# Asia–Africa comparative election-case ledger

**Research cut-off:** 2026-08-11
**Scope:** synthetic media, generative AI, automated/inauthentic coordination, recommender amplification and closely related liar's-dividend mechanisms in elections in Asia and Africa.
**Status:** source-checked research ledger for conversion into the relational dataset. It does **not** assert that any item changed votes or an election result unless causal evidence exists.

## Coding discipline

- `established_evidence`: directly supported by a primary/authoritative record or convergent high-quality reporting.
- `strong_inference`: best explanation of convergent evidence, but not directly measured or adjudicated.
- `plausible_hypothesis`: credible mechanism without sufficient observational evidence.
- `open_question`: evidence does not permit a conclusion.
- Source quality follows the project codebook: `A` authoritative primary/peer-reviewed; `B` high-quality independent reporting or transparent civil-society/technical work; `C` interested-party/platform report or incomplete secondary evidence; `D` research lead only.
- Reach is never treated as behavioural or electoral effect. Platform hashtag totals, vendor counts and view counters are preserved with their denominators and caveats.
- “Deepfake” is used only when a source verifies AI generation or clearly uses the term. Spliced or dubbed media without evidence of generative AI is coded `synthetic_media_non_genAI_or_unresolved`.

## Case ledger

| Case ID | Election | Incident / campaign | Mechanism | Controller / attribution | Classification | Inclusion tier | Maximum warranted conclusion |
|---|---|---|---|---|---|---|---|
| `IN-2024-BOLLYWOOD` | India general election, 2024-04-19–2024-06-01 | Aamir Khan and Ranveer Singh impersonation videos supporting Congress | AI voice/video impersonation; social sharing | Creator unknown; a Congress spokesperson amplified one clip | Harmful manipulation; unauthorised impersonation | A | Two false AI-generated celebrity endorsements circulated at material scale; creator, behavioural effect and electoral effect are unknown. |
| `IN-2024-SANCTIONED-CALLS` | India general election, 2024 | Candidate-authorised voice clones, personalised calls and “resurrected” politicians | Voice cloning, personalisation, multilingual outreach | Candidates, parties and commercial vendors | Boundary case: authorised campaign AI, often weakly disclosed to recipients | B | Sanctioned synthetic outreach operated at very large reported scale; authorisation did not necessarily mean recipient awareness. |
| `TW-2024-STORM1376` | Taiwan presidential/legislative election, 2024-01-13 | False Terry Gou endorsement audio plus AI memes/anchors | AI voice clone, synthetic anchors, coordinated IO | Microsoft attributes the operation to CCP-linked `Storm-1376`; state direction not independently adjudicated | Harmful manipulation; foreign-attributed vendor assessment | A | A coordinated actor deployed AI content in an election influence operation; the specific Gou audio was removed before significant reach and no vote effect is established. |
| `PK-2024-PTI-AUTHORISED` | Pakistan general election, 2024-02-08 | Imran Khan speeches generated from prison-approved notes | AI voice clone and talking image | PTI social-media team, with Khan/party authorisation | Transparent/authorised campaign AI; not harmful manipulation on present evidence | A (boundary comparison) | AI substituted for a physically absent candidate/leader under campaign restrictions; the content was party-confirmed and labelled, but causal electoral impact is unknown. |
| `PK-2024-BOYCOTT-CLONES` | Pakistan general election, 2024-02-08 | Fake AI clips said Khan/PTI figures were boycotting | AI voice cloning and synthetic video | Unknown | Harmful manipulation; participation suppression narrative | B | Fabricated boycott claims targeted participation on/near election day; reach, operator and behavioural effect remain unknown. |
| `ID-2024-OFFICIAL-SYNTHETIC` | Indonesia presidential/legislative election, 2024-02-14 | Golkar's synthetic Suharto endorsement and Prabowo's AI “gemoy” rebranding | Deepfake resurrection, generative images, volunteer distribution, chatbots | Golkar and Prabowo-aligned campaign/volunteer structures | Authorised but ethically contested campaign AI; not covert attribution | A/B | Official campaign actors used synthetic identity and large-scale AI branding; view counts show exposure, not persuasion or vote causation. |
| `BD-2024-PROGOV-AI-DISINFO` | Bangladesh parliamentary election, 2024-01-07 | AI anchors and deepfakes attacking the opposition and the United States | Synthetic news anchors, AI video/voice, pseudo-news distribution | Pro-government outlets/influencers identified by FT; government/party command not proven | Harmful manipulation; domestic alignment, attribution unresolved | B | Multiple AI-fabricated narratives were deployed before a noncompetitive election; direct sponsorship, audience effects and electoral effects are unproven. |
| `KR-2024-DEEPFAKE-ENFORCEMENT` | South Korea National Assembly election, 2024-04-10 | Election authority detected content prohibited by new deepfake ban | AI images/videos; monitoring and removal/enforcement | Multiple/unknown domestic creators; NEC as detector | Aggregate harmful/illegal content set; individual deception varies | B | The NEC reported 129 prohibited AI-media items in the law's first weeks; the count is not 129 unique operations and does not establish reach or effect. |
| `ZA-2024-TRUMP-MK` | South Africa general election, 2024-05-29 | Fake Donald Trump endorsement of uMkhonto weSizwe (MK) | AI celebrity voice/video impersonation; amplification by a party-linked influencer | Origin unknown; amplified without disclosure by Duduzile Zuma-Sambudla | Harmful manipulation; unauthorised impersonation | A/B | A prominent MK figure amplified an AI-generated fake endorsement; overall GenAI volume in the election was nevertheless limited. |
| `NG-2023-ATIKU-AUDIO` | Nigeria presidential election, 2023-02-25 | Fabricated audio of Atiku/Okowa/Tambuwal plotting to rig the election | Spliced audio / synthetic media; multi-platform spread | Unknown | Harmful manipulation; **not established as generative AI** | A/B | The audio was assembled from unrelated recordings and remained online after debunking; its creator and effect are unknown. |
| `GH-2024-AI-BOTS` | Ghana general election, 2024-12-07 | 171-account pro-NPP network using likely AI-generated text | Inauthentic accounts, LLM text, hashtag coordination | Unknown operator; partisan direction inferred from content, not proved | Harmful manipulation; covert partisan amplification | B | A research team identified a coordinated-looking AI-assisted account network; most engagement was internal and impact could not be disentangled. |
| `KE-2022-PLATFORM-AMPLIFICATION` | Kenya general election, 2022-08-09 | TikTok hate/disinformation plus paid influencers and bots across platforms | Recommender amplification, manipulated media, bots, coordinated hashtags | Largely domestic actors; ISS found both leading camps used bots | Harmful manipulation; domestic commercialised influence | A/B | Violative content reached millions and coordinated influence was a market; algorithmic contribution is inferred, while voter and result effects remain unmeasured. |

## Agency-transfer mapping

| Case ID | Capability → controller | Influence vector → target | Capacity placed at risk | Power concentrated in | Democratic harm supported by evidence |
|---|---|---|---|---|---|
| `IN-2024-BOLLYWOOD` | Cheap impersonation → unknown creator/amplifiers | Celebrity endorsement videos → social-media users | Authenticity judgment; candidate evaluation | Anonymous creators and high-reach political accounts | False endorsement and identity appropriation; no demonstrated vote change |
| `IN-2024-SANCTIONED-CALLS` | Voice cloning/personalisation → campaigns/vendors | Named calls/messages → individual voters | Awareness that interaction is synthetic; privacy and autonomous evaluation | Campaigns with data and scalable persuasion infrastructure | Opaque substitution of synthetic presence for human contact; harm is contingent, not established for all uses |
| `TW-2024-STORM1376` | Voice/anchor generation + IO infrastructure → attributed coordinated actor | Fake endorsement and pseudo-news → Taiwanese voters | Source/authenticity judgment and trust | Foreign-attributed influence network/platform distribution | Deceptive foreign-attributed messaging; low known reach for focal audio |
| `PK-2024-PTI-AUTHORISED` | Voice cloning → candidate-authorised party team | Labelled virtual speech → supporters | No clear agency loss if authorisation and disclosure are effective | Party communication apparatus | Primarily access restoration; retained as boundary comparison |
| `PK-2024-BOYCOTT-CLONES` | Voice cloning → unknown actor | False withdrawal/boycott message → PTI voters | Participation decision and identity authentication | Anonymous distributor | False participation instruction; effect unknown |
| `ID-2024-OFFICIAL-SYNTHETIC` | Generative art/deepfake resurrection → official campaigns | AI avatars, synthetic deceased leader → young/nostalgic voters | Historical judgment, source awareness and candidate-image evaluation | Well-resourced campaign/volunteer systems | Identity appropriation and image laundering are plausible; causal harm not established |
| `BD-2024-PROGOV-AI-DISINFO` | Low-cost avatar/video tools → aligned outlets/influencers | Pseudo-news and opposition deepfakes → voters/social-media users | Source judgment, geopolitical and candidate beliefs | Pro-government media ecosystem | False narratives and opposition reputational attack; electoral effect unmeasured |
| `KR-2024-DEEPFAKE-ENFORCEMENT` | Deepfake creation → multiple unknown actors | Prohibited campaign media → online public | Authenticity and candidate evaluation | Creators/distributors; countered by NEC | Existence of illegal AI media established; substantive harm not assessable from aggregate count |
| `ZA-2024-TRUMP-MK` | Celebrity voice/video cloning → unknown creator + party-linked amplifier | False foreign endorsement → South African voters | Endorsement authenticity and candidate evaluation | Influential partisan account | Deceptive endorsement; broader election-level GenAI harm limited |
| `NG-2023-ATIKU-AUDIO` | Audio editing → unknown creator | Fake private conversation → voters immediately before poll | Candidate trust and confidence in process | Anonymous creator and platform amplifiers | False rigging narrative persisted through voting period |
| `GH-2024-AI-BOTS` | LLM text + account automation → unknown network operator | Repetitive hashtag posts → X users | Perceived consensus and salience | Covert network operator/platform ranking | Artificial appearance of support/opposition; external reach and effect unclear |
| `KE-2022-PLATFORM-AMPLIFICATION` | Recommendation + bots + paid influence → platform/campaign ecosystems | Hate/disinformation/hashtags → voters and ethnic communities | Attention, perceived consensus, trust, safety | Campaigns, paid influencers and opaque ranking systems | Violent/ethnic rhetoric and false election information amplified at material scale |

## Atomic claims

### `IN-2024-BOLLYWOOD` — unauthorised celebrity endorsements

| Claim ID | Type | Atomic claim | Evidence label / confidence | Sources | Counterevidence / limit |
|---|---|---|---|---|---|
| `INB-01` | occurrence | Two altered videos depicted Aamir Khan and Ranveer Singh criticising Narendra Modi and ending with a Congress symbol and slogan. | established_evidence / high | `S-IN-01` | None material; both actors and independent verification rejected authenticity. |
| `INB-02` | ai_role | Reuters' digital verification and multiple fact-checkers classified the clips as AI-generated/altered; the Ranveer Singh clip used cloned or manipulated audio over authentic footage. | established_evidence / high | `S-IN-01`, `S-IN-03` | Do not generalise the production method to every copy; many election fakes were cheap edits rather than full video generation. |
| `INB-03` | reach | Reuters counted more than 500,000 social-media views across the two videos; one Congress spokesperson's post recorded 438,000 views, 8,700 likes and 2,900 reshares before removal. | established_evidence / high | `S-IN-01` | View counts are exposure proxies, may overlap across copies, and do not measure belief. |
| `INB-04` | attribution | Reuters could not identify the creator. A Congress spokesperson knowingly retained/amplified the Singh video after X labelled it manipulated media. | established_evidence / high | `S-IN-01` | Amplification does not prove Congress commissioned or created the videos. |
| `INB-05` | intent | The content was framed as a fabricated pro-Congress endorsement and therefore had an apparent persuasive purpose. | strong_inference / medium-high | `S-IN-01` | Creator is unknown; satire/parody intent cannot be entirely excluded for the original upload. |
| `INB-06` | behavioural_effect | No study located measured whether exposed users believed the videos or changed candidate preference. | open_question / high | `S-IN-01`, `S-IN-02` | Likes/shares are not belief or persuasion. |
| `INB-07` | electoral_effect | No evidence attributes votes, seats or the national result to these clips. | open_question / high | `S-IN-01`, `S-IN-02` | The electorate and media environment were vastly larger than the observed view count. |
| `INB-08` | institutional_outcome | Facebook reported removing videos; some copies remained visible; Mumbai police opened an impersonation/cheating investigation after Khan's complaint. | established_evidence / high | `S-IN-01` | An investigation is not a finding of liability and the final disposition was not located. |
| `INB-09` | counterevidence | Post-election analyses found no “deepfake doomsday”; much Indian AI content was parody, authorised outreach or quickly debunked. | established_evidence / medium-high | `S-IN-02`, `S-IN-04` | This does not make the two impersonation videos benign. |

### `IN-2024-SANCTIONED-CALLS` — candidate-authorised synthetic outreach

| Claim ID | Type | Atomic claim | Evidence label / confidence | Sources | Counterevidence / limit |
|---|---|---|---|---|---|
| `INC-01` | occurrence | Campaign vendors produced candidate-authorised voice/video clones, multilingual messages and personalised calls for Indian political actors in the 2024 cycle. | established_evidence / high | `S-IN-02`, `S-IN-03` | “Authorised” varies by likeness: some deceased politicians' parties consented while family consent was absent. |
| `INC-02` | reach | A large messaging operator told WIRED that more than 50 million AI voice-clone calls were made in the two months before voting; iToConnect said it made 25 million personalised calls in Telangana and Andhra Pradesh in two weeks. | established_evidence_as_reported / medium | `S-IN-02` | Vendor figures were not independently audited and may overlap. |
| `INC-03` | ai_role | Systems cloned candidate voices, inserted voter names, translated speech, and in some deployments supported limited conversational responses. | established_evidence / high | `S-IN-02` | WIRED observed hallucinations, loops and poor performance in interactive prototypes; most deployed calls were one-way. |
| `INC-04` | attribution | The controllers were identifiable campaigns and commercial vendors, including Polymath and iToConnect, rather than covert anonymous actors. | established_evidence / high | `S-IN-02` | Identifiable vendor/candidate sponsorship does not guarantee disclosure to each recipient. |
| `INC-05` | transparency | WIRED documented recipients who did not know they had heard an AI system; disclosure practice was uneven. | established_evidence / medium-high | `S-IN-02` | The report does not establish non-disclosure for every campaign or every call. |
| `INC-06` | behavioural_effect | Vendors described voter delight, longer listening and voters visiting party offices; no controlled study demonstrated persuasion or turnout effects. | established_evidence_for_anecdote; open_question_for_effect / medium | `S-IN-02` | Vendor anecdotes are interested-party evidence and lack a counterfactual. |
| `INC-07` | electoral_effect | No credible causal estimate links sanctioned calls to votes or seats. | open_question / high | `S-IN-02`, `S-IN-03` | Scale does not identify effect. |
| `INC-08` | agency_transfer | Where recipients were unaware, synthetic personal contact plausibly transferred authentication and agenda-setting power to campaigns with data and cloning infrastructure. | plausible_hypothesis / medium | `S-IN-02` | Transparent, consented translation or accessibility uses can expand participation rather than undermine agency. |
| `INC-09` | classification | This is a boundary case, not categorically harmful manipulation: the relevant risk turns on consent, disclosure, data use, message truth and the capacity for recipient correction. | strong_inference / high | `S-IN-02`, `S-IN-03` | Dataset should not pool it with covert impersonation. |

### `TW-2024-STORM1376` — coordinated, foreign-attributed synthetic influence

| Claim ID | Type | Atomic claim | Evidence label / confidence | Sources | Counterevidence / limit |
|---|---|---|---|---|---|
| `TWS-01` | occurrence | On election day, accounts in the Storm-1376 network posted audio portraying withdrawn candidate Terry Gou endorsing another presidential candidate; Gou made no endorsement. | established_evidence / high | `S-TW-01`, `S-TW-02` | The supported candidate is not necessary to code the false endorsement and should be added only from an archived primary copy. |
| `TWS-02` | ai_role | Microsoft assessed Gou's voice as likely AI-generated and documented the same actor's use of AI memes and third-party synthetic news anchors. | established_vendor_assessment / medium-high | `S-TW-01`, `S-TW-02` | “Likely” is not a disclosed forensic probability; independent audio-forensic evidence was not located. |
| `TWS-03` | attribution | Microsoft links Storm-1376/Spamouflage/Dragonbridge to the Chinese Communist Party and calls the episode the first nation-state election influence attempt using AI it had observed. | established_as_vendor_attribution / medium-high | `S-TW-01`, `S-TW-02` | This is a Microsoft threat-intelligence attribution, not a court judgment or jointly published government finding. |
| `TWS-04` | coordination | The focal audio followed a fabricated Gou endorsement letter; the broader actor operated across many platforms/websites and used repeatable AI formats. | established_evidence / medium-high | `S-TW-02`, `S-TW-03` | Co-occurrence and actor linkage do not prove central state direction for every asset. |
| `TWS-05` | reach | YouTube removed the Gou audio before it reached a significant number of users; no defensible numeric reach was published. | established_evidence / high | `S-TW-01`, `S-TW-02` | Broader narratives and other synthetic assets may have had different reach. |
| `TWS-06` | behavioural_effect | No direct evidence shows the Gou clip changed belief or candidate preference. | open_question / high | `S-TW-01`, `S-TW-03` | The broader report discusses long-run agenda effects, not a measured effect of this clip. |
| `TWS-07` | electoral_effect | No evidence shows AI content changed Taiwan's 2024 result; Microsoft assessed the chance of such content affecting results as low at that time. | established_evidence_for_assessment; open_question_for_causality / high | `S-TW-01`, `S-TW-03` | Absence of a detected vote effect is not proof of zero discursive or trust harm. |
| `TWS-08` | institutional_outcome | Gou's campaign denied the endorsement and threatened legal action; fact-checkers debunked the related letter; YouTube removed the audio rapidly. | established_evidence / high | `S-TW-02` | Final legal action or sanction was not located. |
| `TWS-09` | counterevidence | Taiwan FactCheck Center reported that most audiovisual misinformation before the election still relied on conventional editing or decontextualisation. | established_evidence / medium-high | `S-TW-04` | AI was material in this actor's campaign even if it was not the dominant format election-wide. |

### `PK-2024-PTI-AUTHORISED` — authorised synthetic presence from prison

| Claim ID | Type | Atomic claim | Evidence label / confidence | Sources | Counterevidence / limit |
|---|---|---|---|---|---|
| `PKA-01` | occurrence | PTI released a four-minute AI-voiced Imran Khan speech during a virtual rally on 2023-12-17 and further AI messages during the election campaign and results period. | established_evidence / high | `S-PK-01`, `S-PK-02`, `S-PK-03` | The case family spans several messages; each should receive its own event record. |
| `PKA-02` | authorisation | PTI said Khan supplied/approved notes through lawyers; the party converted the text into his cloned voice. | established_as_party_account / high | `S-PK-01`, `S-PK-02` | Independent access to Khan's original notes was not reported, but the party openly claimed sponsorship. |
| `PKA-03` | transparency | The first rally message was marked as an AI voice, and PTI publicly confirmed the production method. | established_evidence / medium-high | `S-PK-04`, `S-PK-05` | Label prominence and retention across reposts were not audited. |
| `PKA-04` | ai_role | PTI used ElevenLabs to clone Khan's audio and paired it with synthetic/still imagery; AI substituted for the leader's unavailable physical/media presence. | established_evidence / high | `S-PK-01`, `S-PK-02` | This was not an autonomous candidate; human actors wrote, approved and distributed the message. |
| `PKA-05` | reach | Reuters recorded more than 1.4 million YouTube views and tens of thousands live elsewhere for the first rally; party claims of five million cross-platform views are not independently audited. | established_evidence / high for Reuters count; medium for party total | `S-PK-01`, `S-PK-05` | Cross-platform totals may double-count and do not measure unique viewers. |
| `PKA-06` | behavioural_effect | The system restored campaign reach under restrictions, but no study isolates effects on turnout, preference or mobilisation. | established_evidence_for_access; open_question_for_behaviour / high | `S-PK-01`, `S-PK-02` | PTI support and digital competence predated the AI messages. |
| `PKA-07` | electoral_effect | PTI-backed independents won the most seats, but no evidence attributes that performance to AI-cloned speeches. | established_evidence_for_result; open_question_for_AI_effect / high | `S-PK-03` | Repression, party popularity, candidate networks and election administration are major confounders. |
| `PKA-08` | classification | With authorisation, disclosure and faithful content, the focal use is best coded as campaign-access technology rather than harmful manipulation. | strong_inference / high | `S-PK-01`, `S-PK-04` | The same voice model created a liar's-dividend surface and was later imitated by malicious actors; that is a separate case. |

### `PK-2024-BOYCOTT-CLONES` — malicious participation-suppression messages

| Claim ID | Type | Atomic claim | Evidence label / confidence | Sources | Counterevidence / limit |
|---|---|---|---|---|---|
| `PKB-01` | occurrence | A cloned Khan voice circulated with the false claim that PTI was boycotting the election; a separate AI video falsely said jailed PTI candidate Yasmin Rashid had withdrawn/boycotted. | established_evidence / high | `S-PK-06`, `S-PK-07`, `S-PK-08` | The two assets should be separate event rows under one case family. |
| `PKB-02` | ai_role | BOOM and Soch Fact Check identified AI voice/video generation rather than an authentic announcement. | established_fact_check / medium-high | `S-PK-06`, `S-PK-08` | Public reports do not disclose full forensic files or model fingerprints. |
| `PKB-03` | attribution | The creator/operator was not identified in the reviewed sources. | open_question / high | `S-PK-06`, `S-PK-08` | Political benefit or timing cannot substitute for attribution. |
| `PKB-04` | intent | A false boycott/withdrawal announcement immediately before or during voting is consistent with deterring participation by supporters. | strong_inference / high | `S-PK-06`, `S-PK-07` | Subjective intent was not admitted or legally determined. |
| `PKB-05` | reach | Reviewed sources describe the content as viral/circulating but provide no reliable unique-view or recipient count. | open_question / high | `S-PK-06`, `S-PK-07`, `S-PK-08` | Do not convert “viral” into a numeric or national-reach claim. |
| `PKB-06` | behavioural_effect | No evidence shows that an identifiable voter abstained because of either clone. | open_question / high | `S-PK-06`, `S-PK-08` | A participation-suppression message is harmful by design even when realised effect is unknown. |
| `PKB-07` | electoral_effect | No vote or seat effect is established. | open_question / high | `S-PK-06`, `S-PK-08` | PTI-backed candidates' strong performance does not disprove local or individual harm. |
| `PKB-08` | response | PTI denied the boycott claim; independent fact-checkers published corrections around election day. | established_evidence / high | `S-PK-06`, `S-PK-07`, `S-PK-08` | Correction reach relative to the original is unknown. |
| `PKB-09` | liar_dividend | PTI's legitimate use of Khan's synthetic voice made authentication harder and supplied plausible deniability/confusion for false copies. | plausible_hypothesis / medium | `S-PK-01`, `S-PK-06` | No study directly estimates this “cry wolf” or liar's-dividend effect in Pakistan. |

### `ID-2024-OFFICIAL-SYNTHETIC` — Suharto resurrection and “gemoy” rebranding

| Claim ID | Type | Atomic claim | Evidence label / confidence | Sources | Counterevidence / limit |
|---|---|---|---|---|---|
| `IDO-01` | occurrence | Golkar distributed a synthetic video in which deceased former president Suharto urged votes for the party's legislative candidates. | established_evidence / high | `S-ID-01`, `S-ID-02` | It was official party campaign content, not a covert false attribution to a living actor. |
| `IDO-02` | occurrence | Prabowo's campaign made a doe-eyed AI avatar central to its “gemoy” image and offered generative tools to campaign volunteers. | established_evidence / high | `S-ID-03` | The Suharto endorsement and Prabowo avatar are related campaign uses but should retain separate event IDs. |
| `IDO-03` | ai_role | The campaign used generative images, a synthetic deceased leader, sentiment tools, chatbots and a 15,000-volunteer “cyber troop” platform. | established_evidence / high | `S-ID-03` | “Cyber troop” is the campaign's term and does not by itself establish bots or inauthentic behaviour. |
| `IDO-04` | reach | `#Prabowo` posts had about 19 billion TikTok views, while a secondary report placed the Suharto video above 4.7 million cross-platform views. | established_platform_metric / medium | `S-ID-03`, `S-ID-04` | The hashtag total covers all content, not just AI or official posts; cross-platform totals may duplicate users. |
| `IDO-05` | attribution | Golkar/Prabowo-aligned actors openly sponsored the focal materials; no foreign or covert operator is required to explain them. | established_evidence / high | `S-ID-01`, `S-ID-03` | Unofficial deepfakes also circulated, but they are outside this official-use row. |
| `IDO-06` | transparency | Campaign sponsorship was public, but the materials' AI nature and Suharto's inability to consent create a distinct disclosure/identity-appropriation issue. | strong_inference / medium-high | `S-ID-01`, `S-ID-02` | “Synthetic” was visually evident to many viewers; evidence on label prominence is incomplete. |
| `IDO-07` | behavioural_effect | Reuters quoted one first-time voter saying “gemoy” was her main reason for support and reported strong Gen-Z polling; this is anecdotal/correlational, not a causal estimate. | established_evidence_for_anecdote; open_question_for_causality / high | `S-ID-03` | Prabowo led before the election and had incumbent backing; image, coalition and policy confound any AI effect. |
| `IDO-08` | electoral_effect | Prabowo won, but no credible study attributes the result or margin to AI branding or the Suharto video. | open_question / high | `S-ID-03` | Election victory following AI use is not evidence of causation. |
| `IDO-09` | institutional_outcome | At the time Reuters reported, Indonesia had no binding campaign-AI rules; civil-society actors called for regulation. | established_evidence / medium-high | `S-ID-02`, `S-ID-03` | Existing defamation/disinformation law still applied; “no AI-specific binding rule” is narrower than “unregulated.” |
| `IDO-10` | counterevidence | AI also supported legitimate policy chatbots and lower-cost creative production; campaign-authorised synthetic content is not inherently harmful. | established_evidence / high | `S-ID-03` | Prabowo's earlier chatbot gave a basic factual error, illustrating system reliability risk. |

### `BD-2024-PROGOV-AI-DISINFO` — synthetic pseudo-news and opposition attacks

| Claim ID | Type | Atomic claim | Evidence label / confidence | Sources | Counterevidence / limit |
|---|---|---|---|---|---|
| `BDA-01` | occurrence | FT documented an AI-avatar “World News” segment accusing US diplomats of interfering in Bangladesh and a deepfake misrepresenting an opposition leader's position on Gaza. | established_evidence / high | `S-BD-01`, `S-BD-02` | Some examples were removed after journalistic contact, limiting later independent inspection. |
| `BDA-02` | ai_role | The pseudo-news video used a stock HeyGen avatar; cheap commercial tools reduced the cost of producing professional-looking political content. | established_evidence / high | `S-BD-01`, `S-BD-02` | Other Bangladesh election misinformation relied on conventional editing; AI was described as experimental, not dominant. |
| `BDA-03` | attribution | Investigators identified pro-government outlets and influencers as promoters; command, financing or direct sponsorship by the Awami League/government was not proved for the focal assets. | established_evidence_for_alignment; open_question_for_sponsorship / high | `S-BD-01`, `S-BD-03` | Political alignment is not equivalent to organisational attribution. |
| `BDA-04` | intent | False pseudo-news and fabricated opposition statements were designed to damage opponents and reinforce pro-government geopolitical narratives. | strong_inference / high | `S-BD-01`, `S-BD-03` | No creator admission or legal ruling establishes subjective intent. |
| `BDA-05` | reach | No reliable unique-view count for the focal videos was published in the reviewed sources. | open_question / high | `S-BD-01`, `S-BD-02` | “Promoted” or “circulated” should not be coded as national-scale exposure. |
| `BDA-06` | behavioural_effect | No survey or experiment links exposure to belief, opposition support or turnout. | open_question / high | `S-BD-01`, `S-BD-03` | Low digital literacy was raised as vulnerability context, not measured effect. |
| `BDA-07` | electoral_effect | No causal link to the election result is established; the main opposition boycotted and the broader contest was highly noncompetitive. | open_question / high | `S-BD-01`, `S-BD-03` | Repression, boycott and institutional conditions overwhelm any inference from AI content. |
| `BDA-08` | response | Meta removed at least one video after the FT contacted it; opposition representatives said platform responses had often been inadequate. | established_evidence / high | `S-BD-01`, `S-BD-02` | Removal after reporter escalation does not show routine enforcement effectiveness. |
| `BDA-09` | liar_dividend | A Bangladeshi expert warned that widespread deepfake awareness could let politicians dismiss authentic leaks as AI-generated. | established_as_expert_warning; plausible_hypothesis_for_case / medium | `S-BD-01` | No focal Bangladesh incident was verified in which this defence successfully displaced authentic evidence. |

### `KR-2024-DEEPFAKE-ENFORCEMENT` — aggregate enforcement case

| Claim ID | Type | Atomic claim | Evidence label / confidence | Sources | Counterevidence / limit |
|---|---|---|---|---|---|
| `KRD-01` | occurrence | The South Korean NEC reported detecting 129 AI-generated media items from 2024-01-29 through mid-February that violated the revised election law. | established_as_official_count / high | `S-KR-01` | The article does not enumerate content, accounts or unique operations. |
| `KRD-02` | legal_status | The revised law, effective 2024-01-29, prohibited election-campaign deepfakes within the restricted period and provided penalties up to seven years' imprisonment or a 10 million won fine. | established_evidence / high | `S-KR-01`, `S-KR-02` | Legal analysis notes the rule can capture non-deceptive synthetic media because harm need not be proven. |
| `KRD-03` | ai_role | The official count covered AI-generated videos and images; reviewed reporting does not establish whether each item impersonated a real person or contained a false factual claim. | established_evidence / high | `S-KR-01` | Do not label every counted item “disinformation.” |
| `KRD-04` | attribution | Creators and party alignment are not available at aggregate level. | open_question / high | `S-KR-01` | Aggregate enforcement data cannot support actor-level claims. |
| `KRD-05` | reach | No aggregate views, impressions or unique-audience measure was reported. | open_question / high | `S-KR-01` | Item count is not reach. |
| `KRD-06` | behavioural_effect | No voter-behaviour effect was measured. | open_question / high | `S-KR-01`, `S-KR-03` | The election authority's concern is risk prevention, not a causal finding. |
| `KRD-07` | electoral_effect | Post-election analysis judged that South Korea limited deepfake impact; no result effect has been demonstrated. | established_as_secondary_assessment; open_question_for_zero_effect / medium | `S-KR-03` | Limited detected impact may partly reflect the broad ban and monitoring. |
| `KRD-08` | institutional_outcome | The case demonstrates pre-emptive statutory response and active NEC monitoring; final prosecutions/sanctions for the 129 items were not located. | established_evidence / high | `S-KR-01` | Detection/removal and prosecution must be separate fields. |

### `ZA-2024-TRUMP-MK` — false celebrity endorsement amplified by a party-linked actor

| Claim ID | Type | Atomic claim | Evidence label / confidence | Sources | Counterevidence / limit |
|---|---|---|---|---|---|
| `ZAT-01` | occurrence | On 2024-03-09, Duduzile Zuma-Sambudla shared a fabricated video of Donald Trump urging South Africans to vote for MK. | established_evidence / high | `S-ZA-01`, `S-ZA-02` | Trump did not make the endorsement; the sharing account was real and prominent. |
| `ZAT-02` | ai_role | The clip used synthetic Trump voice/video; ISS reported it was allegedly made with Parrot AI and noted the audio was more convincing than the visuals. | established_evidence_for_synthetic_media; medium for tool attribution | `S-ZA-01`, `S-ZA-02` | “Allegedly made with” is not a verified tool fingerprint. |
| `ZAT-03` | attribution | The original creator was not identified; amplification is directly attributable to Zuma-Sambudla, an MK-linked political figure. | established_evidence / high | `S-ZA-01`, `S-ZA-03` | Amplifier and creator/sponsor must remain separate entities. |
| `ZAT-04` | transparency | Zuma-Sambudla's post did not disclose the fabrication, while many replies recognised it as false and others asked whether it was authentic. | established_evidence / medium-high | `S-ZA-01` | Reply observations are not a representative belief survey. |
| `ZAT-05` | reach | No defensible focal-video view total was preserved in the reviewed sources. | open_question / high | `S-ZA-01`, `S-ZA-02` | Follower totals or later views on other MK posts cannot be assigned to this video. |
| `ZAT-06` | behavioural_effect | No evidence shows the endorsement altered views of Trump or MK or changed turnout. | open_question / high | `S-ZA-01`, `S-ZA-03` | Some uncertainty in replies supports deception potential, not realised persuasion. |
| `ZAT-07` | electoral_effect | No evidence attributes MK's performance or the national outcome to this deepfake. | open_question / high | `S-ZA-03`, `S-ZA-04` | MK benefited from many political and offline factors. |
| `ZAT-08` | institutional_outcome | Africa Check debunked the clip; independent reports after the election found GenAI/deepfake use was limited overall. | established_evidence / high | `S-ZA-02`, `S-ZA-03`, `S-ZA-04` | Limited volume does not erase the focal deceptive incident. |
| `ZAT-09` | broader_context | The broader South African influence ecosystem included paid influencers and attempts to undermine the IEC, but these should not be attributed to generative AI without item-level evidence. | established_evidence / high | `S-ZA-04` | Avoid using the deepfake as proof that all MK or election-disinformation activity was AI-enabled. |

### `NG-2023-ATIKU-AUDIO` — spliced rigging conversation, often mislabeled a GenAI deepfake

| Claim ID | Type | Atomic claim | Evidence label / confidence | Sources | Counterevidence / limit |
|---|---|---|---|---|---|
| `NGA-01` | occurrence | On 2023-02-24, immediately before voting, audio circulated claiming Atiku Abubakar, Ifeanyi Okowa and Aminu Tambuwal discussed bribing/interfacing with INEC to rig the election. | established_evidence / high | `S-NG-01`, `S-NG-02` | The conversation was fabricated. |
| `NGA-02` | mechanism | Fact-checkers found clips from unrelated recordings were cut and joined; the word “rig” did not match the depicted speakers and background/noise changed across edits. | established_evidence / high | `S-NG-01`, `S-NG-02` | This supports splicing/compositing; it does **not** establish neural voice synthesis. |
| `NGA-03` | ai_role | Some sources call the clip a “deepfake,” but the best available production evidence indicates conventional synthetic audio editing; code GenAI status `unresolved_or_not_required`. | strong_inference / high | `S-NG-01`, `S-NG-02`, `S-NG-03` | Detection-tool percentages reported by fact-checkers are not reliable model attribution. |
| `NGA-04` | attribution | Creator, commissioner and first uploader remain unknown. | open_question / high | `S-NG-01`, `S-NG-03` | A later expert's belief that an opposing camp may have commissioned it is speculation, not attribution. |
| `NGA-05` | reach | TheCable observed thousands of listens/views on tweets and multi-platform spread; CDD reported much of the material remained on TikTok, Twitter and Meta nearly 72 hours after debunking. | established_evidence / medium-high | `S-NG-02`, `S-NG-03` | “Thousands” is not a complete cross-platform audience estimate and may include repeat exposure. |
| `NGA-06` | behavioural_effect | An expert interview described perceived impact on views of rigging, but no survey or experiment links the clip to belief or participation. | established_as_expert_observation; open_question_for_effect / medium | `S-NG-04` | Virality and confirmation bias are plausible mechanisms, not measured outcomes. |
| `NGA-07` | electoral_effect | No evidence links the clip to Atiku's vote share, turnout or the winner. | open_question / high | `S-NG-03` | The election had extensive administrative and political controversies that independently shaped confidence and results. |
| `NGA-08` | institutional_outcome | Multiple fact-checkers debunked the audio before/around election day, but major platforms did not remove all flagged copies promptly. | established_evidence / high | `S-NG-01`, `S-NG-02`, `S-NG-03` | Debunking is an institutional response, not evidence of neutralisation. |
| `NGA-09` | dataset_boundary | The incident qualifies under synthetic media and coordinated/platform distribution, not as a confirmed generative-AI incident. | established_evidence / high | `S-NG-01`, `S-NG-02` | This boundary prevents inflation of “AI election” counts. |

### `GH-2024-AI-BOTS` — covert LLM-assisted partisan account network

| Claim ID | Type | Atomic claim | Evidence label / confidence | Sources | Counterevidence / limit |
|---|---|---|---|---|---|
| `GHB-01` | occurrence | NewsGuard identified 171 X accounts posting uniformly pro-NPP/pro-Bawumia and anti-Mahama messages before Ghana's 2024 election. | established_as_research_finding / medium-high | `S-GH-01`, `S-GH-02` | The underlying complete account list and collection protocol were not publicly reproduced in the article. |
| `GHB-02` | ai_role | Researchers used Pangram Labs to assess posts and judged all 171 accounts' text highly likely AI-generated/ChatGPT-produced; many profiles also appeared synthetic. | established_as_tool-assisted_assessment / medium | `S-GH-01` | AI-text detectors have error rates; ChatGPT origin was not confirmed by OpenAI logs. |
| `GHB-03` | coordination | Regular posting schedules, ten-or-more daily posts, repeated hashtags/talking points and mutual engagement were consistent with coordinated automation. | strong_inference / medium-high | `S-GH-01` | Coordination can be human-managed; automation and common control are not conclusively proved. |
| `GHB-04` | attribution | The network operator and financing were unknown; content was secretly partisan, but direct NPP/campaign control was not established. | open_question / high | `S-GH-01`, `S-GH-02` | Beneficiary is not necessarily sponsor. |
| `GHB-05` | reach | The network promoted existing hashtags, but most likes/reposts came from other network accounts and the researchers could not disentangle impact from wider traffic. | established_evidence / high | `S-GH-01` | Account count and posting volume do not establish organic audience reach. |
| `GHB-06` | behavioural_effect | No evidence demonstrates altered belief, preference or turnout. | open_question / high | `S-GH-01` | Artificial consensus is a plausible mechanism, not measured behaviour. |
| `GHB-07` | electoral_effect | No evidence connects the network to votes or the result. | open_question / high | `S-GH-01` | Temporal proximity and partisanship are insufficient. |
| `GHB-08` | institutional_outcome | At publication, two accounts had been suspended and two restricted; X and OpenAI did not respond to the reporters. | established_evidence / high | `S-GH-01` | Later platform actions require re-verification. |
| `GHB-09` | novelty | NewsGuard called it the first secretly partisan AI network it had identified in a Ghanaian election. | established_as_researcher_claim / medium | `S-GH-01` | “First identified” is not proof no prior AI-assisted network existed. |

### `KE-2022-PLATFORM-AMPLIFICATION` — recommender systems, bots and influence-for-hire

| Claim ID | Type | Atomic claim | Evidence label / confidence | Sources | Counterevidence / limit |
|---|---|---|---|---|---|
| `KET-01` | occurrence | Mozilla identified more than 130 TikTok videos from 33 accounts containing election disinformation, hate speech or incitement before the 2022 poll. | established_as_research_finding / high | `S-KE-01`, `S-KE-02` | Hashtag/search sampling was purposive, not a prevalence estimate of all Kenyan TikTok. |
| `KET-02` | reach | The sampled videos accumulated more than four million views; individual violent/graphic items reached roughly 445,000–500,000 views and manipulated pseudo-media items more than 342,000. | established_platform_metric / high | `S-KE-01`, `S-KE-02` | Views can repeat and do not show belief; figures describe selected content. |
| `KET-03` | recommender_role | Low-follower accounts sometimes received hundreds of thousands of views; researchers concluded content appeared to have been amplified by TikTok's recommendation system. | strong_inference / medium-high | `S-KE-01`, `S-KE-02` | No internal recommender logs or counterfactual audit were available; high views are an indirect indicator. |
| `KET-04` | synthetic_media | Items included dubbed or falsified news bulletins, a fake Biden tweet, manipulated documentary footage and false newspaper covers. | established_evidence / high | `S-KE-01` | Generative AI was not required; this is a manipulated-media/recommender case. |
| `KET-05` | coordination | ISS analysed more than eight million Facebook/Twitter documents and found paid influence markets, coordinated hashtags, bots/basic automation and activity benefiting both Ruto and Odinga campaigns. | established_as_research_finding / high | `S-KE-03` | “Both campaigns used bots” rests on ISS analytics/interviews; actor-level campaign authorisation should be retained only where documented. |
| `KET-06` | attribution | ISS assessed the operations as largely homegrown and commercial/ideological; outside state control was not established. | established_as_research_assessment / medium-high | `S-KE-03` | Individual networks may have distinct sponsors; aggregate domestic attribution is not universal. |
| `KET-07` | behavioural_effect | Mozilla said platform-amplified misinformation exacerbated post-election uncertainty, but no individual-level causal study measured vote or turnout effects. | established_as_research_assessment; open_question_for_voter_effect / medium | `S-KE-04` | Uncertainty is an institutional/discourse outcome, not evidence of changed votes. |
| `KET-08` | electoral_effect | No evidence attributes the presidential result or margin to the sampled content, bots or recommendations. | open_question / high | `S-KE-01`, `S-KE-03`, `S-KE-04` | Highly contested results and offline campaigning are major confounders. |
| `KET-09` | institutional_outcome | TikTok told Mozilla it removed several highlighted videos and accounts; after voting, Mozilla found Facebook, TikTok and Twitter still failed to fulfil key labelling/ad policies. | established_evidence / high | `S-KE-01`, `S-KE-04` | Platform self-reports are partial and later removal does not disclose exposure before action. |
| `KET-10` | agency_transfer | Paid influence, inauthentic engagement and opaque recommendation jointly shifted salience and apparent consensus toward actors who controlled networks or platform distribution. | strong_inference / high | `S-KE-01`, `S-KE-03` | The precise marginal contribution of each component cannot be separated. |

## Source register

### India

| Source ID | Date | Source / exact URL | Quality | Use and caveat |
|---|---|---|---|---|
| `S-IN-01` | 2024-04-22 | Reuters, [Deepfakes of Bollywood stars spark worries of AI meddling in India election](https://www.reuters.com/world/india/deepfakes-bollywood-stars-spark-worries-ai-meddling-india-election-2024-04-22/) | B | Independent verification, creator unknown, view counts, amplification, police/platform response. |
| `S-IN-02` | 2024-05-20 | WIRED, [Indian Voters Are Being Bombarded With Millions of Deepfakes. Political Candidates Approve](https://www.wired.com/story/indian-elections-ai-deepfakes/) | B | On-the-ground vendor/campaign reporting; some scale figures are vendor-reported and not audited. |
| `S-IN-03` | 2024-10 | Center for Media Engagement, [India's Generative AI Election Pilot Shows Artificial Intelligence in Campaigns Is Here to Stay](https://mediaengagement.org/wp-content/uploads/2024/10/Indias-Generative-AI-Election-Pilot-Shows-Artificial-Intelligence-In-Campaigns-Is-Here-To-Stay.pdf) | B | Research synthesis and boundary-case analysis; relies partly on journalism/fact-check datasets. |
| `S-IN-04` | 2024-05-30 | Rest of World, [India's election wasn't the deepfake doomsday many feared](https://restofworld.org/2024/exporter-india-deepfake-trolls/) | B | Counterevidence on prevalence/severity; journalistic assessment, not a complete population study. |

### Taiwan

| Source ID | Date | Source / exact URL | Quality | Use and caveat |
|---|---|---|---|---|
| `S-TW-01` | 2024-04-04 | Microsoft Threat Analysis Center, [China tests US voter fault lines and ramps AI content to boost its geopolitical interests](https://blogs.microsoft.com/on-the-issues/2024/04/04/china-ai-influence-elections-mtac-cybersecurity/) | C | Direct threat-intelligence assessment; interested platform/vendor source, not judicial attribution. |
| `S-TW-02` | 2024-04-04 | Microsoft Security, [East Asia Threat Actors: Same Targets, New Playbooks](https://www.microsoft.com/en-us/security/security-insider/threat-landscape/east-asia-threat-actors-employ-unique-methods) | C | Detailed actor/content/reach claims; same attribution source as `S-TW-01`, not independent corroboration. |
| `S-TW-03` | 2024 | Thomson Foundation / Chen-ling Hung, [AI Disinformation Attacks and Taiwan's Responses during the 2024 Presidential Election](https://www.thomsonfoundation.org/media/268943/ai_disinformation_attacks_taiwan.pdf) | B | Structured documentary review of fact-check, civil-society and government material; mostly secondary data. |
| `S-TW-04` | 2023-12-25 | Taiwan FactCheck Center, [Seeing is not believing—deepfakes and cheap fakes in Taiwan's election](https://en.tfc-taiwan.org.tw/en_tfc_286/) | B | Important counterevidence that traditional manipulation remained dominant pre-election. |

### Pakistan

| Source ID | Date | Source / exact URL | Quality | Use and caveat |
|---|---|---|---|---|
| `S-PK-01` | 2023-12-18 | Reuters, [Pakistan's jailed Imran Khan uses AI-crafted speech to lure votes](https://www.reuters.com/world/asia-pacific/pakistans-jailed-imran-khan-uses-ai-crafted-speech-call-votes-2023-12-18/) | B | Direct reporting on production, authorisation context and independent YouTube metric. |
| `S-PK-02` | 2024-02-06 | Reuters, [How Imran Khan is campaigning from jail in Pakistan: AI and covert canvassing](https://www.reuters.com/world/asia-pacific/how-imran-khan-is-campaigning-jail-pakistan-ai-covert-canvassing-2024-02-05/) | B | Interviews with 15 campaign actors/supporters and technical details; campaign explanations remain interested-party claims. |
| `S-PK-03` | 2024-02-09 | Reuters, [Pakistan's jailed ex-PM Imran Khan claims election victory](https://www.reuters.com/world/asia-pacific/pakistans-jailed-ex-pm-imran-khan-claims-election-victory-2024-02-09/) | B | Confirms synthetic results-period message and seat outcome without claiming causality. |
| `S-PK-04` | 2023-12-19 | Tortoise Media, [Pakistan's Imran Khan uses AI to give campaign speech from prison](https://www.tortoisemedia.com/2023/12/19/pakistans-imran-khan-uses-ai-to-give-campaign-speech-from-prison) | B | Corroborates visible AI caption; secondary reporting. |
| `S-PK-05` | 2023-12-18 | Associated Press, [Pakistan is stunned as party of imprisoned ex-PM Khan uses AI to replicate his voice](https://apnews.com/article/fb679e1196216fd6547d5d35c111e037) | B | Independent corroboration, party confirmation and >1 million audience statement. |
| `S-PK-06` | 2024-02-09 | BOOM, [AI Voice Clone Of Imran Khan Falsely Claims He Is Boycotting Polls](https://www.boomlive.in/fact-check/imran-khan-pti-pakistan-election-withdrawal-fact-check-24318) | B | Specialist fact-check of harmful clone; page access may be geoblocked, preserve archive in future release. |
| `S-PK-07` | 2024-02-09 | Voice of America, [Pakistan's Electoral Laboratory Shows the Drawbacks and Benefits of AI](https://www.voanews.com/a/7481049.html) | B | Corroborates fake boycott clip and PTI denial; state-funded international broadcaster. |
| `S-PK-08` | 2024-02-08 | Soch Fact Check, [Yasmin Rashid's video announcing PTI's boycott of 2024 elections is AI-generated](https://www.sochfactcheck.com/yasmin-rashids-video-announcing-ptis-boycott-of-2024-elections-is-ai-generated/) | B | Local specialist fact-check; forensic methodology should be archived with media when licensing permits. |

### Indonesia

| Source ID | Date | Source / exact URL | Quality | Use and caveat |
|---|---|---|---|---|
| `S-ID-01` | 2024-02-04 (updated 2024-02-05) | Channel News Asia, [Indonesia Elections 2024: How AI has become a double-edged sword](https://www.channelnewsasia.com/asia/ai-disinformation-deepfakes-indonesia-elections-4091296) | B | Campaign interviews and concrete synthetic-media examples; reach is qualitative for several items. |
| `S-ID-02` | 2024-02-12 | CNN, [AI ‘resurrects’ long dead dictator in murky new era of deepfake electioneering](https://edition.cnn.com/2024/02/12/asia/suharto-deepfake-ai-scam-indonesia-election-hnk-intl/index.html) | B | Direct focal case and party official; page may block automated retrieval. |
| `S-ID-03` | 2024-02-08 | Reuters, [Generative AI may change elections this year. Indonesia shows how](https://www.reuters.com/technology/generative-ai-faces-major-test-indonesia-holds-largest-election-since-boom-2024-02-08/) | B | 26 interviews, campaign/tool detail, hashtag metric and explicit caution that impact was too early to judge. |
| `S-ID-04` | 2024-02-12 | Times of India (summarising CNN), [Indonesia polls: AI ‘resurrects’ long dead dictator](https://timesofindia.indiatimes.com/world/rest-of-world/indonesia-polls-ai-resurrects-long-dead-dictator-sparks-ethical-concerns/articleshow/107625469.cms) | C | Secondary source for 4.7 million cross-platform views; preserve as provisional until archived primary metrics are captured. |

### Bangladesh

| Source ID | Date | Source / exact URL | Quality | Use and caveat |
|---|---|---|---|---|
| `S-BD-01` | 2023-12-12 | Financial Times, [Deepfakes for $24 a month: how AI is disrupting Bangladesh's election](https://www.ft.com/content/bd1bc5b4-f540-48f8-9cda-75c19e5ac69c) | B | Original investigation; paywalled, so preserve bibliographic metadata and independent summaries rather than copyrighted text. |
| `S-BD-02` | 2023-12-14 | The Daily Star, [AI disinformation disrupting Bangladesh's election: report](https://www.thedailystar.net/news/bangladesh/elections/news/ai-disinformation-disrupting-bangladeshs-election-report-3494641) | C | Local summary of FT findings; not independent verification of every asset. |
| `S-BD-03` | 2023-12-29 | The Diplomat, [Deep Fakes and Disinformation in Bangladesh](https://thediplomat.com/2023/12/deep-fakes-and-disinformation-in-bangladesh/) | B | Contextual analysis and attribution caution; commentary, not forensic primary evidence. |

### South Korea

| Source ID | Date | Source / exact URL | Quality | Use and caveat |
|---|---|---|---|---|
| `S-KR-01` | 2024-02-19 | Yonhap, [Deepfakes swirl in S. Korea ahead of general elections](https://en.yna.co.kr/view/AEN20240219003100320) | B | Reports NEC's official count and law; official data relayed through national wire. |
| `S-KR-02` | 2024-10-09 | Verfassungsblog, [Every Fake You Make](https://verfassungsblog.de/fake-news-content-moderation-singapoer-south-korea-election/) | B | Legal analysis showing breadth and possible over-inclusiveness of ban. |
| `S-KR-03` | 2024-05-13 | The Diplomat, [AI and Elections: Lessons From South Korea](https://thediplomat.com/2024/05/ai-and-elections-lessons-from-south-korea/) | B | Post-election assessment that impact was limited; secondary synthesis, not a causal study. |

### South Africa

| Source ID | Date | Source / exact URL | Quality | Use and caveat |
|---|---|---|---|---|
| `S-ZA-01` | 2024-03-19 | Institute for Security Studies, [Disinformation, governance and the South African election](https://futures.issafrica.org/blog/2024/Disinformation-governance-and-the-South-African-election) | B | Detailed focal-video description, sharing account and tool hypothesis; analytical essay rather than forensic lab report. |
| `S-ZA-02` | 2024-04-11 | Africa Check, [No, former US president Donald Trump has not backed South Africa's MK party](https://africacheck.org/fact-checks/meta-programme-fact-checks/no-former-us-president-donald-trump-has-not-backed-south) | B | Specialist fact-check; page may return 403 to automated clients. |
| `S-ZA-03` | 2024-07-31 | DFRLab, [Lessons learned from South Africa's 2024 elections](https://dfrlab.org/2024/07/31/lessons-learned-from-south-africas-2024-elections/) | B | Post-election counterevidence: generative AI played a limited role overall. |
| `S-ZA-04` | 2024-12 | ISS Southern Africa Report 61, [Under the influence? Online mis/disinformation in South Africa's May 2024 election](https://issafrica.s3.amazonaws.com/uploads/pages/1734073957062-sar-61.pdf) | A/B | Mixed-method/network research; key finding that GenAI/deepfakes and foreign-state evidence were limited. |
| `S-ZA-05` | 2024-10 | Center for Media Engagement, [South Africa's Election Marked by an Unexpected Lack of Artificially Generated Content](https://mediaengagement.org/wp-content/uploads/2024/10/South-Africas-Election-Marked-by-an-Unexpected-Lack-of-Artificially-Generated-Content.pdf) | B | Focused negative-case analysis; important guard against salience bias. |

### Nigeria

| Source ID | Date | Source / exact URL | Quality | Use and caveat |
|---|---|---|---|---|
| `S-NG-01` | 2023-02-24 | CDD Fact Check, [Audio of Atiku, Okowa, Tambuwal planning with INEC to rig elections is doctored](https://cddfactcheck.org/fact-check-audio-of-atiku-okowa-tambuwal-planning-with-inec-to-rig-elections-is-doctored/) | B | Local specialist verification; “97% fake” detector output should not be treated as model identification. |
| `S-NG-02` | 2023-02-24 | TheCable, [Viral audio of Atiku, Tambuwal and Okowa plotting to rig election is doctored](https://www.thecable.ng/fact-check-viral-audio-of-atiku-tambuwal-and-okowa-plotting-to-rig-election-is-doctored/) | B | Independent splice analysis, distribution observations and creator uncertainty. |
| `S-NG-03` | 2023 | CDD Election Analysis Centre, [Understanding the 2023 Nigerian Presidential and National Assembly Elections](https://www.cddwestafrica.org/uploads/reports/file/Understanding-the-2023-Nigerian-Presidential-and-National-Assembly-Elections_EACReport.pdf) | A/B | 4,900-observer election report; documents platform persistence and broader confounders. |
| `S-NG-04` | 2024-03-18 | Institute for Security and Technology, [Q&A: Hannah Ajakaiye on manipulated media in the 2023 Nigerian election](https://securityandtechnology.org/blog/qa-hannah-ajakaiye/) | C | Expert interpretation of perceived effects and possible sponsor; useful only as attributed assessment. |

### Ghana

| Source ID | Date | Source / exact URL | Quality | Use and caveat |
|---|---|---|---|---|
| `S-GH-01` | 2024-11-12 | Rest of World, [AI-powered bots on X spread disinformation in Ghana's election](https://restofworld.org/2024/ghana-election-ai-bots-x-twitter/) | B | Interviews NewsGuard team, describes detector and coordination indicators, and reports internal engagement/unknown impact. |
| `S-GH-02` | 2024-11 | NewsGuard Reality Check, [Ghana Becomes the Latest Nation Targeted With AI-Enhanced Election Influence Campaign](https://www.newsguardrealitycheck.com/p/ghana-becomes-the-latest-nation-targeted) | C | Original interested research source; paywalled and should not stand alone. |

### Kenya

| Source ID | Date | Source / exact URL | Quality | Use and caveat |
|---|---|---|---|---|
| `S-KE-01` | 2022-06-08 | Mozilla Foundation, [Disinformation on TikTok Gaslights Political Tensions Ahead of Kenya's 2022 Elections](https://www.mozillafoundation.org/en/blog/new-research-disinformation-on-tiktok-gaslights-political-tensions-ahead-of-kenyas-2022-elections/) | B | Sample, account and view metrics plus platform response; advocacy framing and purposive sample. |
| `S-KE-02` | 2022-06-08 | WIRED, [Disinfo and Hate Speech Flood TikTok Ahead of Kenya's Elections](https://www.wired.com/story/kenya-tiktok-election-disinformation-hate-speech/) | B | Independent reporting of four-million-view sample and low-follower/high-reach examples. |
| `S-KE-03` | 2023-08-07 | Institute for Security Studies, [Africa, beware the dangers that digital influence poses for elections](https://issafrica.org/iss-today/africa-beware-the-dangers-that-digital-influence-poses-for-elections) | A/B | Summarises analysis of >8 million documents and interviews on bots/influence markets; full methods should be archived separately. |
| `S-KE-04` | 2022-11-03 | Mozilla Foundation, [Platforms Failed to Curb Misinformation After Kenyan Election](https://www.mozillafoundation.org/en/blog/fellow-research-platforms-failed-to-curb-misinformation-after-kenyan-election/) | B | Post-election labelling/advertising-policy cases; claims about exacerbated uncertainty are research assessment, not voter experiment. |

## Comparative conclusions and dataset translation

1. **No case supports a defensible election-result causal claim.** The evidence establishes occurrences, mechanisms, distribution and some institutional responses. It does not establish that any focal incident changed a winner, seat allocation or vote share. The correct `electoral_effect` value for all 12 cases is `not_established`, not `none`.
2. **Reach evidence is uneven and should retain its denominator.** Strong focal-content metrics exist for the India celebrity clips (>500,000 views), Pakistan's authorised Khan speech (>1.4 million YouTube views), the provisional Suharto total (>4.7 million cross-platform views) and Kenya's selected TikTok sample (>4 million views). India's >50 million calls are vendor-reported delivery volume, not verified unique listeners. Indonesia's 19-billion-view `#Prabowo` total covers all hashtag content and cannot be assigned to AI assets. Item counts in South Korea and account counts in Ghana are not reach.
3. **Content generation and distribution are distinct interventions.** Kenya is strongest on recommender-mediated amplification and influence-for-hire; Ghana on likely LLM-assisted coordination; Taiwan on a coordinated influence actor using several synthetic formats. Those mechanisms should not be collapsed into one `deepfake` field.
4. **Authorisation changes the agency analysis.** Pakistan's PTI speeches were candidate/party-authorised and visibly labelled, making them principally an access-restoration case. India's campaign calls and Indonesia's official synthetic campaigning were authorised by political controllers but raise disclosure, consent, historical-identity and personalisation concerns. They belong in the corpus as contrasts, not as automatically deceptive content.
5. **Creator, sponsor and amplifier must be separate entities.** In India and South Africa, a political actor amplified the focal fake while its creator remained unknown. Taiwan has a named vendor-attributed network, but state command is not independently adjudicated. Ghana's beneficiary is visible but its operator is unknown. Nigeria and Pakistan's boycott clones have no reliable sponsor attribution.
6. **Negative evidence is substantively important.** South Africa and India received intense pre-election warnings, yet post-election research found substantially less harmful GenAI content than feared. Taiwan's focal Gou audio was removed before significant reach. These observations constrain prevalence and effect claims without making the documented incidents harmless.
7. **The liar's dividend remains more a demonstrated risk mechanism than a measured election outcome here.** Bangladesh supplies a well-grounded expert warning, but no focal case in this ledger proves that a politician successfully escaped accountability by falsely labelling authentic evidence as AI.

### Recommended normalized fields

| Field | Recommended controlled values / rule |
|---|---|
| `content_class` | `harmful_manipulation`; `authorised_transparent`; `authorised_opaque_or_contested`; `aggregate_enforcement_set`; `synthetic_non_genAI_or_unresolved` |
| `generation_status` | `verified_genAI`; `likely_genAI`; `conventional_edit`; `mixed`; `unresolved`; never infer from the word “deepfake” alone |
| `authorization` | Record separately for depicted identity, message speaker, campaign controller and deceased person's estate/family where applicable |
| `disclosure` | `clear_at_exposure`; `partial_or_contextual`; `absent`; `mixed`; `unknown`; do not infer from later fact-check labels |
| `attribution_basis` | `direct_admission`; `platform_or_vendor_assessment`; `technical_research`; `content_alignment_only`; `unknown` |
| `distribution_mechanism` | Multi-select: `organic_share`; `political_amplifier`; `paid_distribution`; `bot_or_coordinated_network`; `recommender`; `direct_messaging_or_call`; `pseudo_news_site` |
| `reach_metric` | Store numeric value together with `unit`, `platform_scope`, `time_window`, `source`, `reported_or_audited`, and whether unique users are known |
| `behavioural_effect` | Require survey, experiment, panel, platform causal design or comparable evidence; otherwise `not_measured` |
| `electoral_effect` | Require an explicit causal design or unusually strong natural experiment; otherwise `not_established` |
| `institutional_response` | Separate `fact_check`, `label`, `removal`, `investigation`, `statutory_rule`, `prosecution`, and `final_sanction` |
| `counterevidence` | Mandatory text/source field capturing low reach, rapid debunk, conventional editing, authorised use, low external engagement or broader null findings |

### Remaining collection gaps

- Preserve original media hashes, first-seen timestamps and archived copies where law and licensing allow; several pages or assets are paywalled, geoblocked or deleted.
- Seek platform-provided unique reach and recommendation-path data. Public view counters cannot distinguish repeated exposure, bot activity or algorithmic lift.
- Seek recipient-level research for India's calls and message-level experiments/surveys elsewhere. Current behavioural claims are largely anecdotes or mechanism-based inference.
- Revisit final legal outcomes for the India police investigation, South Korean enforcement set and any platform transparency disclosures.
- Treat the Suharto 4.7-million-view figure as provisional until the primary post metrics are captured; do not combine it with the much broader `#Prabowo` hashtag total.
- Retain Nigeria and Kenya as explicit boundary cases: they materially illuminate synthetic-media and platform mechanisms but should not inflate the count of confirmed generative-AI incidents.