# European election manipulation and agency-transfer ledger

**Status:** working research ledger, version 2026-08-11. **Scope:** campaign/incident-level cases involving synthetic media, coordinated influence, covert targeting, recommender gaming, bots/inauthentic amplification, or AI-assisted personalization. The ledger keeps four questions separate: **did it occur; how far did it travel; who directed it; did it change behavior or an election outcome?** A legal ruling is recorded as an institutional outcome, never converted into a causal vote estimate.

## Coding and source-quality legend

- **Evidence classification:** `established` = directly documented by primary/platform/forensic evidence; `strong inference` = multiple convergent indicators but an inferential step remains; `plausible` = credible but materially incomplete; `speculative` = weak or uncorroborated; `open` = evidence does not resolve the claim.
- **Confidence:** `high` ≈ 0.80–1.00; `medium` ≈ 0.50–0.79; `low` < 0.50. Confidence is claim-specific where occurrence, reach, attribution, and effect differ.
- **Source quality:** `A` official election authority, court, intelligence/security authority, regulator, or first-party platform telemetry; `B` peer-reviewed/transparent research or major investigation with reproducible evidence; `C` major reporting with named sources/documents; `D` advocacy, secondary synthesis, or artifact archive useful mainly for discovery/corroboration.
- **Reach caution:** account, post, impression, follower, view, or engagement counts are not unique people, verified voters, belief, persuasion, or vote conversion.
- **Attribution caution:** ideological alignment, Russian-language operation, or pro-Kremlin content is not itself proof of Russian-state command.

## Core records

### 1. `ro-2024-pres-georgescu-tiktok`

| Field | Coded record |
|---|---|
| Country; election/date | Romania; presidential first round, 24 Nov 2024; process annulled 6 Dec 2024 |
| Incident/campaign | Călin Georgescu TikTok/Telegram coordinated amplification and concealed promotion finance |
| Actor and attribution status | Bogdan Peșchir was officially/platform-linked to payments supporting some promotion and became a criminal-investigation subject; guilt was not finally adjudicated in the cited record. Direction of the overall operation by Russia or any state remains **open**. SRI described a state-actor modus operandi; public evidence did not establish the command chain. Russia denied interference. A separate PNL-financed influencer brief, `#EchilibrușiVerticalitate`, is counterevidence against collapsing all pro-Georgescu content into one Russian operation. |
| AI/digital capability | Coordinated dormant-account activation; Telegram tasking (`@propagatorcg`); influencer payments; fake institutional accounts; hashtag/recommender gaming; covert promotion. Generative AI was not established as the campaign's core capability. |
| Influence vector/platform | TikTok, Telegram, influencer networks and cross-platform reposting |
| Target | Romanian voters; candidate visibility and perceptions of Georgescu's popularity/legitimacy |
| Content/deception mechanism | Undeclared paid and inauthentic promotion presented as spontaneous peer enthusiasm; instructions for posting, hashtags and moderation/evasion; false institutional identity cues; declared campaign spend of zero despite documented promotion payments. |
| Occurrence evidence | Declassified SRI material described an initial 25,000 TikTok accounts activated intensely in the final two weeks; 797 were created in 2016 and largely dormant until 11 Nov. SRI/platform material associated roughly US$381,000 in payments/donations by Peșchir with accounts promoting Georgescu, 24 Oct–24 Nov. **Classification: established for organized undeclared promotion; high (0.90).** |
| Scale/reach evidence | The 25,000 figure is an intelligence-reported account count, not unique exposure. Polling chronology—below 1% on 30 Oct–5 Nov, 6.2% on 15–20 Nov, 10.6% on 20–21 Nov, then 22.94% of votes—is temporal correlation only. TikTok's separately disclosed Turkey-based network (4,453 accounts plus 22,764 linked accounts, about three followers/account, primarily AUR and secondarily Georgescu) is a distinct scope and must not be added to SRI's 25,000. **Classification: strong but fragmented; medium-high (0.75).** |
| Behavioral/electoral effect | Georgescu won round one. No individual-level evidence or causal design establishes that the network changed votes or determined the result. The Constitutional Court found distortion of candidate equality and free formation of voter opinion; that is an institutional/legal finding, not a measured vote shift. **Classification: open; low (0.20).** |
| Agency-transfer mechanism | Covert sponsorship, algorithm gaming and manufactured peer consensus removed source/finance cues and delegated judgments of popularity and credibility to an unseen organizer and recommender system. |
| Democratic harm | Campaign-finance opacity; unequal candidate visibility; false institutional endorsement; degraded informed choice and trust. |
| Response/mitigation/outcome | Constitutional Court annulled the process; rerun held in May 2025; Georgescu was barred; criminal investigations followed. The European Commission opened DSA proceedings against TikTok—an investigation, not a liability finding. TikTok disclosed and removed several networks. |
| Overall assessment | **Established occurrence / mixed attribution / open effect. Overall confidence high for occurrence, medium-low for full actor attribution, low for causal effect.** |

**Atomic counterevidence and cautions**

- Do not report “1.4 billion TikTok views” as unique Romanian exposure or as one campaign's reach.
- Do not merge TikTok's Turkey-based network with SRI's 25,000-account estimate.
- The polling surge and vote result cannot be coded as campaign effect.
- Global Witness's fresh-account audit was conducted 5–6 Dec, after round one, and TikTok disputed the method; it is platform-risk evidence, not pre-vote conversion evidence.

**Sources**

1. [Romanian Presidency, declassified SRI report](https://www.presidency.ro/files/userfiles/Documente%20CSAT/Document%20CSAT%20SRI%20I.pdf) — 4 Dec 2024; official intelligence document; **A** for what SRI reported, not an adjudication.
2. [Constitutional Court press release](https://www.ccr.ro/en/press-release-6-dec/) — 6 Dec 2024; official judicial outcome; **A**.
3. [TikTok, continuing election-integrity update](https://newsroom.tiktok.com/en-eu/continuing-to-protect-the-integrity-of-tiktok-during-romanian-elections) — Dec 2024/update series; platform telemetry; **A**, interested party.
4. [European Commission opens DSA proceedings](https://digital-strategy.ec.europa.eu/en/news/commission-opens-formal-proceedings-against-tiktok-election-risks-under-digital-services-act) — 17 Dec 2024; regulator; **A** for procedural status.
5. [Associated Press investigation/counterevidence](https://apnews.com/article/1095e5a6420af8c25208971a8855d664) — Dec 2024; major investigation; **C**.

### 2. `ro-2025-pres-crossplatform-personas`

| Field | Coded record |
|---|---|
| Country; election/date | Romania; repeat presidential election, 4 and 18 May 2025; operation observed Dec 2024–6 Jun 2025 |
| Incident/campaign | Hijacked/fake-persona network supporting Georgescu, then George Simion, and alleging Moldovan/French manipulation |
| Actor and attribution status | Unknown covert operator. DFRLab assessed the network as pro-Russia from narratives and coordination, while explicitly noting that the operator was difficult to establish. No public proof of Russian-state command. Pavel Durov's allegation of French pressure was denied by French intelligence and supported by no disclosed evidence; it is coded as campaign content, not fact. |
| AI/digital capability | At least 215 coordinated accounts; hijacked legacy accounts; synthetic Romanian personas; generative-AI profile images/descriptions; DreamFace and D-ID avatars; stock footage passed off as current protest; coordinated bilingual hashtag sets. |
| Influence vector/platform | Facebook, Instagram, TikTok, Telegram and `ono.news` |
| Target | Romanian voters, especially Georgescu/Simion audiences; Maia Sandu, PAS, Nicușor Dan, EU and France as reputational targets |
| Content/deception mechanism | Borrowed trust from established accounts and fake local identities; synthetic presenters and pseudo-news; false claims that Romanian citizens in Moldova and France stole/manipulated the election; covert pivot from Georgescu to Simion. |
| Occurrence evidence | DFRLab identified 116 Facebook, 79 TikTok and 17 Instagram accounts, reporting a network total of at least 215, with coordinated content and persona reuse. **Established; high (0.91).** |
| Scale/reach evidence | 8,514 posts, at least 16 million views and 681,000 likes; 1,452 videos from 9–28 May; 106 videos on 19 May, 77 sharing one hashtag set; posts entered public Facebook groups with 13,300 and 17,100 members. Counts are non-unique platform metrics, and some Meta accounts fell below public-data thresholds. **Measured platform activity; high (0.84).** |
| Behavioral/electoral effect | None measured; DFRLab says efficacy is difficult to estimate. Dan won 53.60% to 46.40%. The Court's rejection of Simion's annulment complaint as unfounded rejects the legal claim, not the documented network's occurrence. **Open; low (0.15).** |
| Agency-transfer mechanism | Hijacked and synthetic “ordinary Romanian” voices and AI news presenters manufactured local consensus and outsourced fraud judgments to false social proof. |
| Democratic harm | Polluted voter information, cross-border scapegoating, concealed sponsorship and post-result delegitimation. |
| Response/mitigation/outcome | Some assets, including the Facebook `ono.news` page, were removed; TikTok/Telegram assets reportedly remained when studied. ODIHR documented persistent inauthentic behavior and fragmented oversight. Result upheld; Dan took office 2 Jun 2025. |
| Overall assessment | **Established operation and reach; pro-Russia orientation is strong inference; operator/state attribution and effect remain open.** |

**Sources**

1. [DFRLab cross-platform investigation](https://dfrlab.org/2025/08/26/cross-platform-campaign-accuses-moldovas-sandu-of-meddling-in-romanian-elections/) — 26 Aug 2025; transparent investigation; **B**.
2. [ODIHR Romania 2025 final report](https://odihr.osce.org/sites/default/files/f/documents/f/3/600295_0.pdf) — 2025; official election observation; **A**.
3. [Reuters on Durov's allegation and French denial](https://www.reuters.com/world/europe/telegram-founder-says-he-rejects-request-silence-conservative-voices-romania-2025-05-18/) — 18 May 2025; major reporting; **C**.

### 3. `md-2024-presref-shor-psb-votebuying`

| Field | Coded record |
|---|---|
| Country; election/date | Moldova; EU constitutional referendum and presidential first round, 20 Oct 2024; runoff 3 Nov |
| Incident/campaign | Șor/PSB/Telegram hierarchical vote-buying network |
| Actor and attribution status | Operational links to Ilan Șor-affiliated structures, sanctioned Russian Promsvyazbank (PSB), and Russia-based coordinators are supported by searches, accounts and reporting. Named persons' criminal guilt remained alleged where no final judgment is cited. Direct Kremlin command is an authority/investigative attribution, not publicly adjudicated. Șor and Russia denied wrongdoing/interference. |
| AI/digital capability | Hierarchical recruitment; PSB app and bank transfers; Telegram tasking immediately before voting; lists, cards and proof/monitoring processes. AI was not established as core. |
| Influence vector/platform | Telegram, PSB banking infrastructure, territorial organizers and offline payment networks |
| Target | Moldovan voters; a named presidential preference and “No” in the EU referendum |
| Content/deception mechanism | Hidden foreign/oligarchic financing and last-mile digital instructions converted political choice into a paid, monitored patronage transaction. |
| Occurrence evidence | Police searches recovered phones with the PSB app, bank documents/cards, cash and lists. Police described 130 territorial leaders, 1,927 sector heads, over 50,000 activists and over 70,000 recruited supporters; later, over 80,000 alleged recipients of MDL 800–1,500 and US$39 million identified through PSB in Sep–Oct. Because proceedings were ongoing, these figures are authority-reported allegations supported by seizures. **Strong/established operational scheme; high (0.90).** |
| Scale/reach evidence | US$39 million and 80,000 recipients are identified/estimated by police and repeated by ODIHR, not independently audited exposure. Exclude the police projection that eventual involvement might exceed 300,000. ODIHR noted that the initial CEC notification lacked names/concrete evidence; observers directly saw indications of vote buying at only 2 of 1,202 sampled polling stations. **Strong but authority-reported; medium-high (0.76).** |
| Behavioral/electoral effect | Payments and attempts to condition behavior are evidenced; secret ballots make compliance unknowable. “Yes” won 749,719 to 739,155, a 10,564 margin, but neither reversal nor the number of compliant votes can be inferred. Sandu's “150,000 bought votes” statement is not a measured causal estimate. Voting/counting were overwhelmingly positively assessed. **Open; low (0.15).** |
| Agency-transfer mechanism | Cash/bank payments plus hidden Telegram instructions substituted organizer direction and economic pressure for autonomous judgment. |
| Democratic harm | Commodified suffrage, illicit foreign finance, unequal conditions, possible coercion and trust erosion. |
| Response/mitigation/outcome | At least 115 searches, detentions, seizures, fines, anti-vote-selling messaging, and Telegram channel blocking/removal. Referendum and presidential results were validated, not annulled; ODIHR urged stronger enforcement. |
| Overall assessment | **Established/strong scheme; strong inference to Șor-linked/PSB infrastructure; Kremlin command and vote compliance open.** |

**Sources**

1. [Moldovan Police operational account](https://politia.md/ro/print/pdf/node/1524) — Oct 2024; law-enforcement primary source; **A** for reported investigative actions, not guilt.
2. [Moldovan Police English release](https://politia.md/en/node/1519) — Oct 2024; law-enforcement primary source; **A** with same limitation.
3. [Moldovan Police, 115 searches and detentions](https://politia.md/index.php/ro/content/115-perchezitii-si-3-lideri-raionali-retinuti-de-ofiterii-ini-si-procuratura-anticoruptie) — 2024; law-enforcement primary source; **A**.
4. [ODIHR 2024 presidential/referendum final report](https://odihr.osce.org/sites/default/files/documents/official_documents/2025/12/MLD%202024_presidential_const_referendum_FR_%2014.03.2025.pdf) — 14 Mar 2025 report; official observation; **A**.
5. [Reuters, alleged vote-buying scheme](https://www.reuters.com/world/europe/moldova-alleges-pro-russian-vote-buying-scheme-ahead-key-vote-2024-10-03/) — 3 Oct 2024; major reporting; **C**.

### 4. `md-2024-presref-meta-transnistria-cib`

| Field | Coded record |
|---|---|
| Country; election/date | Moldova; presidential election/referendum, 20 Oct and 3 Nov 2024 |
| Incident/campaign | Transnistria-origin fictitious Russian-language news brands using synthetic personas |
| Actor and attribution status | Meta linked the activity to individuals from Russia and Moldova operating in Transnistria, including a fake-engagement service. Meta did not attribute Russian-state control. **Platform attribution to regional operators established; state direction open.** |
| AI/digital capability | Likely GAN profile photos; fake accounts and account-name changes; about a dozen fake news brands; fake engagement; giveaways and paid graffiti; cross-platform funneling. |
| Influence vector/platform | Facebook, Instagram, with funnels to Telegram, Odnoklassniki and TikTok |
| Target | Russian-speaking Moldovan audiences; Sandu, EU integration and pro-Western policy |
| Content/deception mechanism | `Tresh Kich`, `Moldovan Mole`, `Insider Moldova`, `Gagauzia on Air` and other false local-news brands used synthetic identities to launder anti-Sandu/anti-EU, pro-Russia content as community information and fabricate social proof. |
| Occurrence evidence | Meta removed 7 Facebook accounts, 23 Pages, 1 Group and 20 Instagram accounts. **Established through platform telemetry; high (0.96).** |
| Scale/reach evidence | About 4,200 Facebook Page followers and 335,000 Instagram followers; Meta said the vast majority were outside Moldova and indicative of fake engagement; about US$4,000 in ads. Meta said it disrupted the campaign before it built an authentic audience. **Platform-reported gross metrics, low authentic Moldovan reach; medium-high (0.75).** |
| Behavioral/electoral effect | No survey, behavior, turnout or vote effect measured. **Open; low (0.05).** |
| Agency-transfer mechanism | Fake local-news identities and synthetic personas concealed provenance and made operator-authored propaganda appear to be independent community judgment. |
| Democratic harm | Source laundering, false public opinion, polarization and degraded provenance cues. |
| Response/mitigation/outcome | Platform takedown before the election; wider ODIHR monitoring. No actor-specific judicial outcome in the cited record. |
| Overall assessment | **Established CIB; reach limited/contested; state attribution and effect open.** |

**Sources**

1. [Meta, coordinated inauthentic behavior in Moldova](https://about.fb.com/news/2024/10/taking-action-against-coordinated-inauthentic-behavior-in-moldova/) — 11 Oct 2024; platform telemetry; **A**, interested party.
2. [ODIHR 2024 final report](https://odihr.osce.org/sites/default/files/documents/official_documents/2025/12/MLD%202024_presidential_const_referendum_FR_%2014.03.2025.pdf) — 14 Mar 2025; official observation; **A**.
3. [Reuters on Meta removal](https://www.reuters.com/technology/meta-removes-fake-accounts-moldova-ahead-presidential-election-2024-10-11/) — 11 Oct 2024; major reporting; **C**.

### 5. `md-2025-parl-infoleaders`

| Field | Coded record |
|---|---|
| Country; election/date | Moldova; parliamentary election, 28 Sep 2025 |
| Incident/campaign | Paid “InfoLeaders” digital army recruited and instructed through Telegram |
| Actor and attribution status | Undercover investigations linked organizers to Șor-affiliated/Evrazia personnel and Russian-speaking curators; payments were promised through sanctioned Russian state-owned PSB. A curator said funds came “directly from Moscow”; that and operational links support strong inference, not independent proof of Kremlin command. Șor/Evrazia personnel did not answer some inquiries; Russia denied interference. |
| AI/digital capability | Paid local persona management; ChatGPT rephrasing/satirization of briefs while preserving an “organic” voice; fake/unlawful polls; coordinated hashtags and reciprocal engagement; Telegram bots for proof-of-task and payout validation; covert filming instructions. |
| Influence vector/platform | TikTok, Facebook, Instagram and Telegram |
| Target | Moldovan voters, especially persuadable/older audiences; PAS, Sandu, EU and NATO as attack targets |
| Content/deception mechanism | Salaried political workers posed as spontaneous peers; AI adapted central briefs; fake polls and mutual engagement manufactured consensus; themes included fraud, war/conscription, child trafficking/LGBT threats and Western control. |
| Occurrence evidence | BBC/ZDG undercover reporting documented recruitment/training; DFRLab matched 253 inauthentic accounts. **Established/strong; high (0.93).** |
| Scale/reach evidence | BBC: at least 90 TikTok accounts, thousands of videos, over 23 million views and 860,000 likes since Jan. DFRLab: 253 accounts (173 TikTok, 59 Facebook, 16 Instagram), 28,708 posts, 55,367,950 views, 2,207,894 likes, 295,608 shares and 239,016 comments; about 200 recruits by Aug, not necessarily all active. Metrics are non-unique and do not establish authentic Moldovan exposure. **Measured platform activity; high (0.90).** |
| Behavioral/electoral effect | None isolated. PAS won 792,557 votes and 55 seats. On 16 Oct the Constitutional Court found no violations capable of significantly affecting voter will and validated results; that does not negate the network. **Open; low (0.10).** |
| Agency-transfer mechanism | Covert salary and generative AI converted foreign/oligarchic briefs into apparently authentic local speech, hiding the sponsor and substituting manufactured peer judgment. |
| Democratic harm | Hidden foreign finance, false public opinion, potential intimidation/surveillance and pre-emptive fraud narratives. |
| Response/mitigation/outcome | Exposed before voting by BBC/ZDG; most identified accounts later deplatformed, according to DFRLab. Election-wide TikTok removals cannot all be assigned to this network. ODIHR described serious foreign interference/disinformation and inadequate platform response; Court validated result. |
| Overall assessment | **Established operation and measured output; strong operational link to Șor/Evrazia/PSB; direct Kremlin command and electoral effect unresolved.** |

**Sources**

1. [BBC undercover investigation](https://www.bbc.com/news/articles/c4g5kl0n5d2o) — Sep 2025; major undercover investigation; **B/C**.
2. [DFRLab, Russia-linked digital army](https://dfrlab.org/2025/09/24/paid-to-post-russia-linked-digital-army-seeks-to-undermine-moldovan-election/) — 24 Sep 2025; transparent network analysis; **B**.
3. [Ziarul de Gardă undercover investigation](https://www.zdg.md/en/kremlins-digital-army-an-undercover-investigation-its-paid-let-me-tell-you-its-paid-directly-from-moscow/) — 2025; primary undercover evidence; **B**.
4. [ODIHR Moldova 2025 parliamentary final report](https://odihr.osce.org/sites/default/files/documents/official_documents/2026/02/MLD%20Parliamentary%202025%20Final%20Report_18.02.2025_ENG.pdf) — 2026 publication; official observation; **A**. **Second review:** filename date appears inconsistent with the election year; verify metadata before publication.

### 6. `md-2025-parl-storm1516`

| Field | Coded record |
|---|---|
| Country; election/date | Moldova; parliamentary election, 28 Sep 2025 |
| Incident/campaign | Storm-1516 synthetic/staged local-news and ballot-fraud evidence operation |
| Actor and attribution status | G7 Rapid Response Mechanism assessed with high confidence that monitored activity linked to known Russian state-aligned FIMI operations, including Storm-1516. This is governmental analytic attribution, not a judicial finding or full command-chain proof. Russia generally denied interference. |
| AI/digital capability | At least 45 new fake sites; likely AI-generated video; spoofed official/media pages and logos; fake documents, audio and whistleblowers; likely staged ballot-destruction footage; coordinated X accounts, including verified accounts. |
| Influence vector/platform | Fake Romanian/Moldovan-looking sites, X, Telegram, TikTok and Portal Kombat/Pravda ecosystem |
| Target | Moldovan and Romanian audiences; Sandu, PAS, Parliament Speaker Igor Grosu, EU/NATO, diaspora voting and election legitimacy |
| Content/deception mechanism | Fake PAS site/video; fabricated Grosu audio about sending citizens to Ukraine; fake EC/Kaja Kallas conscription letter; false media stories and whistleblowers; discarded/burned opposition-ballot videos; false male travel ban/military emergency. |
| Occurrence evidence | G7 monitoring identified a coherent Storm-1516 artifact set and infrastructure. **Established/strong; high (0.91).** |
| Scale/reach evidence | G7 said Storm had the widest reach/attention among monitored phase-two operations; individual items sometimes exceeded 1 million views. A male travel-ban video exceeded 2.7 million views; an English-subtitled discarded/burned-ballot video received at least 5 million X views. At least 45 new sites were believed part of a network exceeding 250 sites. Do not assign G7's more than 450 total FIMI items to Storm alone. **Strong platform metrics; high (0.86).** |
| Behavioral/electoral effect | No causal estimate for belief, turnout or vote choice. PAS won 55 seats; Court validated mandates after finding no proven violation that significantly changed voter will. **Open; low (0.10).** |
| Agency-transfer mechanism | Impersonated institutions/media and synthetic or staged “evidence” replaced source verification and pre-loaded fraud/war interpretations. |
| Democratic harm | Fear of conscription/war; distrust in ballots, diaspora voting, legitimate media and institutions; post-election denialism. |
| Response/mitigation/outcome | G7 and fact-checkers exposed artifacts; platforms conducted wider election enforcement but aggregate removals cannot be mapped to Storm without artifact-level data. ODIHR documented AI videos, troll farms and bots and judged response inadequate. Result validated. |
| Overall assessment | **Strong, officially assessed Russian state-aligned operation; direct command chain and effect unresolved.** |

**Sources**

1. [G7 RRM, 2025 Moldovan election FIMI assessment](https://international.canada.ca/en/global-affairs/corporate/reports/rapid-response-mechanism/news/2025-moldovan-election) — 2025; multi-government analytic assessment; **A**.
2. [ODIHR Moldova 2025 final report](https://odihr.osce.org/sites/default/files/documents/official_documents/2026/02/MLD%20Parliamentary%202025%20Final%20Report_18.02.2025_ENG.pdf) — 2026; official observation; **A**; metadata should be checked.
3. [TikTok Moldova election-safety update](https://newsroom.tiktok.com/safety-during-elections-en?from_seo_redirect=1&lang=ro-MD) — 2025; platform primary source; **A**, but aggregate enforcement is not campaign-specific.

### 7. `md-2025-parl-telegram-ai-bots`

| Field | Coded record |
|---|---|
| Country; election/date | Moldova; parliamentary election, 28 Sep 2025; observed 1 Jul–15 Sep |
| Incident/campaign | Context-adaptive AI bot commenting network on Moldovan Telegram |
| Actor and attribution status | Unknown operators. OpenMinds assessed a Russian operation because 347 of 462 accounts also operated in Russian/Ukrainian channels, including eight channels in Russian-occupied Ukrainian territories, and echoed Kremlin war narratives. This is strong inference from cross-deployment, not official/forensic proof of Russian-state command. |
| AI/digital capability | Context-aware, overwhelmingly non-duplicate generated comments; automated high-volume posting; observable generation failure modes; link-spam promoting an AI anti-Sandu video. |
| Influence vector/platform | Telegram comment sections in 76 of 88 open-comment political/news channels |
| Target | Sandu, PAS, election legitimacy and protest mobilization; mainly Russian-language audiences, with 57 accounts also posting Romanian |
| Content/deception mechanism | Synthetic pseudo-citizens adapted attacks to host posts and discussion, making coordinated propaganda look like spontaneous public opinion; themes alleged betrayal, hunger/war, vote rigging and need for protest. At least 1,800 comments explicitly called for protest. |
| Occurrence evidence | OpenMinds' algorithmic detection plus human verification identified 462 bot-like accounts and recurrent AI failure modes; one account posted 2,815 comments, 2,620 unique, under a Sandu post. **Strong inference, high (0.88), but non-platform study.** |
| Scale/reach evidence | Exactly 62,229 comments, 12.8% of observed comment activity, over 14,000 reactions; in 16 channels bots supplied over half of comments. Counterevidence: 118 bots posted one AI-video link 3,639 times, yet the video had only about 1,400 views on 16 Sep. Output is not readership. **Measured activity high (0.90); authentic exposure medium (0.55).** |
| Behavioral/electoral effect | No evidence of changed belief, protest, turnout or votes. Court validation does not disprove the network. **Open; low (0.05).** |
| Agency-transfer mechanism | Adaptive synthetic peers created a false social norm and apparent community judgment, transferring agenda-setting from authentic participants to an unseen operator. |
| Democratic harm | Crowding out genuine discussion; fabricated opposition consensus; normalization of fraud and protest claims. |
| Response/mitigation/outcome | Telegram periodically removed accounts and one channel erased/closed comments; replacements reportedly appeared. No public actor-specific sanction. ODIHR later documented election-wide AI video, troll farms and automated bots and found platform response inadequate. |
| Overall assessment | **Strongly evidenced network and output; Russian-operation attribution is strong inference; state direction, authentic reach and effect open.** |

**Sources**

1. [OpenMinds, AI-driven Telegram bots in Moldova](https://www.openminds.ltd/reports/how-russia-uses-ai-driven-bots-on-telegram-to-meddle-in-moldovas-elections) — 2025; method-led research report; **B**.
2. [ODIHR Moldova 2025 final report](https://odihr.osce.org/sites/default/files/documents/official_documents/2026/02/MLD%20Parliamentary%202025%20Final%20Report_18.02.2025_ENG.pdf) — 2026; official observation; **A**; metadata should be checked.

### 8. `sk-2023-parl-simecka-todova-audio`

| Field | Coded record |
|---|---|
| Country; election/date | Slovakia; parliamentary election, 30 Sep 2023; artifact released 28 Sep during the moratorium |
| Incident/campaign | Synthetic audio impersonating Michal Šimečka and journalist Monika Tódová discussing vote buying/rigging |
| Actor and attribution status | Origin unknown. Amplification by Štefan Harabin and former MP Peter Marček is documented. Similarity to an SVR narrative and pro-Kremlin amplification do not establish Russian-state authorship; a Slovak official said proof was absent. |
| AI/digital capability | AI voice cloning, “leaked call” framing and rapid cross-platform amplification during a period when candidates/media had constrained response options |
| Influence vector/platform | Telegram, Facebook, TikTok, YouTube, email and WhatsApp |
| Target | Šimečka/Progressive Slovakia, Tódová/media credibility, Roma voters and confidence in electoral integrity |
| Content/deception mechanism | Authentic-sounding cloned voices falsely discussed buying Roma votes and manipulating results; release during the moratorium increased ambiguity and reduced corrective capacity. |
| Occurrence evidence | Audio was forensically/ journalistically identified as AI-generated and debunked. **Established; high (0.96).** |
| Scale/reach evidence | Harabin's Telegram post received roughly 13,000 views/500 shares; Marček's Facebook post about 50,000; total Facebook exposure exceeded 100,000 and Instagram about 30,000 in cited research. Non-unique and cross-platform overlap unknown. **Measured/estimated; medium-high (0.80).** |
| Behavioral/electoral effect | No causal study. SMER's victory is not effect evidence; major parties accepted the result. Scholarship calls the claim that the deepfake “swung” the election reductive. **Open; low (0.10).** |
| Agency-transfer mechanism | Cloned voice and leak conventions transferred the credibility of known people and private evidence to a synthetic artifact. |
| Democratic harm | Reputational harm, voter-fraud narratives, anti-Roma manipulation and impaired correction during moratorium. |
| Response/mitigation/outcome | AFP debunked same day; Facebook labeled content; police investigations were halted/reopened, with a third investigation reported by Nov 2024 under defamation because no specific deepfake offence applied. |
| Overall assessment | **Established synthetic-media incident and substantial circulation; actor and behavioral/electoral effect open.** |

**Sources**

1. [CEDMO, police investigation and reach analysis](https://cedmohub.eu/how-ai-generated-content-influenced-parliamentary-elections-in-slovakia-the-slovak-police-will-investigate-the-recording-for-a-third-time/) — 2024; regional research/fact-check synthesis; **B**.
2. [Harvard Misinformation Review, Slovak case](https://misinforeview.hks.harvard.edu/article/beyond-the-deepfake-hype-ai-democracy-and-the-slovak-case/) — 2024; scholarly analysis; **B**.
3. [VSquare investigation](https://vsquare.org/slovak-election-targeted-by-pro-kremlin-deepfake-hoax/) — 2023; major regional investigation; **B/C**.
4. [ODIHR Slovakia 2023 final report](https://odihr.osce.org/sites/default/files/f/documents/6/b/563625_0.pdf) — 2023/2024; official observation; **A**.

### 9. `fr-2017-pres-macronleaks`

| Field | Coded record |
|---|---|
| Country; election/date | France; presidential runoff, 7 May 2017; leak released 5 May |
| Incident/campaign | `#MacronLeaks` hack-and-leak with forged/misleading material and automated amplification |
| Actor and attribution status | In 2025 France officially attributed APT28/GRU cyber operations intended to destabilize the 2017 election. The precise authorship of every forged item, troll cluster and amplifier remains mixed; US alt-right actors and WikiLeaks independently amplified. Do not force all activity into a single actor model. |
| AI/digital capability | Cyber theft, bulk release, admixture of authentic and forged/miscontextualized material, troll/bot amplification and search/social trending; no core generative-AI role established. |
| Influence vector/platform | File-sharing sites, Twitter/X hashtags, WikiLeaks, forums and news/search ecosystems |
| Target | Emmanuel Macron, campaign legitimacy and French voters immediately before the runoff |
| Content/deception mechanism | A large stolen corpus created authenticity cover for rumors and altered/forged items; timing exploited the electoral blackout to constrain rebuttal and seize attention. |
| Occurrence evidence | About 15 GB/21,075 emails were released; forensic and official records document the operation. **Established; high (0.97).** |
| Scale/reach evidence | `#MacronLeaks` approached 500,000 tweets in 24 hours in cited analyses; tweet volume is not unique French exposure and much activity came from foreign/ideological communities. **Measured online activity; high (0.85).** |
| Behavioral/electoral effect | Post-mortem research assesses the operation as failing to sway the result; Macron won 66.1%. No individual causal estimate, so code electoral effect as not detected rather than a quantified zero. **Not detected/open; medium (0.65) for no discernible aggregate effect.** |
| Agency-transfer mechanism | Stolen authentic data plus fabricated content and automated salience transferred agenda control to attackers while exploiting the evidentiary authority of a “leak.” |
| Democratic harm | Privacy/campaign security breach, source confusion, agenda capture and compressed correction window. |
| Response/mitigation/outcome | Macron campaign warned of mixed authenticity; election commission cautioned media/public; French media restraint and blackout reduced spread; investigation and later official attribution followed. |
| Overall assessment | **Established occurrence and official GRU/APT28 campaign-level attribution; component attribution mixed; no demonstrated electoral effect.** |

**Sources**

1. [French Foreign Ministry attribution of cyberattacks to GRU](https://www.diplomatie.gouv.fr/en/presse-et-ressources/decouvrir-et-informer/actualites/russie-attribution-de-cyberattaques-contre-la-france-au-service-de-renseignement-militaire-russe) — 29 Apr 2025; official attribution; **A**.
2. [Atlantic Council, Macron Leaks post-mortem](https://www.atlanticcouncil.org/wp-content/uploads/2019/06/The_Macron_Leaks_Operation-A_Post-Mortem.pdf) — 2019; detailed research report; **B**.
3. [Ferrara, Disinformation and Social Bot Operations in the Run Up to the 2017 French Presidential Election](https://arxiv.org/abs/1707.00086) — 2017; computational research preprint; **B**.
4. [The Guardian, French warning not to publish leak](https://www.theguardian.com/world/2017/may/06/french-warned-not-to-publish-emmanuel-macron-leaks) — 6 May 2017; contemporaneous reporting; **C**.

### 10. `fr-2024-eu-lepen-nieces`

| Field | Coded record |
|---|---|
| Country; election/date | France; European Parliament election, 9 Jun 2024; accounts surfaced in Apr |
| Incident/campaign | Fake “Amandine Le Pen,” “Léna Maréchal” and “Chloé” personas |
| Actor and attribution status | Operator unknown. RN and Reconquête denied involvement. An anonymous administrator reportedly characterized it as a “social experiment,” but that self-description is not reliable attribution. |
| AI/digital capability | Face swaps and synthetic personas built from stolen influencer videos; fictitious family relationships; AI-enhanced glamour/parasocial content |
| Influence vector/platform | TikTok and reposting ecosystems |
| Target | Young/social-media voters; Marine Le Pen/Marion Maréchal audiences; RN/Reconquête support |
| Content/deception mechanism | Attractive fake relatives appeared to endorse far-right politics, exploiting kinship, celebrity and parasocial trust; some content included racial mockery. |
| Occurrence evidence | Journalists preserved and examined the accounts and source videos; some accounts were removed 16 Apr. **Established; high (0.92).** |
| Scale/reach evidence | “Amandine” exceeded 30,000 followers, with some videos over 600,000 views; “Léna” approached 30,000, with some videos exceeding 1 million views. Counts are not unique users/voters. **Platform-reported/observed; medium-high (0.80).** |
| Behavioral/electoral effect | No measured persuasion/vote effect. CETaS found no evidence that AI content affected the French/EU outcome. **Open/not detected; low (0.10).** |
| Agency-transfer mechanism | False family identity and parasocial glamour converted celebrity/kinship cues into political trust without disclosing the real speaker. |
| Democratic harm | Identity deception, reputational harm, covert political persuasion and racialized content. |
| Response/mitigation/outcome | Account removal and media/fact-check exposure; party denials. |
| Overall assessment | **Established synthetic-persona incident with material views; actor and effect unknown.** |

**Sources**

1. [Euractiv, viral Le Pen-family deepfakes](https://www.euractiv.com/news/viral-deepfake-videos-of-le-pen-family-reminder-that-content-moderation-is-still-not-up-to-par-ahead-of-eu-elections/) — Apr 2024; major EU-policy reporting; **C**.
2. [Le Monde, AI enters the European-election campaign](https://www.lemonde.fr/politique/article/2024/04/30/europeennes-quand-l-intelligence-artificielle-se-mele-a-la-campagne_6230817_823448.html) — 30 Apr 2024; major investigation; **C**.
3. [CETaS, AI-enabled influence operations threat analysis](https://cetas.turing.ac.uk/publications/ai-enabled-influence-operations-threat-analysis-2024-uk-and-european-elections) — 2024; research synthesis; **B**.
4. [AI Incident Database record 737](https://incidentdatabase.ai/cite/737/) — 2024 entry; artifact index; **D**.

### 11. `fr-2024-leg-macron-bonus-copycop`

| Field | Coded record |
|---|---|
| Country; election/date | France; snap legislative election, first round 30 Jun 2024; artifact active by 26 Jun |
| Incident/campaign | CopyCop/Russia-linked fake Renaissance “€100 Macron voting bonus” site |
| Actor and attribution status | DFRLab identified the operation as a later CopyCop iteration and Russia-linked through infrastructure/content indicators. This is strong research inference; Cyrillic artifacts, source patterns and time-zone clues are not themselves official state attribution. |
| AI/digital capability | Large language model rewriting; exposed prompts; cloned/typosquatted campaign interface; automated laundering of copied journalism; cross-platform seeding |
| Influence vector/platform | `ensemble-24.fr`, `franceencolere.fr`, Telegram/Slavyangrad and influencer amplification |
| Target | French voters and Macron/Renaissance coalition credibility |
| Content/deception mechanism | A typo-domain impersonated the governing coalition and falsely offered €100 to social-benefit recipients who voted for it; supporting sites used LLM prompts to rewrite conservative/anti-Macron material while mimicking news outlets. |
| Occurrence evidence | DFRLab documented the site/infrastructure. `franceencolere.fr` contained 5,486 articles, most reposts from a previously Russia-sponsored site; 237 exposed LLM prompts; 52 copied outlets, 14 exposed prompts and 38 successful rewrites in the analyzed subset. **Established; high (0.95).** |
| Scale/reach evidence | Corpus size and exposed prompts measure production, not audience. Unique visits/exposure remain unknown. **Production measured, reach unknown; medium-low (0.40).** |
| Behavioral/electoral effect | No measured effect; broader CETaS assessment found no evidence AI-enabled operations changed the result. **Open; low (0.05).** |
| Agency-transfer mechanism | A fake official interface and purported material inducement commandeered institutional trust, while automated rewriting disguised operator-controlled propaganda as independent journalism. |
| Democratic harm | Candidate impersonation, bribery narrative, source laundering and scalable information pollution. |
| Response/mitigation/outcome | Public exposure and fact-checking; no cited actor-specific judicial remedy. |
| Overall assessment | **Established AI-assisted deception; Russia-link strong inference; audience and effect unresolved.** |

**Sources**

1. [DFRLab, operation targets French snap elections with AI-generated content](https://dfrlab.org/2024/07/04/operation-targets-french-snap-elections-using-ai-generated-content/) — 4 Jul 2024; transparent technical investigation; **B**.
2. [Columbia Journalism Review/Tow Center, Russian disinformation networks](https://www.cjr.org/tow_center/russian-disinformation-networks-ramp-up-attacks-on-european-elections.php) — 2024; research journalism; **B/C**.
3. [CETaS 2024 threat analysis](https://cetas.turing.ac.uk/publications/ai-enabled-influence-operations-threat-analysis-2024-uk-and-european-elections) — 2024; research synthesis; **B**.

### 12. `uk-2024-ge-x-botlike-network`

| Field | Coded record |
|---|---|
| Country; election/date | United Kingdom; general election, 4 Jul 2024; monitored from 22 May |
| Incident/campaign | Ten high-volume bot-like X accounts amplifying Reform/pro-Kremlin and anti-Labour themes |
| Actor and attribution status | Operator unknown. “Bot-like” is based on behavioral red flags and volume, not an X confirmation of automation. No evidence Reform UK funded or directed the accounts. |
| AI/digital capability | High-volume posting and possible automation; coordinated hashtag flooding; recommender/trending manipulation. Generative AI not established. |
| Influence vector/platform | X/Twitter hashtags and recommender feeds |
| Target | UK voters; Labour, immigration debate, Reform visibility and pro-Ukraine consensus |
| Content/deception mechanism | Flooding fabricated apparent grassroots volume and social proof; content included hate, conspiracies and pro-Putin messages. All five accounts using `#LabourLosing` promoted Reform; two of three using `#stoptheboats` did likewise. |
| Occurrence evidence | Global Witness identified 10 bot-like accounts posting over 60,000 times since 22 May; eight were political. Behavioral classification remains partly inferential. **Partly confirmed/strong inference; medium-high (0.76).** |
| Scale/reach evidence | Estimated 150 million displays/impressions. This is modeled/non-unique, not 150 million people, UK voters or authentic engagements. **Estimated; medium-high (0.70).** |
| Behavioral/electoral effect | No causal evidence. CETaS observed exposure largely among aligned minorities and possible reinforcement/emboldening, but no result impact. **Open; low (0.10).** |
| Agency-transfer mechanism | Automated flooding and false social proof made a small number of operators appear like a large public constituency and manipulated attention allocation. |
| Democratic harm | Distorted perceived consensus, hate amplification, foreign-policy polarization and opaque agenda setting. |
| Response/mitigation/outcome | Public investigation; X reportedly did not respond before publication. No party-level liability established. |
| Overall assessment | **High-volume activity and estimated impressions credible; automation degree, actor and effect unresolved.** |

**Sources**

1. [Global Witness investigation](https://globalwitness.org/en/campaigns/digital-threats/investigation-reveals-content-posted-by-bot-like-accounts-on-x-has-been-seen-150-million-times-ahead-of-the-uk-elections/) — 2024; advocacy investigation with disclosed criteria; **B/D**.
2. [ITV News coverage](https://www.itv.com/news/2024-07-01/bot-like-x-accounts-having-oversized-influence-on-general-election-study) — 1 Jul 2024; major reporting; **C**.
3. [CETaS 2024 threat analysis](https://cetas.turing.ac.uk/publications/ai-enabled-influence-operations-threat-analysis-2024-uk-and-european-elections) — 2024; research synthesis; **B**.

### 13. `uk-2024-ge-nigeria-facebook-network`

| Field | Coded record |
|---|---|
| Country; election/date | United Kingdom; general election, 4 Jul 2024 |
| Incident/campaign | Nigeria-administered Facebook network using AI images and microtargeted pro-Farage content |
| Actor and attribution status | Meta/ABC established that administrators were mostly in Nigeria. Pro-Kremlin narratives and operational hallmarks do not prove Russian direction. No evidence Reform UK was involved; local branches sometimes shared memes. |
| AI/digital capability | AI-generated anti-immigration imagery; coordinated Pages; deceptive location/identity; paid demographic targeting |
| Influence vector/platform | Five Facebook Pages, organic posts and paid ads |
| Target | UK voters; paid pro-Farage ads targeted British men aged 65+ |
| Content/deception mechanism | Foreign-administered Pages posed as UK grassroots voices; synthetic imagery and inflated/fake follower signals promoted anti-immigration, pro-Kremlin and some pro-Reform content. Ads were paid in Nigerian currency. |
| Occurrence evidence | ABC identified the coordinated Pages and administrators; Meta removed all five for deceptive inauthentic behavior. **Established; high (0.95).** |
| Scale/reach evidence | About 190,000 combined followers, but Meta said most were outside the UK, primarily Vietnam, consistent with fake engagement. Genuine UK exposure is unknown. **Gross platform metric established, authentic reach qualitative/unknown; medium (0.62).** |
| Behavioral/electoral effect | No evidence of belief, turnout or vote change. **Open; low (0.05).** |
| Agency-transfer mechanism | Microtargeting older men, fake UK identity and synthetic/fake social proof concealed the real speaker and substituted covertly curated cues for independent judgment. |
| Democratic harm | Covert foreign participation, demographic targeting, xenophobic content and distorted popularity. |
| Response/mitigation/outcome | Meta removed all five Pages following ABC inquiries; party involvement not established. |
| Overall assessment | **Established Nigeria-operated deceptive network; Russian-state direction and effect open; gross followers overstate UK reach.** |

**Sources**

1. [ABC News investigation](https://www.abc.net.au/news/2024-06-29/uk-election-pro-russian-facebook-pages-coordinating/104038246) — 29 Jun 2024; major investigation; **B/C**.
2. [ABC News, Meta takedown](https://www.abc.net.au/news/2024-07-01/meta-shuts-down-pro-russian-facebook-pages-in-uk-elections/104045286) — 1 Jul 2024; platform-response reporting; **C**.
3. [CETaS 2024 threat analysis](https://cetas.turing.ac.uk/publications/ai-enabled-influence-operations-threat-analysis-2024-uk-and-european-elections) — 2024; research synthesis; **B**.

### 14. `de-2025-bund-storm1516-rfbi`

| Field | Coded record |
|---|---|
| Country; election/date | Germany; Bundestag election, 23 Feb 2025 |
| Incident/campaign | Storm-1516/R-FBI synthetic witness and fake-news-site smear operation |
| Actor and attribution status | Researchers linked infrastructure/content to the John Mark Dougan/R-FBI ecosystem; Dougan denied involvement. On 20 Dec 2025 the German federal government publicly attributed Storm-1516 to the Russian Federation. Exact responsibility for every R-FBI asset/domestic amplifier may differ; keep artifact-level attribution scoped. |
| AI/digital capability | Approximately 100–102 AI-generated fake-news sites; synthetic witnesses; fabricated scandals; fake documents/media; coordinated X/Telegram/TikTok dissemination |
| Influence vector/platform | Fake domains, X, Telegram, TikTok and domestic political/influencer amplification |
| Target | Robert Habeck, Friedrich Merz, CDU/Greens, pro-Ukraine policy and confidence in German media/elections |
| Content/deception mechanism | Synthetic pseudo-media and witnesses alleged scandals/corruption and supplied false “corroboration”; domestic and AfD-linked accounts amplified some artifacts. |
| Occurrence evidence | Correctiv/Alliance4Europe and official German material documented the network and artifacts. **Established; high (0.96).** |
| Scale/reach evidence | 3,280 X posts linking German domains, 1 Jul–11 Feb; a TikTok account's five videos included four viral clips totaling about 3.9 million views, one at 1.7 million; 692 Telegram posts in 395 channels generated about 4.7 million message views. Non-unique and cross-platform overlap unknown. **Measured; high (0.90).** |
| Behavioral/electoral effect | No causal evidence of belief/vote/result effects. **Open; low (0.05).** |
| Agency-transfer mechanism | Synthetic witnesses, fake outlets and domestic laundering manufactured independent corroboration for operator-written claims. |
| Democratic harm | Candidate defamation, false epistemic authority, foreign interference and trust degradation. |
| Response/mitigation/outcome | Fact-checks, public warnings, platform actions and later official attribution; no demonstrated tally effect. |
| Overall assessment | **Established network with high-confidence/official campaign attribution to Russia; exact sub-asset ownership and effect unresolved.** |

**Sources**

1. [German Interior Ministry, disinformation and hybrid threats](https://www.bmi.bund.de/SharedDocs/schwerpunkte/EN/disinformation/article-disinformation-hybrid-threat.html) — updated 2025; official government material; **A**.
2. [Alliance4Europe, Storm-1516 and German election](https://alliance4europe.eu/storm-1516-german-elections) — 2025; network research; **B/D**.
3. [Correctiv investigation](https://correctiv.org/en/fact-checking-en/2025/01/24/disinformation-operation-russian-meddling-in-german-election-campaign-exposed/) — 24 Jan 2025; major forensic/fact-check investigation; **B**.
4. [Reuters, Russia-linked AI sites](https://www.reuters.com/world/europe/russia-linked-ai-websites-aim-dupe-german-voters-study-finds-2025-01-23/) — 23 Jan 2025; major reporting; **C**.

### 15. `hu-2026-parl-proorban-tiktok-cio`

| Field | Coded record |
|---|---|
| Country; election/date | Hungary; parliamentary election, 12 Apr 2026 |
| Incident/campaign | Anonymous pro-Orbán TikTok covert influence operation using generative-AI videos |
| Actor and attribution status | TikTok confirmed the accounts were part of a covert influence operation and banned them; origin/commissioner was not publicly traced. Do not conflate with the separate Russian Matryoshka campaign. |
| AI/digital capability | AI profile pictures; synthetic presenters/experts; talking animals; celebrity deepfakes; high-volume anonymous account creation and video production |
| Influence vector/platform | TikTok |
| Target | Péter Magyar/Tisza, anti-Fidesz voters and perceptions of war/security; pro-Orbán persuasion |
| Content/deception mechanism | Anonymous pseudo-news and synthetic authority figures attacked Magyar and promoted Orbán; AI celebrities/characters supplied false authority, familiarity and emotional fear cues. |
| Occurrence evidence | NewsGuard identified 34 anonymous accounts and hundreds of AI videos; 22 accounts were created within two days in January. TikTok confirmed CIO status and removed them. **Established; high (0.96).** |
| Scale/reach evidence | Roughly 10 million total views; six fake news reports drew about 385,000 views/5,400 likes; celebrity deepfakes about 105,000 views. These are non-unique platform views. **Measured/research reported; high (0.88).** |
| Behavioral/electoral effect | None measured. Tisza reportedly won a landslide/137 seats; the network's failure to prevent that outcome does not show no attitude effects. **Open; low (0.05).** |
| Agency-transfer mechanism | Synthetic anchors, experts and celebrities commandeered authority/familiarity cues, while anonymity concealed the campaign principal. |
| Democratic harm | Covert campaigning, opponent defamation, fear manipulation and false source authority. |
| Response/mitigation/outcome | TikTok removed the confirmed network; ODIHR noted generative-AI/manipulative content used to discredit opponents. Election proceeded and opposition won. |
| Overall assessment | **Established platform-confirmed CIO with substantial views; sponsor and behavioral/electoral effect open.** |

**Sources**

1. [NewsGuard special report](https://www.newsguardtech.com/special-reports/influence-campaign-uses-ai-tiktok-videos-to-boost-hungarys-viktor-orban) — 2026; transparent monitoring investigation; **B**.
2. [The Record, TikTok removes covert networks](https://therecord.media/tiktok-removes-covert-networks-hungary-vote) — 2026; specialist major reporting with platform confirmation; **C**.
3. [ODIHR Hungary 2026 preliminary statement](https://odihr.osce.org/sites/default/files/documents/official_documents/2026/04/HUN%20Parliamentary%20Elections%202026%20PS_13.04.2026.pdf) — 13 Apr 2026; official observation; **A**.
4. [Reuters, Hungary election](https://www.reuters.com/world/europe/hungarians-vote-landmark-election-closely-watched-by-eu-russia-us-2026-04-11/) — 11/12 Apr 2026; major reporting; **C**. **Second review:** verify that final seat figure is supported by a final-results URL rather than this pre-election story.

## Additional comparator records (include with review flags)

These four records bring the package to **seven Romania/Moldova anchors plus twelve European comparators**. They are useful but have a narrower evidence base or a scope/attribution issue that requires a second publication pass.

### 16. `sk-2024-pres-korcok-stb-hoax` — **review flag: actor/network scope**

| Field | Coded record |
|---|---|
| Country; election/date | Slovakia; presidential election, 23 Mar and 6 Apr 2024 |
| Incident/campaign | Fabricated StB documents and fake-account amplification accusing Ivan Korčok |
| Actor and attribution status | The fabricated allegation was promoted by lawyer/influencer Zoroslav Kollár and amplified through accounts displaying inauthentic indicators. No foreign-state direction established. |
| AI/digital capability | Fabricated documentary “evidence,” coordinated/fake-looking Facebook accounts and high-reach video; generative AI not established. |
| Vector/target/mechanism | YouTube/Facebook; targeted Korčok by laundering a false secret-police history through documents and apparent peer corroboration. |
| Occurrence and reach | Investigation reports roughly 230,000 YouTube views, at least 9,500 Facebook shares, and a most-viral post with about 2,600 shares. Of sampled commenters, 74/194 on the allegation and 80/239 on a debunk reportedly showed fake-account signs. **Occurrence established (0.88); reach measured (0.78).** |
| Effect | No causal evidence of belief or votes; Pellegrini's victory is not effect evidence. **Open (0.05).** |
| Agency transfer/harm | False archival evidence plus inauthentic consensus displaced provenance checking; reputational and historical-memory manipulation. |
| Response | Debunks and investigation; no cited actor-specific judicial outcome. |

**Source:** [VSquare, fake accounts and StB hoax in Slovak election](https://vsquare.org/slovakia-election-ivan-korcok-peter-pellegrini-fake-accounts-facebook-hoax/) — 2024; regional investigation; **B/C**. **Second review required:** add election-authority/platform source and validate the account-classification method.

### 17. `fr-2024-leg-zelenska-bugatti` — **review flag: direct election targeting not proven**

| Field | Coded record |
|---|---|
| Country; election/date | France; legislative election, between rounds, 1 Jul 2024 |
| Incident/campaign | False claim that Olena Zelenska bought a Bugatti in Paris using Western aid |
| Actor and attribution status | Researchers linked the fake outlet/video ecosystem to a Russia/Dougan-aligned operation; exact state command remains inferential. Dealership denied the transaction and reportedly filed a complaint. |
| AI/digital capability | Synthetic/staged “dealership employee,” fake invoice, pseudo-news site (`Vérité Cachée`), coordinated X amplification and search gaming |
| Vector/target/mechanism | Fake site/X/Google; French and international audiences; used synthetic testimony and documents to attack Ukraine/aid credibility and French policy. |
| Occurrence and reach | The false story was documented; reporting cites about 18 million X views in 24 hours and high Google ranking. Non-unique and likely substantially non-French. **Occurrence established (0.93); gross reach strong (0.82).** |
| Effect | No evidence of French vote effects; direct French-election intent is not established despite timing. **Open (0.03).** |
| Agency transfer/harm | Synthetic witness and invoice converted false artifacts into apparent firsthand evidence; foreign-policy trust harm. |
| Response | Dealership denial/complaint, fact-checking and media exposure. |

**Sources:** [Wired investigation](https://www.wired.com/story/ai-generated-russian-disinformation-zelensky-bugatti/) — Jul 2024; major investigation, **B/C**; [CNN/WRAL fact check](https://www.wral.com/archive/21508090/) — Jul 2024, **C**; [Reuters Fact Check](https://www.reuters.com/fact-check/no-evidence-ukraines-first-lady-bought-bugatti-sportscar-paris-2024-07-13/) — 13 Jul 2024, **C**. Keep as comparator/reserve unless direct election targeting is independently established.

### 18. `de-2025-bund-fake-ballots` — **review flag: limited reach**

| Field | Coded record |
|---|---|
| Country; election/date | Germany; Bundestag election, 23 Feb 2025 |
| Incident/campaign | Fabricated Leipzig ballot omitting AfD and fake Hamburg video of shredded AfD ballots |
| Actor and attribution status | Artifacts displayed reported Storm-1516/Russian-operation hallmarks; artifact-level command chain was not publicly adjudicated. German authorities warned of Russian disinformation and investigated. |
| AI/digital capability | Staged/fabricated ballot evidence, video, false local context and coordinated dissemination; AI role not conclusively established. |
| Vector/target/mechanism | Social platforms/Telegram; AfD voters and election legitimacy; fake physical “proof” of suppression/fraud. |
| Occurrence and reach | Fabricated artifacts and official debunks are established; reach was limited/uncertain in the reviewed sources. **Occurrence high (0.90); reach unknown (0.35).** |
| Effect | No belief, turnout, vote or tally effect established. **Open (0.03).** |
| Agency transfer/harm | Staged ballot evidence pre-empted institutional verification and promoted a fraud interpretation; trust and post-result legitimacy harm. |
| Response | Rapid election-authority/media debunk and state-security investigation. |

**Sources:** [Tagesschau on Russian disinformation/fake postal ballots](https://www.tagesschau.de/inland/bundestagswahl/desinformation-russland-briefwahl-100.html) — Feb 2025; public-service reporting with official sources, **A/C**; [Reuters, Germany warns of Russian disinformation](https://www.reuters.com/world/europe/germany-warns-russian-disinformation-targeting-election-2025-02-21/) — 21 Feb 2025, **C**. **Second review required:** obtain platform metrics and artifact-specific attribution.

### 19. `hu-2026-parl-fidesz-execution-ad` — **review flag: official campaign content, reach unresolved**

| Field | Coded record |
|---|---|
| Country; election/date | Hungary; parliamentary election, 12 Apr 2026; video reported 19 Feb |
| Incident/campaign | Fidesz generative-AI execution-scene advertisement implying a Tisza win would bring war |
| Actor and attribution status | Official Fidesz campaign content; actor admitted/established. |
| AI/digital capability | Google/AI-generated photorealistic execution scene and synthetic political forecasting |
| Vector/target/mechanism | Campaign/social video; Hungarian voters and Tisza/Péter Magyar; vivid synthetic atrocity scenario used to convert an opponent victory into an emotionally immediate prediction of war. |
| Occurrence and reach | Reuters documented the acknowledged campaign video; public reach metrics are unresolved in the reviewed source. **Occurrence established (0.98); reach unknown (0.30).** |
| Effect | No ad-level causal estimate. A broader survey finding that 23% believed a related war narrative, with large partisan differences, cannot be attributed to this artifact. Fidesz ultimately lost, which does not prove zero attitudinal effect. **Open (0.05).** |
| Agency transfer/harm | Synthetic future imagery bypassed deliberation with fear and purported visual evidence; harm includes intimidation, opponent demonization and blurred reality/simulation cues. |
| Response | Opposition condemnation, press scrutiny and ODIHR observation of AI/manipulative campaigning; no cited ban. |

**Sources:** [Reuters, opposition condemns Fidesz AI execution video](https://www.reuters.com/world/hungary-opposition-condemns-fidesz-election-video-with-fictitious-execution-scene-2026-02-19/) — 19 Feb 2026; major reporting, **C**; [ODIHR Hungary 2026 preliminary statement](https://odihr.osce.org/sites/default/files/documents/official_documents/2026/04/HUN%20Parliamentary%20Elections%202026%20PS_13.04.2026.pdf) — 13 Apr 2026, official observation, **A**. **Second review required:** platform reach, targeting spend and ad-library record.

## Event and institutional-outcome index

| Date | Event | Case(s) | Evidentiary meaning |
|---|---|---|---|
| 5–7 May 2017 | MacronLeaks released during blackout; Macron wins runoff | `fr-2017-pres-macronleaks` | Occurrence/timing and election result; not a causal effect estimate |
| 28–30 Sep 2023 | Šimečka/Tódová audio released and debunked; election held | `sk-2023-parl-simecka-todova-audio` | Artifact and correction established; result not attributable |
| 9 Jun–7 Jul 2024 | EU, UK and French legislative elections | France/UK 2024 records | Temporal scope only |
| 20 Oct–3 Nov 2024 | Moldova referendum/presidential rounds | Moldova 2024 records | Vote-buying/CIB exposure window; validation does not negate occurrence |
| 24 Nov–6 Dec 2024 | Romania first round, declassification and annulment | `ro-2024-pres-georgescu-tiktok` | Court/legal outcome; no quantified vote shift |
| 23 Feb 2025 | German Bundestag election | German records | No tally effect demonstrated |
| 4–18 May 2025 | Romania repeat election | `ro-2025-pres-crossplatform-personas` | Network pivot/delegitimation period; result upheld |
| 28 Sep–16 Oct 2025 | Moldova parliamentary vote and validation | Moldova 2025 records | Occurrence can coexist with court finding of no proven result-changing violation |
| 12–13 Apr 2026 | Hungary election and ODIHR preliminary statement | Hungary records | Fidesz loss does not quantify campaign effects |

## Excluded or non-standalone weak cases

1. **Romania “85,000 cyberattacks.”** Technical disruption belongs in a cyber-infrastructure ledger unless linked to voter-agency transfer; no tally effect shown.
2. **Romania “1.4 billion TikTok views.”** Aggregate hashtag/content volume is not unique exposure and cannot be assigned to a single operation.
3. **Romania Georgescu polling surge as effect.** Chronology is not causal evidence.
4. **Turkey-based TikTok network merged with SRI network.** Scopes, account populations and sponsors differ; no summing.
5. **Moldova “150,000 bought votes,” over-300,000 projection, or €200m/1%-GDP claims.** These are political/authority estimates or projections, not observed compliant votes or fully accounted expenditure. Prefer identified PSB transfers and label estimates.
6. **Election-wide platform enforcement mapped to named networks.** TikTok's 7,593 network accounts/134,000 fake-account removals and G7's 450+ FIMI items cannot be wholesale assigned to InfoLeaders, Telegram bots or Storm-1516.
7. **Cloudflare 898 million malicious requests/DDoS.** Infrastructure attack reportedly mitigated; separate cyber record, not content influence or voter-agency transfer.
8. **Isolated deepfake anecdotes without preserved artifact, provenance or reach.** Exclude until auditable evidence exists.
9. **Court validation or annulment as behavioral effect.** Courts establish legal/institutional outcomes, not individual persuasion or a numeric vote shift.
10. **UK Keir Starmer Oct 2023 audio.** Authenticity/source remained inconclusive and it fell outside the 2024 campaign window.
11. **Reform candidate Mark Matlock “AI candidate” claim.** The person was real; this is a liar's-dividend/misidentification episode, not an election-manipulation operation.
12. **Slovakia “beer-price” audio as separate row.** Lower/unclear reach and same campaign context; treat as a supporting artifact, not a distinct incident.
13. **Hungary Péter Magyar pension video.** Alleged deepfake and criminal complaint without sufficient public forensic establishment; keep excluded pending evidence.
14. **Hungary voting-advice chatbot audit.** Useful systemic-risk comparator, but no demonstrated real voter exposure or intentional manipulation campaign.
15. **France Zelenska/Bugatti as a core election row.** Preserved above only as a flagged comparator because timing is strong but direct French-election targeting is unproven.
16. **Broad Portal Kombat, Doppelgänger or Pravda ecosystem totals.** Avoid vague standalone rows without artifact-level linkage, date, target and reach.

## Integration status for v0.1.0

The Romania/Moldova checklist below was applied to `research/seed.json` before release and is retained as a historical integration record, not as an outstanding task list:

- Romania's promotion, platform-network, domestic-financing, institutional, and cyber components are separated at claim/source level; incompatible account and view totals are not summed.
- Generative AI is not coded as a material driver of the 2024 Romanian annulment; the confirmed 2025 synthetic-persona network remains a distinct linked case.
- Polling chronology, court action, behavioural effect, and causal electoral effect remain separate.
- Mixed Romanian attribution and the PNL-financed brief are preserved as counterevidence; the EC action remains an opened proceeding, not a final liability finding.
- Moldova's vote-buying, CIB, synthetic-persona, and 2025 InfoLeaders evidence are separated; secret-ballot compliance and causal vote effect remain unknown.
- Gross platform metrics are not treated as authenticated in-country reach, and aggregate enforcement totals are not assigned wholesale to named networks.
- Cyber-infrastructure totals without a demonstrated voter-agency pathway are excluded from the core content-influence rows.
- The Moldova 2025 ODIHR source was resolved to its canonical landing page and verified publication date before ingestion.

Post-release work remains limited to permitted archival capture of ephemeral artifacts and the explicitly flagged second-source/platform-metric pass for comparator records that remain candidates rather than core cases.