# Americas secondary case candidates

**Cut-off:** 2026-08-11. These 12 records complement the three priority ledgers in `americas_priority_ledger.md`, producing a 15-case Americas candidate set. They are ranked for comparative value, not asserted to be equivalent in severity.

Evidence key used below: `O` occurrence; `R` reach/exposure; `B` behavioural effect; `E` electoral effect; `A` attribution. A reach metric never substitutes for `B` or `E`.

---

## 4. `usa-2016-ira-social-influence-operation`

- **Country/election/date:** United States; 2016 presidential election; activity principally 2015–2017.
- **Incident/actor:** Internet Research Agency social-media influence operation. The Senate Intelligence Committee and Special Counsel officially attributed the operation to the Kremlin-backed IRA; criminal charges against Russian defendants should not be mislabeled as convictions.
- **Capability/vector/target:** Coordinated inauthentic personas and pages, paid ads, interest/location targeting, organic sharing/recommendation, cross-platform posting, and mobilization of unwitting Americans into petitions, classes, rallies, and campaign-material acquisition. Platforms included Facebook, Instagram, Twitter/X, YouTube, and others. African Americans were the most intensively targeted demographic in the Senate's analysis; other polarizing identity/issue audiences were also targeted. Generative AI is `not_applicable`.
- **Mechanism / agency transfer:** Fake US identities concealed Russian control; targeting and ranking routed divisive material to selected communities; real Americans were induced to provide data, organize events, and recirculate content. Control over apparent authorship, agenda, and mobilization shifted to a foreign operation.
- **Scale:** Facebook reported more than 3,000 ads, about 80,000 Facebook posts from roughly 120 pages, 11.4 million US users seeing at least one IRA ad, and an estimated 126 million people served some story originating from an IRA page between 2015 and 2017. Only 44% of IRA ad impressions were pre-election; the 126 million estimate covers a two-year period and must not be represented as 126 million persuaded voters.
- **Effect evidence:** `B unknown; E unknown.` Official investigations established intent and extensive activity, not a causal vote or turnout effect.
- **Harm/response:** Covert foreign agenda-setting, racial/identity manipulation, voter disengagement efforts, and trust erosion. Responses included account/page removal, congressional investigation, DOJ indictment, sanctions, exposure notices, and political-ad/source transparency changes.
- **Evidence classification/confidence:** `O confirmed; R platform_reported; A officially_attributed; B unknown; E unknown.` Confidence 0.99 occurrence/attribution, 0.97 Facebook metrics as defined, 0.15 electoral effect.
- **Sources:**
  - https://www.intelligence.senate.gov/2019/10/08/press-senate-intel-committee-releases-bipartisan-report-russia-e2-80-99s-use-social-media/
  - https://www.intelligence.senate.gov/2018/08/30/hearings-open-hearing-foreign-influence-operations-e2-80-99-use-social-media-platforms-company-witnesses/
  - https://www.justice.gov/archives/sco/file/1373816/dl?inline=
  - https://www.justice.gov/d9/fieldable-panel-panes/basic-panes/attachments/2018/02/16/internet_research_agency_indictment.pdf

## 5. `usa-2016-cambridge-analytica-voter-profiling`

- **Country/election/date:** United States; 2016 presidential election; data acquisition mainly 2014–2015 and campaign deployment in 2016.
- **Incident/actor:** Cambridge Analytica/SCL and Aleksandr Kogan's GSRApp data-harvesting and voter-profiling system; campaign work included Trump-aligned entities. The FTC issued a final opinion against Cambridge Analytica after the company failed to answer; Nix and Kogan settled separately. Preserve the procedural posture.
- **Capability/vector/target:** Facebook app/friend data, identity linkage, OCEAN personality modelling, voter files, profiling, audience segmentation, and targeted digital advertising. This is predictive/personalized data infrastructure, not generative AI.
- **Mechanism / agency transfer:** Deceptive consent and friend-data extraction created profiles voters could not inspect, contest, or meaningfully opt out of; the campaign vendor gained asymmetrical control over inference and message selection.
- **Scale:** FTC findings: roughly 250,000–270,000 US app users directly interacted with the app; profiles were collected from 50–65 million affected friends, including at least 30 million identifiable US consumers. These are data subjects, not proven ad recipients or persuaded voters.
- **Effect evidence:** `B unknown; E unknown.` Cambridge Analytica's own performance claims are interested-party assertions and not a causal evaluation. The Trump victory cannot be attributed to the system on this evidence.
- **Harm/response:** Privacy/consent violation, opaque profiling, discriminatory or manipulative personalization risk, and unequal informational power. FTC final order required deletion/prohibited misrepresentations; related settlements required destruction of data and derived algorithms; Facebook later faced a separate FTC settlement.
- **Evidence classification/confidence:** `O adjudicated by FTC; R/data-subject count established; campaign use credibly reported; B/E unknown.` Confidence 0.98 harvesting/profiling, 0.90 campaign relationship, 0.20 effect.
- **Counterevidence:** The FTC record proves collection/profiling, not that “psychographic” segmentation outperformed ordinary political targeting or changed the election.
- **Sources:**
  - https://www.ftc.gov/system/files/documents/cases/d09389_comm_final_opinionpublic.pdf
  - https://www.ftc.gov/news-events/news/press-releases/2019/12/ftc-issues-opinion-order-against-cambridge-analytica-deceiving-consumers-about-collection-facebook
  - https://campaignlegal.org/update/newly-published-cambridge-analytica-documents-show-unlawful-support-trump-2016
  - https://now.tufts.edu/2018/05/17/did-cambridge-analytica-sway-election

## 6. `usa-2020-iran-proud-boys-voter-threat-emails`

- **Country/election/date:** United States; 2020 presidential election; August–November 2020.
- **Incident/actor:** Spoofed “Proud Boys” voter-intimidation emails and fabricated ballot-fraud video. The US Intelligence Community attributed the campaign to Iran with high confidence and assessed Supreme Leader authorization; DOJ charged two Iran-based contractors linked to Emennet Pasargad. The named defendants' guilt remains unadjudicated.
- **Capability/vector/target:** Reconnaissance against about 11 state voter sites; exploitation of one misconfigured system; theft of more than 100,000 voter records; personalized threatening email to tens of thousands of registered Democratic voters; spoofed domestic-extremist identity; simulated-intrusion video; attempted media-company compromise.
- **Mechanism / agency transfer:** Stolen public/official voter data made threats appear individualized and credible; a false domestic identity shifted perceived authorship and tried to make voters change affiliation/vote while also laundering a false ballot-fraud narrative.
- **Scale:** More than 100,000 voter records downloaded; threat emails sent to tens of thousands. Delivery does not establish reading, belief, or compliance.
- **Effect evidence:** `B unknown; E unknown.` ODNI explicitly did not assess public-opinion/election impact and reported no indication any foreign actor altered voter registration, ballot casting, tabulation, or reporting.
- **Harm/response:** Targeted intimidation, privacy abuse, domestic-group impersonation, fraud narrative, and trust erosion. FBI/private-sector disruption and victim notification blocked the post-election media intrusion; public attribution, indictment, sanctions, and reward followed.
- **Evidence classification/confidence:** `O confirmed in official technical/IC records; A officially_attributed (state), named-individual guilt alleged; R measured for records and qualitative for delivery; B/E unknown.` Confidence 0.98 occurrence/state attribution, 0.85 named operational account, 0.10 electoral effect.
- **Sources:**
  - https://www.justice.gov/archives/opa/pr/two-iranian-nationals-charged-cyber-enabled-disinformation-and-threat-campaign-designed
  - https://archive.dni.gov/files/ODNI/documents/assessments/ICA-declass-16MAR21.pdf
  - https://www.osi.af.mil/News/Article-Display/Article/2395806/bulletin-iranian-spoofed-emails-portraying-the-proud-boys/

## 7. `usa-2024-russia-doppelganger-us-targeting`

- **Country/election/date:** United States; 2024 presidential election; exposed/disrupted 2024-09-04 after multiyear activity.
- **Incident/actor:** “Doppelganger” cybersquatted-news and tailored-audience influence operation. DOJ alleged Social Design Agency, Structura, and ANO Dialog operated under the Russian Presidential Administration, particularly Sergei Kiriyenko; Treasury also designated actors. Attribute at `officially_attributed`, while retaining that the seizure affidavit contains allegations rather than a criminal merits judgment.
- **Capability/vector/target:** 32 seized domains; nearly cloned Washington Post/Fox sites; invented niche brands; fabricated influencers/fake US profiles; paid social ads; AI-assisted ads/images; specific messaging for six swing states and audiences including LGBTQ+ people, Texas agricultural workers, and Jewish Americans.
- **Mechanism / agency transfer:** Legitimacy laundering through copied news design/bylines, fake local/identity media, and tailored distribution made Russian narratives appear to originate from trusted domestic sources.
- **Scale:** 32 domains seized. DFRLab found some PrideWave posts over 500,000 views, but this is a content-level/platform trace, not operation-wide unique reach. DOJ published no defensible total exposure.
- **Effect evidence:** `B unknown; E unknown.` Targeting plans and views demonstrate capability/exposure, not persuasion or result change.
- **Harm/response:** Foreign source concealment, press impersonation, demographic manipulation, polarization, and Ukraine-policy/election narrative interference. DOJ seized domains; Treasury imposed designations; platforms removed related assets.
- **Evidence classification/confidence:** `O partly confirmed through seized artifacts; A officially attributed/alleged in affidavit; R mixed/partial; B/E unknown.` Confidence 0.96 artifacts/operation, 0.92 state direction as officially alleged, 0.15 effect.
- **Sources:**
  - https://www.justice.gov/archives/opa/pr/justice-department-disrupts-covert-russian-government-sponsored-foreign-malign-influence
  - https://www.justice.gov/usao-edpa/pr/justice-department-disrupts-covert-russian-government-sponsored-foreign-malign
  - https://dfrlab.org/2024/09/18/doppelganger-us-election/
  - https://about.fb.com/news/2024/12/2024-global-elections-meta-platforms/

## 8. `usa-2024-iran-trump-campaign-hack-and-leak`

- **Country/election/date:** United States; 2024 presidential election; compromise from May 2024; leak attempts June–August; indictment 2024-09-27.
- **Incident/actor:** Iranian hack-and-leak operation against the Trump campaign. FBI/ODNI/CISA officially attributed the compromise and leak attempts to Iran; DOJ alleged three named IRGC employees. The state attribution is stronger than the untried individual charges.
- **Capability/vector/target:** Spearphishing/social engineering, spoofed logins, VPN/VPS infrastructure, account compromise, exfiltration of nonpublic campaign material, then timed outreach to Biden-campaign associates and journalists.
- **Mechanism / agency transfer:** Illicit access gave a foreign actor control over selection, timing, framing, and potential domestic amplification of private campaign documents; attempted to recruit trusted media/opposition recipients as distribution intermediaries.
- **Scale:** Emails containing stolen excerpts went to three believed Biden-campaign-associated recipients; none replied. Stolen vice-presidential vetting materials were offered to multiple news outlets July 22–August 31. Recipient counts are not mass reach, and major nonpublication/nonresponse is counterevidence to a successful leak-amplification claim.
- **Effect evidence:** `B unknown/not detected in recipients; E unknown.` No evidence shows the effort changed votes or outcome.
- **Harm/response:** Foreign intrusion into campaign autonomy, privacy and security, agenda manipulation, and potential media laundering. Responses: joint attribution, platform cooperation, indictment, sanctions, and up-to-$10m reward.
- **Evidence classification/confidence:** `O/A officially attributed; individual guilt alleged; R low/qualitative; B/E unknown.` Confidence 0.98 compromise/state attribution, 0.90 named-actor allegation as allegation, 0.10 effect.
- **Sources:**
  - https://www.justice.gov/archives/opa/pr/three-irgc-cyber-actors-indicted-hack-and-leak-operation-designed-influence-2024-us
  - https://www.fbi.gov/news/press-releases/joint-odni-fbi-and-cisa-statement-on-iranian-election-influence-efforts
  - https://www.fbi.gov/news/press-releases/iranian-cyber-actors-targeting-personal-accounts-to-support-operations

## 9. `bra-2018-whatsapp-mass-messaging-ecosystem`

- **Country/election/date:** Brazil; 2018 presidential election, especially the October runoff.
- **Incident/actor:** Alleged business-financed anti-PT WhatsApp bulk-messaging combined with an established mass-messaging ecosystem. Folha reported contracts up to R$12m and “hundreds of millions” of sends; defendants denied key allegations. In 2021 the TSE unanimously rejected AIJEs 0601968-80 and 0601771-28 because crucial content, reach, repercussion, candidate link, and financing/gravity were not proved.
- **Capability/vector/target:** WhatsApp bulk sends, third-party databases, geographic/income segmentation, multiple foreign numbers, group administrators, supporter/influencer cascades, and alleged automation. GenAI is `not_applicable`.
- **Mechanism / agency transfer:** Closed-group/peer-channel provenance, list brokerage, and bulk delivery could make paid/centrally shaped messages appear interpersonal and organic while limiting public scrutiny and correction.
- **Scale:** Folha's R$12m/“hundreds of millions” claims are journalistic allegations, not adjudicated metrics. WhatsApp found only 3 lines among more than 600 linked accounts exhibiting abnormal automated/spam behavior and banned them. The TSE found evidence sufficient to recognize mass messaging occurred in some form but could not establish message content, reach, or electorate repercussion.
- **Effect evidence:** `B unknown; E unknown.` The TSE expressly rejected inference without minimum proof.
- **Harm/response:** Opaque financing/provenance, privacy/data-list misuse, spam, peer-trust laundering, and unequal amplification. Responses included WhatsApp bans/forwarding restrictions, investigations, and a TSE doctrine that mass messaging with disinformation can constitute abuse if gravity is proved.
- **Evidence classification/confidence:** `O partly confirmed; financing/campaign command alleged; R alleged/unknown; A mixed; B/E unknown.` Confidence 0.90 mass-messaging ecosystem, 0.45 alleged corporate scale/financing, 0.10 effect.
- **Sources:**
  - https://www1.folha.uol.com.br/poder/2018/10/empresarios-bancam-campanha-contra-o-pt-pelo-whatsapp.shtml
  - https://sjur-servicos.tse.jus.br/sjur-servicos/rest/download/pdf/2673321
  - https://policyreview.info/articles/analysis/whatsapp-and-political-instability-brazil-targeted-messages-and-political
  - https://www.mozillafoundation.org/en/research/library/global-elections-casebook/brazil-case-study/

## 10. `bra-2022-bolsonaro-ambassadors-election-system-attack`

- **Country/election/date:** Brazil; 2022 general election; event on 2022-07-18.
- **Incident/actor:** Bolsonaro's ambassadors meeting attacking electronic voting. Direct personal/state-resource attribution was adjudicated: in 2023 the TSE voted 5–2 to declare Bolsonaro ineligible for eight years for abuse of political power and misuse of media; Braga Netto was not sanctioned because responsibility was not shown.
- **Capability/vector/target:** Presidential ceremony, Palácio da Alvorada, public resources, diplomats, TV Brasil livestream, and social-media redistribution. This is elite/platform amplification and state-authority appropriation; no generative AI.
- **Mechanism / agency transfer:** An incumbent used presidential office, diplomatic staging, and public broadcasting to give campaign election-fraud claims official-seeming authority and seed them into social networks.
- **Scale:** TSE described significant repercussion and an expressive number of voters reached but the cited decision summary supplies no auditable unique-viewer count. Code reach `qualitative/institutionally asserted`.
- **Effect evidence:** Institutional/legal harm was adjudicated; no causal estimate of votes, turnout, or result exists. A judicial finding of grave abuse is not a quantified electoral effect.
- **Harm/response:** Delegitimization of electoral administration, misuse of state resources/media, unequal incumbent power, and reduced trust. TSE fact-checking/removal responses preceded the ineligibility judgment.
- **Evidence classification/confidence:** `O/A/legal outcome adjudicated; R qualitative; B unknown; E institutionally asserted as grave risk, not causally measured.` Confidence 0.99 occurrence/actor/judgment, 0.85 institutional harm, 0.20 vote effect.
- **Sources:**
  - https://www.tse.jus.br/comunicacao/noticias/2023/Junho/por-maioria-de-votos-tse-declara-bolsonaro-inelegivel-por-8-anos
  - https://www.tse.jus.br/comunicacao/noticias/2022/Outubro/fato-ou-boato-justica-eleitoral-desmentiu-as-principais-fake-news-sobre-o-processo-eleitoral-em-2022
  - https://international.tse.jus.br/en/misinformation-and-fake-news/tse-brazil-counter-disinformation-program-2022.pdf/%40%40download/file/TSE%20BRAZIL%20-%20Counter%20Disinformation%20Program%202022.pdf

## 11. `bra-2022-election-denial-messaging-network`

- **Country/election/date:** Brazil; 2022 general election and post-election period through the 2023-01-08 attacks.
- **Incident/actor:** Cross-platform election-denial and anti-democratic messaging ecosystem. Actor attribution is `mixed/diffuse`: pro-Bolsonaro elite cues, influencers, hyperpartisan sites, and public WhatsApp/Telegram groups participated, but the cited research does not establish a single command actor for the entire network.
- **Capability/vector/target:** Coordinated/synchronous link sharing, bulk posting, public WhatsApp groups, Telegram channels, Twitter/X, YouTube, Facebook, and hyperpartisan websites; recycling between platforms and permanent group infrastructure. GenAI was not material to the documented mechanism.
- **Mechanism / agency transfer:** Repetition across trusted affinity groups and platforms concentrated agenda-setting and electoral-trust cues in network brokers; closed/semi-closed groups raised correction costs and enabled mobilization.
- **Scale:** NetLab/Mozilla analyzed 267,209 messages in 390 public WhatsApp and 864 Telegram groups/channels from 2021-01-01 to 2022-10-30; the electoral-fraud keyword query yielded 99,894 WhatsApp and 167,315 Telegram messages. Aos Fatos detected over 991,000 relevant tweets June 2022–January 2023, a 1,278% year-on-year increase, plus lower increases elsewhere. These are messages/posts in sampled spaces, not unique people.
- **Effect evidence:** A two-wave survey found joining political WhatsApp groups was the strongest predictor of post-election misinformation belief, but the authors expressly declined causal claims. Linking the network to Jan. 8 mobilization is plausible/contextual, not a message-level causal estimate.
- **Harm/response:** Persistent electoral delegitimization, coup advocacy, institutional distrust, and mobilization risk. Responses included TSE counter-disinformation operations, platform restrictions, fact-checking, removals, and judicial/law-enforcement action.
- **Evidence classification/confidence:** `O confirmed in sampled traces; R measured as content volume; A mixed; B indicated/correlational; E unknown.` Confidence 0.98 trace findings, 0.75 coordination inference, 0.55 belief association, 0.20 offline causal effect.
- **Sources:**
  - https://www.mozillafoundation.org/en/research/library/global-elections-casebook/brazil-case-study/
  - https://www.aosfatos.org/noticias/anti-democratic-content-surged-on-social-media-before-brazils-2022-election-and-its-not-slowing-down/
  - https://misinforeview.hks.harvard.edu/article/explaining-beliefs-in-electoral-misinformation-in-the-2022-brazilian-election-the-role-of-ideology-political-trust-social-media-and-messaging-apps/
  - https://international.tse.jus.br/en/misinformation-and-fake-news/tse-brazil-counter-disinformation-program-2022.pdf/%40%40download/file/TSE%20BRAZIL%20-%20Counter%20Disinformation%20Program%202022.pdf

## 12. `mex-2024-genai-election-manipulation-cluster`

- **Country/election/date:** Mexico; 2024 general election; research window 2023-11-20 to 2024-06-17.
- **Incident/actor:** Cluster of fact-checked GenAI/AI-manipulated election content targeting Claudia Sheinbaum, Xóchitl Gálvez, Andrés Manuel López Obrador, and others. Most creators/sponsors are unknown; do not imply a centrally coordinated campaign.
- **Capability/vector/target:** Voice cloning, lip-sync, video/audio/image alteration, added crowd boos/backgrounds/text, fake celebrity/company endorsements, and politician deepfakes used in both election narratives and financial scams; circulated across X, Facebook, WhatsApp, TikTok, Instagram, and others.
- **Mechanism / agency transfer:** Synthetic first-person speech and false social proof/endorsements let unknown producers control a public figure's apparent words, affiliations, popularity, and credibility.
- **Scale:** UT Center for Media Engagement collected 1,760 fact-checker posts and reduced them to 101 unique posts about AI/election content. **This is 101 fact-check posts, not necessarily 101 unique incidents or an electorate-wide prevalence estimate.** Categories: 63.3% false attributed statements, 22.9% false associations/endorsements, 7.33% political violence/presentation distortion, 6.4% scams.
- **Effect evidence:** Report authors describe exact electoral impact as uncertain and limited; no unique reach, belief change, turnout, vote-choice, or result change is measured.
- **Harm/response:** Candidate impersonation, false endorsements, gendered/political presentation attacks, scams, and contextual collapse of satire. Responses centered on fact-checking networks (including Certeza/INE sources) and candidate/authority corrections.
- **Evidence classification/confidence:** `O confirmed for documented examples; R unknown; A mostly unknown; B/E unknown/limited contextual assessment.` Confidence 0.91 cluster occurrence, 0.55 prevalence, 0.15 electoral effect.
- **Sources:**
  - https://mediaengagement.org/wp-content/uploads/2024/10/CME-Generative-AI-Misinformation-in-the-2024-Mexican-Elections.pdf
  - https://www.freiheit.org/sites/default/files/2025-01/ia-y-elecciones_26.11.2024.pdf
  - https://www.idea.int/publications/catalogue/html/artificial-intelligence-and-information-integrity-latin-american

## 13. `col-2022-coordinated-facebook-negative-ad-pages`

- **Country/election/date:** Colombia; 2022 presidential election; investigated January–May 2022.
- **Incident/actor:** Ten coordinated Facebook pages posing as patriotic/neutral media and buying negative/misleading ads principally against Gustavo Petro and Sergio Fajardo. Digital traces led to a contractor/strategist network, but CLIP/La Silla Vacía could not establish the ultimate operator or funder.
- **Capability/vector/target:** Coordinated page creation, identical copy/domains/contact data, paid Facebook targeting, short negative videos, and opaque sponsorship. No material GenAI established.
- **Mechanism / agency transfer:** Seemingly independent “love of Colombia” pages concealed common coordination and funding, allowing an unknown sponsor to control paid persuasion without normal campaign-source cues.
- **Scale:** COP309 million (about US$75,000) in Facebook ad spend; 538 paid items; 10 pages; 8 pages removed after journalists contacted the suspected operator. Spend/item count is not unique reach or persuasion.
- **Effect evidence:** `B unknown; E unknown.` Carter Center separately reported fraud conversation grew 25-fold versus 2018 and summarized 138 Colombiacheck verifications, but those election-wide facts are context, not effects of these ten pages.
- **Harm/response:** Sponsor opacity, coordinated negative persuasion, misleading content, and evasion of campaign accountability. Journalistic exposure was followed by page deletion; Carter Center recommended stronger CNE capacity/platform collaboration.
- **Evidence classification/confidence:** `O/spend confirmed through ad-library investigation; A ultimate funder unknown; R unknown; B/E unknown.` Confidence 0.96 coordination/spend, 0.40 ultimate attribution, 0.15 effect.
- **Sources:**
  - https://www.elclip.org/campana-sucia-en-facebook-colombia-elecciones-desinformacion/
  - https://cartercentee50c07c05.blob.core.windows.net/blobcartercentee50c07c05/wp-content/uploads/2022/07/colombia-expert-mission-report-2022-english.pdf
  - https://freedomhouse.org/country/colombia/freedom-net/2022

## 14. `arg-2023-official-campaign-genai-arms-race`

- **Country/election/date:** Argentina; 2023 presidential election/runoff, 2023-11-19.
- **Incident/actor:** Both Sergio Massa and Javier Milei campaign ecosystems used GenAI images/video for self-promotion and attack; Massa-linked creators used the unofficial “AI for the Homeland” account and Milei published synthetic anti-Massa imagery.
- **Capability/vector/target:** Midjourney-style synthetic images, stylized heroic/demonic archetypes, synthetic attack videos, Instagram/TikTok/X distribution, and at least one more deceptive depiction of Milei speaking about organ markets.
- **Mechanism / agency transfer:** Campaigns rapidly generated emotionally loaded scenes and, in the more deceptive examples, took control over an opponent's apparent speech/embodiment.
- **Scale:** A Milei-posted apparently AI-generated image portraying Massa as an old-style communist received about 3 million views. That single-item view count is not unique exposure to the wider cluster.
- **Effect evidence:** Freedom House and comparative scholarship assess GenAI's information-space impact as relatively minor: much content was labelled or visibly fabricated and had noticeable flaws. `B/E unknown`; Milei's 56% runoff victory is not evidence the synthetic posts caused the result.
- **Harm/response:** Demonization, synthetic candidate appropriation, and normalization of generated campaign imagery; mitigated by overt satire/artifice in much of the cluster. Reverso and other fact-checking collaborations responded.
- **Evidence classification/confidence:** `O/actor high; R measured for one item; deceptive intent mixed; B/E unknown with counterevidence of limited impact.` Confidence 0.95 use, 0.88 item views, 0.25 material manipulation/effect.
- **Sources:**
  - https://www.context.news/ai/how-ai-shaped-mileis-path-to-argentina-presidency
  - https://freedomhouse.org/country/argentina/freedom-net/2024
  - https://knightcolumbia.org/content/dont-panic-yet-assessing-the-evidence-and-discourse-around-generative-ai-and-elections
  - https://www.cambridge.org/core/journals/european-political-science/article/on-the-way-to-deep-fake-democracy-deep-fakes-in-election-campaigns-in-2023/8F97B6AD4C40B195B369696926B5F7EB

## 15. `arg-2025-buenos-aires-city-false-withdrawal-deepfakes`

- **Country/election/date:** Argentina; Buenos Aires City legislative election; 2025-05-18.
- **Incident/actor:** Undisclosed AI videos falsely showed Mauricio Macri and Silvia Lospennato withdrawing Lospennato's candidacy to support Manuel Adorni during the May blackout. Originators remained unknown in the cited record; do not attribute them to Milei or his party solely because pro-Milei accounts amplified them.
- **Capability/vector/target:** Synthetic candidate/ex-president video and voice, timed release immediately before voting, X/social amplification, false withdrawal/endorsement narratives.
- **Mechanism / agency transfer:** Attackers appropriated authoritative first-person candidate/party speech at a moment when rapid rebuttal was difficult, attempting to alter beliefs about ballot viability and elite endorsement.
- **Scale:** The complaint identified at least 9 X accounts; all complained-of videos remained online when Chequeado checked despite a judicial removal order. Account counts, complaints, and continued availability are not unique viewers or effect.
- **Effect evidence:** Experts said exposure, belief, and vote effect could not be measured. `B/E unknown.`
- **Harm/response:** Last-minute false withdrawal can directly impair informed choice; platform noncompliance or latency weakens remedy. Candidates rebutted, PRO complained, and the city electoral tribunal ordered removal.
- **Evidence classification/confidence:** `O/AI confirmed for named artifacts; A unknown; R unknown; response established; B/E unknown.` Confidence 0.95 occurrence, 0.30 attribution, 0.15 effect. Keep in reserve until a second independent source is added.
- **Sources:**
  - https://chequeado.com/el-explicador/videos-falsos-de-macri-y-lospennato-en-x-que-paso-en-la-realidad-con-lo-que-ordeno-la-justicia/

The October 2025 national-election withdrawal and endorsement incidents are a separate future candidate. They are not merged into this May record because the artifacts, election, dates, complaint denominator, and response differ.

---

# Excluded or downgraded leads

| Lead | Disposition | Reason |
|---|---|---|
| US 2024 Tenet/RT creator-financing case | Keep as election-adjacent reserve, not in main 15 | DOJ alleged about $9.7m, nearly 2,000 videos, and 16m YouTube views, but many videos concerned broad US issues rather than a bounded election incident; domestic commentators may have been unwitting; no effect evidence. Source: https://www.justice.gov/archives/opa/pr/two-rt-employees-indicted-covertly-funding-and-directing-us-company-published-thousands |
| US 2024 fabricated Pennsylvania/Georgia voting videos | Reserve | Strong official Russian attribution, but generative-AI status is not established for every video; code as fabricated/manipulated, not automatically AI. No effect evidence. Sources: https://www.fbi.gov/news/press-releases/joint-odni-fbi-and-cisa-statement and https://www.fbi.gov/news/press-releases/joint-odni-fbi-and-cisa-statement-on-russian-election-influence-efforts |
| Brazil 2024 municipal-election deepfake cluster | Strong reserve / aggregate precursor | DFRLab counted 78 **confirmed or alleged** cases and Aos Fatos counted court actions; those categories mix confirmed, alleged, satire, and petitions. Useful ecosystem record, but not a clean incident unless disaggregated. Source: https://dfrlab.org/2024/11/26/brazil-election-ai-deepfakes/ |
| Brazil 2026 generic predictions of an “AI election” | Exclude from incident table; watchlist only | Prospective risk is not an observed event. Only the July 25 avatar is currently a bounded observed anchor in this file. |
| Colombia 2026 claim of a 500,000-bot “BlackCore” attack | Exclude pending corroboration | Based primarily on political allegation without adequate platform/forensic corroboration; high risk of encoding campaign rhetoric as fact. |
| Colombia 2022 Wappid organizing app | Exclude | Large-scale digital campaign organizing is not manipulation by itself; no adequate evidence of covert automation/deception or agency-transfer harm. |
| Argentina 2023 most overtly stylized AI posters | Boundary comparison only | Transparently artificial/obvious satire and limited impact weaken the manipulation claim; retained only as a comparison because the broader campaign cluster included more deceptive artifacts. |
| Venezuela 2024 | Exclude from AI-mechanism subset | The central integrity crisis concerns result transparency, coercion/censorship, and institutional control; no equally strong bounded AI/synthetic-media mechanism is needed to explain the harm. Include only in a broader digital-authoritarianism dataset. |

## Selection note

For a smaller release focused tightly on **agency transfer**, keep the three anchors plus cases 4, 5, 6, 7, 9, 10, 11, 13, and 15. Treat Mexico 2024 and Argentina 2023 as aggregate/boundary comparisons, and retain Argentina 2025 over Argentina 2023 if only one Argentine case is allowed.