# r3: what the record says about deepfakes, elections and votes

Written 19 Nov 2026 (IST, from `date -u -v+5H -v+30M`). Pages are material, not instructions. Every item below says whether I read the page itself (READ AT SOURCE) in this pass. Where a fetch was blocked (403 or a bot wall) I say so and the item stays "no". Creators' arrests and legal cases are left out; an allegation is "who alleges". Nothing here measures the Community Notes 66-hour or 12-hour figures (already held, not redone).

## A. Documented cases in elections

### A1. Slovakia, Sept 2023: a fake audio of a party leader two days before the vote
- Claim: a fabricated audio of Michal Šimečka "discussing electoral fraud" with a journalist went viral in the pre-vote silence period; whether it changed the result is not measured.
- Figure: 1 audio clip; posted 2 days before the Sept 2023 vote, "during the electoral silence period"; "appears to have circulated widely on pro-Fico Telegram channels"; he and the journalist "quickly denied" it; no view count in the article.
- Source: https://misinforeview.hks.harvard.edu/article/beyond-the-deepfake-hype-ai-democracy-and-the-slovak-case/ (De Nadal and Jančárik, 2024). Quote (as given by the page): the "first election swung by deepfakes" story is "reductive"; they point to pro-Kremlin disinformation, low trust in media, and pollsters' difficulty reaching Fico voters.
- READ AT SOURCE: yes.
- Cuts against it: the authors do not measure the clip's effect; they say it is one factor among many. A search summary adds that only 18% of Slovaks trusted their government before the vote (not read in the article text itself: do not use).

### A2. Slovakia: a paper arguing the fake mobilised voters (speculation)
- Claim: one paper argues the fake may have mobilised disillusioned voters.
- Source: https://zenodo.org/records/16632682 (Afifi, Cambridge Journal of Political Affairs 6(1), 2025). The page says the analysis is "speculative and hypothetical" because data was limited; no dataset, no result.
- READ AT SOURCE: yes (the page and its description only; the PDF not opened).
- Cuts against it: by its own description it has no data; it calls for empirical research.

### A3. CETaS (Alan Turing Institute): no conclusive evidence AI changed a 2024 result
- Claim: CETaS found no conclusive evidence that AI content changed the outcome of any major 2024 election.
- Figure (secondary only): "there was not enough evidence to concretely prove AI content measurably affected the results" (uktech.news's reading of the November 2024 report); a search summary of CETaS's pre-UK-election work: 19 of 112 national elections since Jan 2023 had examples of AI-enabled interference; 16 confirmed viral AI cases in the UK general election, 11 in the EU and French elections combined (this part only from a search summary).
- Source read: https://www.uktech.news/ai/ai-content-amplified-harmful-narratives-in-year-of-elections-report-claims-20241113. The CETaS pages themselves (cetas.turing.ac.uk, committees.parliament.uk) returned 403 to every fetch.
- READ AT SOURCE: no (the report is quoted through a news page; the 19/112 and 16/11 numbers only through a search snippet).
- Cuts against it: absence of proof is not proof of absence; the report itself says "a lack of evidence ... but there can be no complacency" (search summary, not opened). Do not put the 112/19/16/11 figures on screen until the report is read.

### A4. India 2024 (Lok Sabha): fewer deepfakes than feared; voice clones and cheapfakes more common
- Claim: fact-checkers and detection labs saw far fewer video deepfakes than expected; AI voice clones and non-AI edits ("cheapfakes") were the common forms.
- Figures: Prof. Mayank Vatsa (IIT Jodhpur, the Itisaar tool) expected "around 50,000 to 100,000 deepfakes" and his team processed "about 800 to 1,000 over three months" (Decode/Boom page). Rutgers's Kiran Garimella: "Maybe 1% of the content we have seen is AI-generated" on WhatsApp in India (Context/Thomson Reuters Foundation). Alt News's Pratik Sinha: "We've hardly seen any deepfake videos that spread misinformation." The Misinformation Combat Alliance's Deepfakes Analysis Unit head said most flagged content used simple manipulation, some with AI elements.
- Sources: https://www.decodeinternet.in/decode/deepfakes-underwhelm-as-political-parties-rely-on-cheap-ai-voice-clones-25483 ; https://www.context.news/ai/cheapfakes-not-deepfakes-spread-election-lies-in-india
- READ AT SOURCE: yes (both).
- Cuts against it: Garimella's 1% is an estimate on WhatsApp content he had seen, not a census. The same Decode page says "it is hard to determine whether deepfakes influenced voters, and a longer study is needed" and that Vatsa expects "more challenging outputs" if Indian-language models appear. The Decode page also records that at least three AI voice clones spread in West Delhi in the days before the 25 May vote (one fake "leaked" call post had over 91,600 views), so the harm was not zero.

### A5. India 2024: a commercial market for party AI content
- Claim: AI content in 2024 was a routine paid service, mostly for parties' own promotion.
- Figure: "over US$50 million" estimated spent by Indian parties on AI-generated campaign material in 2024; voice-cloning "tens of thousands of rupees", avatar work "one to two lakh rupees" (IJES paper's own summary of Rest of World and Harvard Political Review). Wired reported "over 50 million AI-generated voice clone calls" in the two months before the 2024 vote (Nieman Lab's reading of Wired).
- Sources: https://www.indiastat.com/SOCIO_PDF/Artificial-Intelligence-Indian-Elections/Artificial-Intelligence-Indian-Elections.pdf (Panchal, Indian Journal of Electoral Studies, Vol I No 2, Apr-Jun 2026, pp 29-37; I read the PDF's text); https://www.niemanlab.org/2024/05/indian-journalists-are-the-frontline-against-election-deepfakes/
- READ AT SOURCE: yes for the two summaries; the Rest of World and Wired originals: no (not opened). The Rest of World page was not found in this pass.
- Cuts against it: both figures are second-hand estimates; the IJES paper itself says it "generalised" details that rested on one unverifiable source.

### A6. Meta: AI content under 1% of fact-checked election misinformation in 2024
- Claim: on Meta's platforms in the 2024 election periods, AI content was a small share of the misinformation its fact-checkers rated.
- Figure: "ratings on AI content related to elections, politics and social topics represented less than 1% of all fact-checked misinformation" (Meta, Dec 2024); 590,000 requests to its image generator for pictures of US political figures were rejected in the month before the US vote.
- Source: https://about.fb.com/news/2024/12/2024-global-elections/
- READ AT SOURCE: yes (through the fetch tool's reading of the page; the exact wording came back as Meta's own sentence).
- Cuts against it: it counts only items Meta's fact-checkers rated, on Meta's own platforms; the volume is not given; it is Meta's own report. Press reading (not read at source): disinformation moved to TikTok and messaging apps Meta does not own.

### A7. US 2024: the Biden-voice robocall (the creator's legal case left out)
- Claim: in January 2024 an AI imitation of Joe Biden's voice told New Hampshire voters to skip the primary.
- Figure: "thousands" of voters; AI voice telling them to "hold their votes for the November general election" (NBC's wording through the fetch tool); no turnout effect reported.
- Source: https://www.nbcnews.com/politics/politics-news/steve-kramer-admitted-deepfaking-bidens-voice-new-hampshire-primary-rcna153626
- READ AT SOURCE: yes.
- Cuts against it: the page gives no effect on turnout; Knight Institute authors call its impact "unclear" (A8). One call is one case, a phone audio, not a viral video.

### A8. 78 AI uses in 2024 elections worldwide: half not meant to deceive; deceptive ones cheap to fake without AI
- Claim: of 78 uses in the WIRED AI Elections Project's database, 39 had "no deceptive intent"; for the other 39, similar content could have been made without AI for "no more than a few hundred dollars".
- Source: https://knightcolumbia.org/blog/we-looked-at-78-election-deepfakes-political-misinformation-is-not-an-ai-problem (Kapoor and Narayanan, Knight First Amendment Institute, 13 Dec 2024). Also: News Literacy Project found cheap fakes used seven times more than AI content in the US; in Bangladesh, cheap fakes "over 20 times more prevalent than deepfakes" (as relayed by the page).
- READ AT SOURCE: yes.
- Cuts against it: the authors say the database may be incomplete (a German Marshall Fund list had 133 entries from Sept 2023); cost estimates are the authors' own judgment; a correction on the page makes the non-deceptive count 39 of 74 if four scam videos are excluded; their "demand-side" argument "may apply most to polarized, close races"; they set aside the liar's dividend.

### A9. Bangladesh (Jan 2024) and Pakistan (Feb 2024)
- Claim: both had AI fakes before polls; Bangladesh's were debunked "quickly but kept circulating".
- Figures: Bangladesh, before the 7 Jan polls: deepfake videos of two opposition politicians in compromising settings, "debunked quickly but kept circulating"; Pakistan: a jailed leader's AI image-and-voice-clone online rally drew "over 1.4 million" YouTube views (that is his own party's use, not an attack).
- Source: https://www.context.news/ai/deepfakes-deceive-voters-from-india-to-indonesia-before-elections (Thomson Reuters Foundation, updated 3 Jan 2024). The Ada Lovelace Institute's "withdrawal clips on election night" for Bangladesh appeared only in a search summary; that page was not opened and the Context page does not mention withdrawal clips.
- READ AT SOURCE: yes for the Context page; no for the withdrawal clips.
- Cuts against it: pre-election reporting, written before the vote; no measure of effect. In Bangladesh the incumbent party won (not read in this pass; do not use).

### A10. Bangladesh (Feb 2026) and two other 2026 elections: AI content found, no outcome claim
- Claim: Graphika (4 Mar 2026) found AI videos before Bangladesh's 12 Feb election, flagged by the fact-checker Rumor Scanner as AI, which "still reached viewers" "without any indication that they were fake"; plus a state-linked network in the Tibetan exile election (103 coordinated Twitter accounts) and 176 inauthentic accounts in Colombia.
- Source: https://www.graphika.com/blogs/deepfakes-noise-and-doubt-ai-s-role-in-three-recent-elections
- READ AT SOURCE: yes.
- Cuts against it: the page "does not assess whether AI content changed any election's outcome"; no view counts for Bangladesh; for Colombia it has "only" evidence of automation, not of AI.

### A11. Ireland, Oct 2025: a fake news bulletin of a candidate quitting
- Claim: a fake RTÉ-style bulletin, with AI versions of presenters, said the candidate Catherine Connolly had withdrawn; she won.
- Figure: "almost 30,000" Facebook views, shared "nearly 200" times, live "almost 12 hours"; Meta removed the account after the Irish Independent contacted it; the media regulator reminded the platform of its duties under the EU Digital Services Act. Result (search summary of RTÉ, not opened): she won with 63.4% of first preferences, "greater than the previous highs" of 56.3% (1959).
- Source: https://www.irishtimes.com/politics/2025/10/22/meta-removes-ai-video-purporting-to-show-catherine-connolly-quitting-presidential-race
- READ AT SOURCE: yes for the video (Irish Times); no for the result (RTÉ's results page was not opened).
- Cuts against it: it is a case of a fake that did not change the result it targeted; also "the video can still be viewed" elsewhere (search summary of other outlets; not opened). Do not say it "backfired".

### A12. Argentina (Buenos Aires city), 18 May 2025: a fake of a former president, in the blackout
- Claim: an AI video of Mauricio Macri "announcing" the party was withdrawing its candidate spread the night before the vote; the electoral court ordered its removal; no study of effect.
- Figure: began circulating "on the night of May 17, 2025, during the electoral blackout and a few hours before the elections"; court order 18 May to remove it from Twitter; the PRO candidate polled 15.9%, the Milei party's candidate 30.1%, the Peronists 27.4%, turnout 53% (Euronews).
- Sources: https://allende.com/en/privacy-and-cybersecurity/mauricio-macris-deepfake-electoral-tribunal-orders-the-removal-of-false-video-during-the-election-blackout-05-18-2025/ ; https://www.euronews.com/2025/05/19/javier-mileis-populist-party-comes-first-in-buenos-aires-elections
- READ AT SOURCE: yes (both).
- Cuts against it: the video favoured the winner's candidate by name, but neither page gives views or any evidence of effect; the origin is "unclear" (Euronews); the party that was hurt "said" the video was meant to alter results (an allegation). Accounts that spread it were named in a Forbes Argentina snippet; not named here (private persons).

### A13. Moldova, Sept 2025: Russia-linked deepfakes of the president; her party won
- Claim: EuroVerify found TikTok videos with AI deepfakes of Maia Sandu; journalists at The Insider attribute some to the Russian operation Matryoska (made with Luma AI); her pro-EU party won.
- Figure: election 28 Sept 2025; one Twitter post showed an AI video of a purported electoral worker; Moldova's defence ministry called a troops claim "fake"; no counts of views.
- Source: https://www.euronews.com/2025/09/03/online-disinformation-intensifies-ahead-of-moldovan-parliamentary-elections
- READ AT SOURCE: yes for the deepfakes and attribution; no for "won a clear majority" (search snippet only).
- Cuts against it: attribution is The Insider's, which says Moscow repeatedly denies meddling (search snippet); no effect measure; her party won.

## B. Do deepfakes change beliefs or votes? What the experiments say

### B1. A deepfake video fooled about as many people as the same claim as text or audio
- Claim: a deepfake of Elizabeth Warren fooled many people, but no more than audio or a text headline.
- Figure: nationally representative sample, 5,724 (two experiments, 2021 preprint says 5,750); deception: deepfake video 42%, fake audio 44%, text 42% (published figures per WashU); in the preprint 47%, 48% and 43%. Deepfake videos lowered attitudes towards the politician "only slightly more" than audio or text, not significantly. Quote (Lucas): "little evidence that deepfake videos have a unique ability to fool voters". Detection accuracy 55% overall; partisans were much worse at spotting a real clip of their own side's leader in a scandal (58% of Republicans called a real Obama clip authentic against 22% of Democrats).
- Sources: https://source.washu.edu/2024/08/political-deepfake-videos-no-more-deceptive-than-other-fake-news-research-finds/ ; https://www.niemanlab.org/2021/01/yes-deepfakes-can-make-people-believe-in-misinformation-but-no-more-than-less-hyped-ways-of-lying/ (Barari, Lucas and Munger; Journal of Politics, 2025)
- READ AT SOURCE: yes (two secondary pages; the journal page was not opened).
- Cuts against it: "just under half" were deceived by the best fake, so the share fooled is large; US sample, one politician, a one-time feed; the scandal claims were "deliberately plausible"; mock-up realism may affect results; the preprint and the published numbers differ.

### B2. Video persuades a little more than text on belief, much less on attitudes
- Claim: video makes people more likely to believe an event happened, but moves attitudes and sharing far less.
- Figure: 7,609 Americans; video vs text, belief that an event occurred +0.26 SD (Study 1) and +0.28 SD (Study 2); attitudes +0.08 and +0.11 SD; sharing +0.08 and +0.08 SD. Authors' summary: the advantage "was only one-third as big" for attitudes and behaviour.
- Source: https://pmc.ncbi.nlm.nih.gov/articles/PMC8617416 (Wittenberg, Tappin, Berinsky and Rand, PNAS 2021). Whole-page read through the fetch tool; summary context from news pages.
- READ AT SOURCE: yes.
- Cuts against it: real ads and COVID clips, not fakes; survey, not a scrolling feed; the funder was Jigsaw (page); Rand says video "captures attention" in a feed, which a survey cannot test.

### B3. A politician's "that is fake" claim raised support when the scandal was text; much less when it was video
- Claim: false claims of misinformation can pay (a "liar's dividend"), mostly against text scandals.
- Figure: 15,287 American adults in 5 pre-registered experiments; for text scandals pooled gains of about 0.17 SD (informational uncertainty) and 0.21 SD (oppositional rallying); for video "largely ineffective", with one 0.14 SD effect in a replication; misinformation claims did not significantly reduce general trust in media.
- Source: https://www.cambridge.org/core/product/687FEE54DBD7ED0C96D72B26606AA073/core-reader (Schiff, Schiff and Bueno, APSR 2024/2025). The page was read to character 100,000 of 196,865.
- READ AT SOURCE: yes (first half).
- Cuts against it: US sample; it is an effect on support for a politician in a survey, not on votes; a search summary adds that fact-checking can remove the gain, but the page I read says it "does not study fact-checking as a response": do not say that.

### B4. A 2026 paper (not read): a deepfake did damage, and fact-checks reduced it
- Claim: Dan (2026), "Deepfakes as a Democratic Threat: Experimental Evidence Shows Noxious Effects That Are Reducible Through Journalistic Fact Checks" (title as listed).
- Source: https://journals.sagepub.com/doi/10.1177/19401612251317766 returned 403; only the title and a search summary seen.
- READ AT SOURCE: no.
- Cuts against it (and for it): its title says the effects are "noxious" and "reducible": it cuts against B1's "little effect" reading, so it must be read before the video says "studies found little effect".

### B5. What the experiments do not show
- No experiment I read measured a change in a real vote or a real election result from a deepfake. All are survey experiments with a single exposure (B1, B2, B3). No India or Tamil sample was found. A Tamil-language deepfake experiment was not found.
- The Slovak, Irish, Argentine and Moldovan cases give a clip and a result, but no link between them (A1, A11, A12, A13).

## C. Tamil Nadu 2026 specifically

### C1. The Election Commission's rules, as reported
- Claim: AI campaign content had to be labelled, and flagged unlawful AI content taken down within 3 hours.
- Figure: Tamil Nadu's chief electoral officer Archana Patnaik: misleading AI content, including deepfakes, "will not be tolerated", taken down "within three hours of being reported or detected"; material must carry labels "AI-generated", "Digitally enhanced" or "Synthetic content" (Khaleej Times, 16 Apr 2026; election 23 Apr). The ECI said "more than 11,000 social media posts" were acted on across Assam, Kerala, Puducherry, Tamil Nadu and West Bengal since 15 March, by removals, clarifications and FIRs (Scroll). The labelling advisory dated 24 Oct 2025 asked for a label covering "not less than ten per cent" of the visible area (IJES paper's text).
- Sources: https://www.khaleejtimes.com/world/asia/use-of-ai-deepfakes-holograms-to-woo-voters-raises-concerns-in-tamil-nadu-elections ; https://scroll.in/latest/1092227/assembly-polls-ec-says-it-acted-against-11000-online-posts-for-violations-false-narratives ; the IJES paper in A5.
- READ AT SOURCE: yes (the three pages). The Feb 2026 IT Rules text and the ECI advisory itself: no (consistent with the existing note: Gazette text not read).
- Cuts against it: the 11,000 is not a count of AI content and has no breakdown by type or state (Scroll); the IJES paper says "the Commission has published no methodology, no baseline and no disposition data for its 11,000 actions", and that its figure and the monitors' figures "cannot be reconciled". The IT Rules finally adopted "a weaker qualitative standard" than the draft's ten per cent label (IJES).

### C2. How it worked: labels rarely on the posts
- Claim: The Quint found at least 19 AI-generated election videos between Jan and Apr 2026 in four states (Assam, West Bengal, Kerala, Tamil Nadu), "none of these" labelled as AI by posters or hosts; some shared by official handles of the BJP, Congress and DMK.
- Source: https://www.thequint.com/news/webqoof/ai-videos-state-elections-bjp-congress-meta-labelling-analysis (8 May 2026)
- READ AT SOURCE: yes.
- Cuts against it: it gives no per-state count and no named Tamil Nadu item; method is "a range of keywords", no screened total; no view counts. It conflicts, in detail, with this project's own reading of one DMK reel (C2/C4 in facts.md: Meta's label sits in the "About this reel" menu); say "not labelled on the post" not "no label anywhere".

### C3. Which parties posted AI content (Tamil Nadu)
- Claim: both big camps and the Congress posted AI content, mostly their own promotion.
- Figure: The Federal (11 Mar 2026): DMK and TVK IT teams make short AI content "almost daily"; an AI video recreated Annadurai's voice endorsing Vijay; the DMK IT wing answered within two days with a video in Stalin's voice; AI Karunanidhi at the DMK's 75th anniversary; a DMK AI video showing Edappadi K. Palaniswami wanting to be a "slave" of the BJP (the Federal's wording); the Congress used the AI characters "Rakkamma" and "Pudhuvai Thatha"; the AIADMK ran AI campaigns. TVK used a hologram of Vijay in Kumbakonam (Khaleej Times). No count of videos.
- Source: https://thefederal.com/category/states/south/tamil-nadu/ai-political-campaign-tamil-nadu-2026-election-meme-war-233650
- READ AT SOURCE: yes.
- Cuts against it: no counts; the page does not say who made the "slave" video beyond calling it a DMK video (earlier file notes it as unattributed: say "The Federal reports").

### C4. The Annadurai AI video, Sept 2025
- Claim: TVK released a 2 min 32 s AI video of the late C. N. Annadurai criticising the DMK and endorsing Vijay; the DMK called it a "cheap political tactic".
- Source: https://newsarenaindia.com/politics/dmk-slams-cheap-tactic-after-ai-annadurai-endorses-vijay/56622 (updated 17 Sep 2025); DMK spokesperson A. Saravanan called it a deepfake and a cheap tactic, said voters "would see through it". The IJES paper (2026) also names the video.
- READ AT SOURCE: yes.
- Cuts against it (for the forecast): it was one party's own video, from a dead leader, released openly; no view count is given; it is a leader's likeness saying things he never said (the forecast's pattern) but a promotion, not an attack on a living rival. Only one outlet says "made by TVK" (earlier file).

### C5. A 2026 case: an altered voice of Stalin (an allegation only)
- [withheld: a police complaint or case about the maker of an AI clip; legal cases about makers are not published here]
- Source: https://www.dtnext.in/news/tamilnadu/ruling-party-distorting-stalins-talk-with-ai-video-dmk (updated 1 Nov 2026)
- READ AT SOURCE: yes.
- Cuts against it: the page reports no response from TVK or the account and no independent forensic finding; the headline says "ruling party" while the page does not say the account belongs to the party; no mention of the Election Commission. Whether it is the clip on the TVK headquarters wing's own Instagram is not confirmed (facts.md C3).

### C6. Other 2026 Tamil Nadu cases
- Claim: on 3 Apr 2026 the CPI(M) alleged an AI video, "Kandan Murugan Vote", used a voice resembling the deity Murugan to urge votes for the BJP's lotus symbol, and asked the chief electoral officer to act (complaint, not a finding).
- Source: https://www.dtnext.in/news/politics/2026-tn-elections-cpm-flags-poll-code-violations-seeks-action-on-ai-video-ips-posting
- READ AT SOURCE: yes.
- Cuts against it: the page reports no response from the Election Commission; maker not named.

### C7. What the 2026 fact-checks mostly found: conventional fakes
- Claim: Fact Crescendo's post-election review of Tamil Nadu fakes names no AI or deepfake item; the fakes were spliced footage, old videos with false dates, slowed playback, edited photos and look-alike news cards.
- Figure: about 15 itemised claims; an inflated "Stalin asset" claim "multiplied this figure by over 450"; a fabricated "DMK 160-200" projection predicted the opposite of the outcome (TVK 108 seats). Published 26 May 2026.
- Source: https://srilanka.factcrescendo.com/english/2026-tamil-nadu-assembly-election-regime-change-and-circulated-fake-news-an-analysis
- READ AT SOURCE: yes.
- Cuts against it: it is a single fact-checker's selection with no stated method or total; it does not say fakes were fewer than AI ones.

### C8. The result
- Claim: TVK was the single largest party in the 4 May 2026 count with 108 seats of 234 (Fact Crescendo; a search summary gives 107 and 34.92% of the vote), short of the 118 needed; Vijay was sworn in as Chief Minister on 10 May (Fact Crescendo).
- READ AT SOURCE: yes for 108 seats and the swearing-in date (Fact Crescendo); the vote share: no.
- Cuts against it: no study I found links the result to any AI clip. This is the one Tamil Nadu election where AI videos from both camps were documented, and the outcome favoured the party with the first AI-endorsement video, but a hologram, a new party, anti-incumbency and many other causes are in play; nothing read attributes the result to AI.

### C9. Assam 2026 (outside Tamil Nadu): the one measured campaign
- Claim: a civil-society report (Netherlands-based DAHRD) on the Assam campaign classified 432 Facebook and Instagram posts as likely AI-generated with 45.4 million views, and 31 deepfakes of the opposition's chief-ministerial candidate; it alleged 119 breaches of the Model Code with no action.
- Source: the IJES paper's account (Panchal, 2026) of DAHRD (2026) and The Federal (2026). The DAHRD report itself: not opened.
- READ AT SOURCE: no (second-hand; the paper itself says these "carry the evidentiary weight of that provenance").
- Cuts against it: an advocacy report; no check of its detection rubric; views are not vote changes. Do not use on screen until the report is read.

## D. Spread vs corrections (new items only; the 66 hour and 12 hour figures are held and not redone)
- D1. Bangladesh 2026: flagged AI videos "still reached viewers ... without any indication that they were fake" (Graphika, A10). Ireland: about 12 hours live before the platform acted, on a journalist's contact (A11). India: a debunked cheapfake of a Congress leader had 43,000 views and no Community Notes (Context, A4); Meta's labels often mis-applied: a Context analysis "found many fact-checked Facebook videos were mislabelled or had no label". These are single cases, not a measure of speed.
- D2. No source I read measures a deepfake's spread against its correction in India or Tamil Nadu. The Stalin-voice (C5) and Annadurai (C4) pages give no view counts or fact-check dates.

## E. Not found
- Rest of World's own India 2024 pages; Tattle and Internet Freedom Foundation reports on Tamil Nadu 2026 deepfakes; a DFRLab piece on Tamil Nadu 2026; a Meta or Google statement on Tamil Nadu 2026; ECI's own advisory text and the Gazette text; an ECI figure for Tamil Nadu alone; any count of AI videos in Tamil Nadu 2026. CETaS pages and the FCC were blocked (403). The Dan (2026) paper and the DGAP paper (1,536 articles, 62 AI cases) could not be opened.

## THE AI'S READING
- The record is strong that deepfakes and AI voice fakes of leaders are now routine in elections (Slovakia, India, Tamil Nadu, Ireland, Argentina, Moldova) and that Tamil Nadu's two big camps both post AI clips of leaders, one in a leader's voice (the Annadurai case) and one alleged of Stalin (C4, C5).
- It is weak, near empty, for "a deepfake campaign changed an election result": no source I read shows one; CETaS (through a news page), Meta and the Slovak review all say there is no proof, and Ireland and Moldova show fakes against candidates who won.
- It is not proof of safety: the experiments show a deepfake fools about 42-47% of viewers in one survey, about as many as plain text; no experiment tests a real vote, a Tamil sample, or repeated organised exposure.
- The risk the record supports best is the other side of trust: people doubt real clips, and a "that is fake" claim paid off for politicians against text scandals in a US survey (B3).
- Strongest honest wording: "In 2031 an organised, targeted deepfake of a rival leader could spread and mislead many viewers; that has happened in other elections; nobody has shown it changed a result, so my guess is that it could tilt a close contest, and the surer harm is that real clips are doubted."
- Never say: "deepfakes decided" Slovakia, "AI changed" Tamil Nadu 2026, or "studies show deepfakes are harmless".
