel, did you pass your baccalaureate, class dunce?" transform elements of his public image into formulas of mockery. Direct attacks include "What a scoundrel you are, you rat," "An idiot he was, an idiot he remains!," "still running on 2 neurons," and "How low can you stoop!" Unlike the cases of Doseanu and Ungureanu, hostility in this corpus is largely concentrated on the post's author. The 111 comments that do not target him are directed primarily at other political actors and the justice institutions.

Comparing the results shows that the overall intensity of hostility and its personalisation are distinct dimensions. Emanuel Ungureanu has 632 hostile comments, but only 53 target him, while Marcel Ciolacu is the target of 835 of the 946 hostile comments. In the case of the two Carla Cristina Tanasie posts, hostility is also concentrated on the author — 626 and 686 comments respectively. Qualitative differences appear in the vocabulary used: attacks against the journalist frequently include challenges to intelligence and competence through the "blonde" stereotype, as well as sexualised or condescending insults. Attacks against Ciolacu are centred primarily on the ridicule of political identity, education, morality, and public image, while those against Ungureanu and Doseanu primarily contest integrity, professional competence, and presumed complicity or financial motivation.

Conclusion

The analysis of the two weeks of crisis shows that public discourse in Romania cannot be understood merely through the prism of post volume or expressed sentiment, but through its target. The Bihor care homes case was treated predominantly informatively (nearly three quarters of posts had a neutral tone), but in the comments hostility was almost generalised and, more importantly, convergent: regardless of the orientation of the host post, it was directed primarily at institutions. The data support this observation across all five comment analyses. On the post by the Pașca family's lawyer, 68.6% of comments are positive, but 97.5% of the hostile ones attack targets other than the author: the state, the press, the justice system. Among the posts classified as negative toward Pașca appear formulations such as "GENOCIDE by the ROMANIAN STATE," which does not target him but the state.

The public is therefore not divided into two comparable camps. What varies from one post to another is not the intensity of hostility, but the direction in which it discharges according to the author's identity. This finding has a direct methodological consequence: identifying the target of sentiment — which we had until now treated as a limitation of the analysis — proves to be the main dimension to be measured. Valence alone (positive, negative, mixed, neutral) describes the tone, but does not say against whom it is oriented, and in this case it is precisely the target that is the relevant information.

On top of this layer is superimposed a separately documented phenomenon: networks of accounts that synchronously disseminated, in windows of a few minutes, the same texts about the case, adding to the targeted narratives up to 13% additional exposure. The coincidence between the topics with the greatest potential to divide (domestic politics, justice and corruption) and the topics most intensely amplified through coordination is not coincidental and confirms a pattern already observed in other crisis episodes in the Romanian public space.

The Dumbrava case was therefore not merely a criminal investigation followed by the public, but a terrain on which authentic user reactions, a pre-existing and generalised distrust in state institutions, and the coordinated amplification infrastructure identified in previous crises all converged.

In such a context, amplification networks do not need to construct hostility — only to direct and accelerate it. Separating these layers (who reacts organically, at whom it is directed, and who artificially amplifies) remains the necessary step for correctly understanding not just this case, but any other sensitive topic that may appear in the Romanian public space in the period ahead.

Methodology

The research began from the hypothesis of a classic polarisation around the Dumbrava care centres case — a pro camp and a con camp. The data, however, showed a different pattern: the public is divided only in its assessment of Pașca, while hostility toward institutions remains constant regardless of the orientation of the post being commented on. The methodology below was adapted to capture this distinction, by separating sentiment from the target of hostility.

1. Thematic analysis

We collected a corpus of 696,202 posts in Romanian from Facebook, disseminated in the period June 29 – July 13, 2026. Following dataset processing, we excluded 71,161 posts that did not fit any of the identified themes, some of them containing too little information to be interpreted. For thematic analysis we used a Machine Learning (ML) model for grouping similar texts, interpreted through a hybrid Human-In-The-Loop system based on NLP (Natural Language Processing) and RAG (Retrieval-Augmented Generation).

The reach_proxy index, described at point 3 of this section, was used to measure audience. We summed this index of audience metrics (likes, comments, shares) for all unique messages belonging to a theme, obtaining the "Unique Proxy" score. In the case of multiple identical messages, we included the message with the highest values. The "Total Proxy" score was obtained by adding all reach_proxy values for all texts belonging to a theme. The difference between "Total Proxy" and "Unique Proxy" reflects the audience obtained through verbatim-distributed messages. Thus, the Unique / Total ratio shows the percentage of audience obtained from original posts, out of the total posts.

2. Sentiment analysis

Each post was analysed independently, exclusively on the basis of the content of the text column. The number of likes, comments, shares, views, and reactions was not used in classification. Texts were interpreted both in forms with and without diacritics, with negations, intensifiers, emojis, and formulations distributed throughout the text taken into account.

For informative posts, a distinction was made between the gravity of the event reported and the attitude expressed by the author: a news item about an arrest or a death may remain neutral when it does not contain an explicit emotional assessment.

The initial classification was performed through a supervised machine learning method, based on TF-IDF representations and a linear classifier trained on the Romanian corpus LaRoSeDa. Because the training corpus contains mainly positive and negative examples, the identification of the neutral and mixed categories was completed through contextual rules. The results were subsequently manually reviewed, including posts initially classified as unclassifiable.

Also extracted and analysed were 7,000 comments, according to the same criteria: positive, negative, neutral, mixed, and unclassifiable. The main limitations are the identification of subtle sarcasm and the establishment of the exact target of sentiment when the same post supports Viorel Pașca but criticises the authorities or the press.

Comments were extracted in order of posting, with a sample of 1,000 per post, to allow comparison.

3. How reach_proxy was calculated

In the absence of a direct measure of unique audience, a synthetic index of posts' amplification potential was constructed. Likes were assigned a weight of 2, comments a weight of 4, and shares a weight of 6. The weights reflect a progressive level of user engagement: liking involves a quick action and minimal effort, commenting involves formulating and expressing an opinion, and sharing directly contributes to extending the post's circulation to other user networks. The constant distance of two points between weights keeps the formula simple, transparent, and easy to apply to all posts.

The resulting value does not represent the actual number of persons who saw the post, but a comparative score of interaction intensity and propagation potential. As a sensitivity check, the indicator was also recalculated using equal weights for likes, comments, and shares. The hierarchy of the main themes was preserved, indicating that the comparative results are not determined by the choice of the above-mentioned weights.

4. Hostility analysis

The hostility analysis was conducted through a method of textual analysis assisted by a large language model (LLM), combined with human-in-the-loop verification. Each comment was analysed individually, exclusively on the basis of its text, without the sentiment label previously assigned automatically determining the result.

In a first stage, comments were classified as non-hostile, hostile, or unclear. Formulations expressing aggression, contempt, disparagement, ridicule, general accusations, or delegitimisation toward a person, institution, organisation, or other target were considered hostile. Firm but reasoned criticisms lacking the degradation of a person were not automatically classified as hostile.

For each hostile comment, the main target and any secondary targets were identified. Only comments that effectively targeted the post's author were then divided into exclusively content-based criticism, exclusively personal attack, mixed attack, or hostility without substantial criticism. Personal and mixed attacks were additionally coded by subtype, such as general insult, attack on intelligence, professional competence, or integrity, ridicule, command to be silenced, threat, or wish for punishment. Possible forms of coded gendered language were also analysed separately.

The human component included the re-verification of ambiguous comments, sarcasm, threats, coded gender markers, cases with low certainty, and harsh criticisms that could be confused with personal attacks. Following this verification, classifications were corrected where the context or meaning of the formulation required a different categorisation. Exact duplicates were not automatically removed, but were flagged and reported separately.

The sentiment, comment, and coordination analyses focused on Facebook, the platform with the most complete public interaction data.

This material is funded through the Tech Accountability Grants 2026 Programme of the European Fact-Checking Standards Network (EFCSN), under the CERV grant (no. 101236606) awarded by EACEA.

The views and opinions expressed in this material belong exclusively to the authors and do not necessarily reflect the position of the European Union or the European Education and Culture Executive Agency (EACEA). Neither the European Union nor the funding authority can be held responsible for them.