# c7. The closing beat: government, platforms, outside services (what is true on 19 Nov 2026)

The request's line: the GOVERNMENT will help by putting restrictions; PLATFORMS (Instagram, Twitter) do a FIRST LAYER of check and bot
assistance; OUTSIDE SERVICES like SynthID help. Method: 40-odd web searches, then the pages opened. Marks: READ = I opened the primary
source (or a primary file in the repo) and saw the words; READ-VIA-READER = opened through the fetch tool's page reader (it returns the
tool's paraphrase and short quotes, so exact wording should be re-checked at the page before a quote goes on screen); SECOND-HAND =
only a search summary or an outlet's report of the primary. Not opened by me: .env, .prod.vars, collector/state.
Clock: the TN election was 23 Apr 2026 (count 4 May, from a report); a TVK government is in office (The News Minute, 2 Aug 2026).

## A. GOVERNMENT (India first)

A1. IT Rules amendment on synthetic media. G.S.R. 120(E), notified 10 Feb 2026. READ the consolidated text in the repo
    (data/research/legal/it-rules-2026-02-10/it-rules-2021-updated-2026-02-10.txt, from meity.gov.in).
    - Definition, rule 2(1)(wa): audio, visual or audio-visual information "artificially or algorithmically created, generated, modified or
      altered" so that it "appears to be real, authentic or true" and depicts a person or event "indistinguishable from a natural person or
      real-world event". Excluded: routine or good-faith editing, colour, noise reduction, compression, documents, PDFs, teaching material.
      So cartoons and obvious fantasy are mostly outside; text is outside.
    - Rule 3(3)(a)(i): an intermediary that offers a tool to make such content must use "reasonable and appropriate technical measures" so the
      tool does not produce content that "falsely depicts or portrays a natural person or real-world event by misrepresenting, in a manner
      that is likely to deceive, such person's identity, voice, conduct, action, statement". (Also child abuse material, false documents, etc.)
    - Rule 3(3)(a)(ii),(b): other synthetic content must be "prominently labelled" (audio: "prominently prefixed audio disclosure") and "embedded
      with a permanent metadata or other appropriate technical provenance mechanisms, to the extent technically feasible, including a unique
      identifier" of the tool used; the platform "shall not enable the modification, suppression or removal of the label, permanent metadata".
    - Rule 4(1A), significant social media intermediaries: before display, require the user to declare whether content is synthetic; "deploy
      appropriate technical measures ... to verify the accuracy of such declaration"; label where declared or verified. If it "knowingly
      permitted, promoted, or failed to act" it "shall be deemed to have failed to exercise due diligence".
    - Clocks: rule 3(1)(d) takedown now "three hours" (was thirty-six) after a court order or a reasoned written intimation from an officer
      of the Government not below Joint Secretary; rule 3(2)(b) complaint by the person depicted about nudity/sexual acts/"impersonation in an
      electronic form, including artificially morphed images": "two hours" (was 24); grievance resolution 7 days (was 15).
    - Effective 20 Feb 2026: SECOND-HAND (ipleaders, via reader; TeamLease, Simpliance, S.S. Rana by search). The repo text shows only
      "Ins. by G.S.R. 120(E), dated 10.02.2026"; the commencement clause of the amending notification itself was not read.
    - Penalty: no fine in the text. The consequence is loss of Section 79 safe harbour (the platform becomes liable for user content): the text
      says "deemed to have failed to exercise due diligence"; the safe-harbour reading is from ipleaders (READ-VIA-READER) and rule 1(1B) READ.
    - What it does NOT do (reading of the text): no duty on the person who posts, except the declaration; no deadline for platforms to act on an
      ordinary viewer's report other than the 7-day grievance clock; the 3-hour clock needs a court or an officer; metadata duty is "to the
      extent technically feasible"; a tool that does not write the mark (open-source, made abroad) leaves nothing to read.
    - Draft follow-on (21 Apr 2026, a new rule 3(4) for binding MeitY directions): SECOND-HAND (a search summary); status unknown.

A2. Election Commission of India. Oct 2025 advisory (Bihar, but "all general and bye-elections"): label "AI-Generated, Digitally Enhanced,
    or Synthetic Content" on "at least 10 per cent of the visible display area" (audio: first 10% of duration); parties act within 3 hours of
    a report or detection; keep internal records; no misrepresenting a person's identity or voice without consent.
    READ-VIA-READER of All India Radio (newsonair.gov.in, 25 Oct 2025); ECI's own PDF not opened (earlier pass: connection refused).
    - PIB/ECI release 19 Apr 2026 (Release ID 2253528): READ-VIA-READER through India Today NE (PIB itself returned 403 to me; an earlier pass
      in the repo read it): "more than 11,000 social media posts and URLs have been identified and acted upon" since the 15 Mar announcement,
      actions = removal, FIRs, "clarifications and rebuttals"; platforms directed to act on misleading AI-generated or manipulated content
      "within three hours of it being flagged"; C-Vigil: 3,23,099 complaints, "96.01% resolved within the 100-minute timeframe" (all complaint
      kinds). No split by state or by AI content. Five states together (Assam, Kerala, TN, Puducherry, West Bengal).
    - The 3 hours: for parties it runs from "coming to their notice"; for platforms from being flagged. It binds parties and platforms in a
      campaign; it is a response clock, not a detection duty.

A3. Tamil Nadu. CEO Archana Patnaik, 20 Mar 2026 (READ-VIA-READER of devdiscourse; Khaleej Times 16 Apr 2026): "misleading or unlawful" AI
    content taken down from party handles within 3 hours of detection or report; label as above; political ads pre-certified by the MCMC
    (recognised parties apply 3 days ahead, others 7; processed within 48 h). No FIR, arrest or takedown count was reported in either article.
    - 3 Apr 2026: CPM complained to the CEO about an AI clip "Kandan Murugan Vote" (a voice like the deity urging the lotus symbol); the
      article (dtnext, READ-VIA-READER) reports no action.
    - [withheld: a police complaint or case about the maker of an AI clip; legal cases about makers are not published here]
    - State fact-check unit: made Oct 2023 by the DMK government (Iyan Karthikeyan). After the change of government it was split and
      renamed "Information Integrity Desk"; headed by Guru, founder of the Online Thalapathy Fans Club, with a TVK IT wing member as project
      lead (The News Minute, 2 Aug 2026, READ-VIA-READER). No tipline or public submission channel in that article. It does not have takedown
      power (earlier reports, SECOND-HAND).
    - Wording risk: the state's own fact-check body is staffed from the governing party's wing; do not present it as neutral help.

A4. Enforcement record (what is known).
    - Bihar Congress video of PM Modi and his late mother: posted 10 Sep 2025, 7.1 million views by 17 Sep, when the Patna High Court (Acting CJ
      P B Bajanthri) ordered it removed and told Meta, Google India and X to stop circulation, in a PIL (several outlets, SECOND-HAND; Livelaw,
      Scroll, newsonair in the search list, none opened). So "up 7 days" = 10 to 17 Sep, until a court acted: that is arithmetic from the
      reports, and it was BEFORE the Oct 2025 ECI advisory and the Feb 2026 Rules. Whether the platforms took copies down, and when, was not found.
    - The Quint (8 May 2026), READ-VIA-READER: "around 19 AI-generated videos" tied to the Assam, WB, Kerala, TN elections, Jan to Apr 2026,
      "None ... labelled as AI-generated" by poster or platform; some from official BJP, Congress and DMK handles; keyword sample, not a census;
      no Meta reply in the article.
    - No report found of MeitY or the ECI penalising a platform for missing a label under the new Rules. Absence in search, not a statement.
    - The repo's own readings (facts.md V11): the two TN reels carry their maker's AI mark on the frame.

A5. Other countries, one line each.
    - EU AI Act Article 50: READ the text via artificialintelligenceact.eu (reader): providers must mark outputs "in a machine-readable format"
      "detectable as artificially generated or manipulated"; deployers "shall disclose that the content has been artificially generated or
      manipulated". Applies from 2 Aug 2026; the machine-readable marking got a grace to 2 Dec 2026 for systems already on the market, under the
      Digital Omnibus (SECOND-HAND, law-firm notes; one note said not yet binding on 7 Jul 2026). Voluntary Code of Practice, June 2026 (SECOND-HAND).
    - China: CAC, MIIT, MPS, NRTA labelling measures plus standard GB 45438-2025, in force 1 Sep 2025: visible labels and metadata labels;
      platforms check for the hidden label and add a visible one (SECOND-HAND, law-firm summaries; one later note says enforcement is uneven).
    - Denmark: bill L 49 on a copyright-style right over one's face and voice. Euronews (READ-VIA-READER) 8 Oct 2026: a bill "not passed";
      minister hopes it takes effect at the start of 2027; satire and parody kept. Some outlets say "passed": they are wrong or early.

## B. PLATFORMS

B1. Instagram / Meta label. Meta, 5 Apr 2024 post (page updated 23 Oct 2025) and 19 Feb 2026 US-midterm post (both READ by the earlier
    pass, ledger items 10, 11): "AI info" is added from "industry-shared signals" (C2PA, IPTC) or because the poster "self-disclosed"; Meta
    requires disclosure for photorealistic video or realistic audio, "may apply penalties". 12 Sep 2024: for content only edited with AI the label
    moved into the post's menu. Meta said (2024) it "can't yet detect those signals" from other companies' video and audio. Where the label sits on a reel: not stated in
    any Meta text read; the repo's own check (facts.md C2) saw it under "..." > "About this reel". Meta Help Center page: opened, returned only a cookie
    wall, not read.
    - Known misses: The Quint (A4): 19 clips, none labelled. 2024: BOOM and Context found political deepfakes on Meta without the "Made with
      AI" label; a report said Meta approved 14 of 22 AI-manipulated test ads in India (The Wire via search, SECOND-HAND).
    - C2PA is stripped by many re-uploads: Washington Post test, none of eight platforms kept a Sora video's credentials (SECOND-HAND, via
      search summary). So the file signal usually does not reach the label.
B2. Instagram Community Notes: Meta's own wording (19 Feb 2026) says contributors are people "in the US". No India launch found (a search
    summary says it was delayed in India; undated; SECOND-HAND). Do not say Instagram has notes in India.
B3. Twitter Community Notes (say Twitter, as the request asked; the company's name is X).
    - Rule: a note shows only if rated helpful by people who "have sometimes disagreed" (READ by the earlier pass, GitHub docs); minimum 30
      minutes delay; at least 5 ratings. No typical time-to-show published there.
    - PNAS 2025, Slaughter, Peytavin, Ugander, Saveski: READ (arXiv v2 PDF downloaded, text searched). 40,078 posts, 16 Mar to 23 Jun 2023.
      "-46.1% reposts ... after being attached", but "11.6%" over the post's whole life; views -13.5% after, -5.5% lifetime. By when the note
      came: within 1 to 12 h total repost cut 24.9%; 12 to 23 h 12.3%; 23 to 47 h 4.3%; after 47 h 0.1%. Limits stated by the authors: only
      posts that had a note, 2023 data, the system changed since; they could not estimate how many misleading posts got a note at all.
    - Indian languages: CSOH report 30 Jun 2025 (READ-VIA-READER): of 1.85 million notes, "1,737 are written in South Asian languages",
      0.094%; fewer than 40% of South Asian drafts reach "helpful + bridging", about 65% for English. No Tamil figure on that page. A Trusted
      Information Alliance comment to the Oversight Board (Dec 2025) reportedly lists about 100 Tamil notes (0.028%): SECOND-HAND (the PDF
      would not parse). Tamil notes exist but are very few.
    - AI Note Writer API: launched 1 Jul 2025 as a pilot; AI-written notes still need human raters across viewpoints (SECOND-HAND, several
      outlets by search). Later an experiment with humans requesting an AI-drafted note (Engadget, 2026, SECOND-HAND). This is the real "bot
      assistance" on Twitter: bots draft, humans decide.
B4. Grok as "bot assistance". It is not a platform control; people tag it. Evidence:
    - DFRLab, 24 Jun 2025 (READ-VIA-READER): about 130,000 posts; Grok replied 312 times to 353 tags on one AI airport video; swung between
      "damaged" and "not"; before a Community Note, 31% of its replies said the AI video showed damage at Tel Aviv airport; called an AI
      building-collapse video real; DFRLab: it "struggles with verifying already-confirmed facts, analyzing fake visuals".
    - Tow Center (CJR), Mar 2025: Grok 3 answered 94% of citation queries wrongly (eWeek/CJR via search, SECOND-HAND). That test is on news-citation, not on fakes; do not use it as a deepfake figure.
    - Full Fact, 28 Aug 2026 (READ-VIA-READER): tested Grok 4, ChatGPT 5 and three Gemini models by API: 39 major errors, 67 wrong replies; Grok said a
      fake image of burning ambulances and a fake cost-of-living protest video were real; some models corrected after Full Fact published; the
      companies said the API test is not the consumer app (OpenAI, Google); X did not reply.
    - Maldita and University of Jaen, Oct 2026 (search summary, SECOND-HAND; the page was opened but was image data): 766 queries on 46 World
      Cup claims; with image or video, right in 121 of 245 (49.4%); verdict changed between replies in 21 of 40 tweets. Re-check at the page.
    - No India or Tamil evaluation of Grok's fact-checks was found.
B5. YouTube. Disclosure required for realistic altered content; automatic labels for YouTube's own AI tools, C2PA metadata, or "internal
    systems"; label on the video itself for election and news topics (YouTube Help and blog 18 Mar 2024; READ by the earlier pass, ledger 13, 14).
    Likeness detection: expanded 10 Mar 2026 to a pilot of government officials, political candidates, journalists (TechCrunch, READ-VIA-READER):
    face only, voice not yet; they may request removal, YouTube reviews it, parody and critique protected; names of testers not disclosed; no
    country list, no India detail. A report that it later opened to all adults is SECOND-HAND (exchange4media headline). A citizen cannot use it
    to check someone else's video; it protects those enrolled.
B6. WhatsApp. Forward limit: since Apr 2020 a message forwarded 5 or more times (double arrow) can go to one chat at a time; "India ...
    five chats at once" in 2018 (WhatsApp/Meta blogs, READ by the earlier pass); end-to-end encrypted, so WhatsApp cannot read content. Current Help
    Center not readable (script page). New: an "AI content" label for Channels, mainly set by channel admins, may be automatic, Android beta 2.26.34.5,
    31 Aug 2026 (WABetaInfo, READ-VIA-READER); Channels only, not private chats or groups; beta. No WhatsApp takedown of a deepfake in private chats is
    described anywhere I read. The Rules' traceability (rule 4(2)) exists for messaging apps but it is for police, not a viewer check.
B7. Indian-language gaps (what is shown): Twitter notes (B3); Meta's label rests on file signals, language-blind but only on signals; Meta
    fact-checking in India works with "nearly a dozen" fact-checkers (TIA, SECOND-HAND). Nothing found that shows Meta or Twitter label
    accuracy by language. Do not claim a Tamil-specific failure of the labels.

## C. OUTSIDE SERVICES

C1. SynthID. What it is: an invisible watermark that Google and a few partners write into their AI output.
    - Gemini app check: Google blog 18 Dec 2025 (READ): upload a video "and ask ... 'Was this generated using Google AI?'"; it scans "the
      imperceptible SynthID watermark across both the audio and visual tracks" and can say which seconds ("SynthID detected within the audio
      between 10-20 secs"); files "up to 100 MB and 90 seconds"; "available in all languages and countries supported by the Gemini app". Image
      check came first (Nov 2025). The blog says nothing about what it cannot detect or accuracy.
    - SynthID Detector portal: announced May 2025 for journalists; opened to everyone 7 Oct 2026 (MacRumors, Hans India READ-VIA-READER; PetaPixel,
      TechCrunch blocked/404; Google's own post not opened). Images, video, audio; sign-in; free; "worldwide in English". Detects Google,
      OpenAI, Nvidia and Kakao marks; Apple "coming". Google: it "cannot identify every piece of AI-generated content", "not a general AI
      detector", cannot tell AI-made from AI-edited. Google says its checks in Search, Gemini and Chrome handle over a million requests a day
      (Hans India). ppc.land (SECOND-HAND) adds: results are only "Watermark detected" / "not detected", no accuracy figure published, daily
      quota undisclosed, files kept as a signature for 24 h. ElevenLabs as an adopter (May 2026): SECOND-HAND, unclear; matters for Tamil voice clones, so not to be said.
    - What a layperson can do: forward a clip to the Gemini app, ask in Tamil or English (Gemini supports Tamil: general knowledge, not tested
      here). A "detected" answer is strong; "not detected" is nothing.
C2. C2PA Content Credentials. Pixel 10 (Google blog 10 Sep 2025, READ-VIA-READER): Pixel Camera attaches credentials to every JPEG, Assurance
    Level 2 "the highest ... currently defined"; Google Photos shows them in "About"; Google: they "are not the sole solution". C2PA 2.2 says a
    manifest can be "removed from an asset" (repo, F27). Platforms often strip them on upload (B1). In India no platform or phone app readable by
    an ordinary Tamil user shows credentials on a WhatsApp-forwarded clip; Adobe's Verify site and Google Photos' About panel are the checks that exist
    (Adobe: from general knowledge, not opened). So absence of credentials proves nothing; presence on an authentic camera file helps the maker prove it is real.
C3. Deepfakes Analysis Unit (Misinformation Combat Alliance, now at dau.trustedinfoalliance.in; the old dau.mcaindia.in redirects 301). READ
    (reader): the homepage lists reports up to 5 Jun 2026 (Jaishankar investment scam clip), then 21 May, 24 Mar, 9 Mar, 12 Feb, 23 Jan, 13 Jan,
    2 Jan 2026; button bit.ly/dauwatipline (earlier pass: redirects to wa.me/919999025044); no languages, no turnaround on the page. English, Hindi, Tamil,
    Telugu comes from the Mar 2024 launch coverage (SECOND-HAND). Roughly one report a month in 2026, mostly synthetic audio of national
    figures (army chief, ministers, Tharoor); none about a Tamil Nadu leader seen. Whether the tipline answers today: NOT tested.
    So: it exists, reports are weeks apart, it is not a fast checker, and a Tamil clip may never be picked.
C4. Tamil fact-checkers: Factly tipline listed with English only (+91 92470 52470), BOOM's tipline English/Hindi (+91 77009 06588), Fact Crescendo
    lists Tamil among nine languages (+91 90490 53770) (a search summary of old listings, SECOND-HAND, numbers not verified, none tested).
    Tamil sites exist (Fact Crescendo Tamil, BOOM Tamil). No Tamil WhatsApp tipline confirmed by me. YouTurn (Iyan Karthikeyan's, earlier) was not examined.
C5. Detector tools. Deepfake-Eval-2024 (Chandra et al., arXiv 2503.02857, READ abstract): 45 hours of video, 56.5 hours of audio, 1,975 images
    collected in 2024 from social media and detection-platform users; open-source detectors' AUC fell "by 50% for video, 48% for audio, and 45%
    for image models compared to previous benchmarks"; commercial and fine-tuned models do better but "do not yet reach the accuracy of
    deepfake forensic analysts". I read only the abstract; per-tool numbers not read. Not a Tamil test.
    Free detector sites (many) are not evaluated here; no Tamil-audio test found.

## D. WHAT A TAMIL NADU PHONE USER CAN ACTUALLY DO TODAY
1. Ask: who posted, is it on the leader's own channel? (the repo's controls: official channels exist, but see 8b).
2. Gemini app: upload (100 MB, 90 s) and ask whether Google AI made it. Works only for SynthID marks.
3. Instagram: "..." menu > About this reel; look for the AI label. Twitter: look for a note (Tamil notes are rare).
4. DAU tipline on WhatsApp (+91 9999025044): exists, reply time unknown.
5. Fact-check sites in Tamil (Fact Crescendo Tamil, BOOM Tamil); no confirmed Tamil WhatsApp line.
6. The ECI's C-Vigil app for election-time complaints (it exists; results in A2 are not split by AI).
7. WhatsApp: a double-arrow forward is a prompt to stop; there is no deepfake check inside WhatsApp.

## E. WHAT MAY BE SAID AS FACT
- India's IT Rules, amended 10 Feb 2026 (in force 20 Feb, outlets), tell platforms to label realistic AI audio and video, ask big platforms to have
  users declare AI content and to check the declaration, and cut the takedown time on a court or government order to 3 hours from 36.
- The consequence for a platform that fails is losing its legal protection for what users post (safe harbour), not a fixed fine.
- The Election Commission told parties (Oct 2025) to label AI campaign material (10% of the screen) and to remove flagged fakes within 3 hours;
  the Tamil Nadu CEO repeated it on 20 Mar 2026; the ECI said 11,000+ posts across five states were acted on by 19 Apr.
- The Quint found about 19 AI videos around four state elections with no label from poster or platform, some from official party handles.
- Instagram's AI label comes from hidden file signals or the poster's own say; for other companies' video and audio Meta said it could not yet read them.
- A Twitter note, when it comes, cut reposts: 46% after it appeared; 24.9% of a post's total reposts if it came within 12 hours; 0.1% after 47 hours.
- On Twitter, Tamil-language notes are very few (South Asian languages: 0.094% of 1.85 million notes).
- Grok, asked to check AI fakes, got it wrong in published tests (DFRLab 2025; Full Fact 2026).
- SynthID is a watermark; the Gemini app (Dec 2025) and the SynthID Detector (public 7 Oct 2026) find it, for Google, OpenAI, Nvidia, Kakao output.
- A real DAU tipline exists with reports in 2026 on synthetic audio of national figures.
- Detectors lose about half their score on real-world fakes (Deepfake-Eval-2024).

## F. MUST BE SAID AS A LIMIT
- A missing SynthID mark proves nothing: it is found only on output from firms that adopted it; Google says it is not a general AI detector and cannot tell AI-made from AI-edited.
- A missing Instagram label proves nothing: it depends on file signals or the poster's word, and re-uploads strip signals.
- The 3-hour clock in the Rules starts from a court order or an officer’s written intimation, not from a viewer’s casual report (the Election Commission advisory is reported to speak of three hours after a flag); the ordinary grievance clock is 7 days.
- The Rules cannot reach WhatsApp private chats and groups (end-to-end encrypted; the forward limit slows, does not remove), a clip made with a foreign or open-source tool that writes no mark, or a maker who does not declare.
- The Rules label realistic content; satire or cartoon use may fall outside the definition.
- No record found of any platform penalised under the new labelling rules; the Bihar case needed a High Court, a week after posting.
- The parties are themselves the main posters; the state fact-check unit is staffed from the governing party's wing.
- Instagram Community Notes is for US contributors; Twitter's Tamil notes are rare; the PNAS data are 2023.
- DAU: no stated turnaround; no Tamil-specific check seen; not tested by me.
- Gemini's check: 100 MB, 90 s; English-only portal.
- Detector accuracy: not Tamil-tested; "AI detector says fake" is not proof either way.

## G. WORDING RISKS
- Do not say "the government bans deepfakes": the Rules label and speed takedown; they do not ban realistic synthetic content that is labelled (rule 3(3)(a)(i) bars the deceptive impersonation kind at the tool level).
- Do not say "3-hour takedown" without "on a court or government order" (rule) or "when flagged to a party/platform" (ECI).
- Do not say "penalty/fine": the stated consequence is losing safe harbour.
- Do not say "Instagram detects AI": it reads hidden signals or takes the poster's word.
- Do not say "X" for the platform (the request's wish); the studies name it X or Twitter.
- Do not say "Instagram has Community Notes" in India or "Grok fact-checks it" as help. "Bot assistance" is true only of the AI Note Writer pilot (bots draft, humans rate).
- Do not give "94% wrong" for Grok as a deepfake figure (it is a news-citation test, 2025, Grok 3).
- Do not say SynthID "detects deepfakes": say "finds Google's and a few partners' watermark".
- Do not say "the DAU answers in minutes" or "in Tamil" as fact; Tamil comes from 2024 launch coverage.
- Do not say Denmark "has passed" its law; the EU's labelling "applies from 2 Aug 2026" but marking has a grace to 2 Dec 2026 (secondary).
- Do not say the Bihar video "stayed up 7 days despite the rules": it was Sep 2025, before the advisory and the Rules.
- The request's three-layer line is true as intent and partly as practice: government (rules, no record of enforcement on a platform), platforms (a first layer that misses), outside services (only where a mark exists); say each with its limit.
