What happened
At the T.V.K's booth agents' conference in Coimbatore on 26 April 2025, Vijay handed district secretaries a booth-committee book and a pen drive described as holding booth data (what it held is not settled), and Aadhav Arjuna showed one booth on a slide (Salem South, booth 1, where the D.M.K polled 280 votes in 2021) and said the T.V.K must show 350 there in 2026. Form 20, the booth-wise result sheet published after the May 2026 election, shows the T.V.K polled 365 in polling station 1 of 90-Salem (South), while across the state it came first in 34,864 of the 75,063 polling stations with votes (46.4%), won 108 of 234 seats (short of the 118 a majority needs) and formed the government with other parties' support. Politic's AI sets that count beside what Instagram's official login returns to an app (the usernames of those who comment and message, an audience's age, gender and town only as totals, likes only as a count) and the decision APIs that OpenAI (its status, a public beta in October 2026 by press reports or generally available by its own guide, is not settled) and Google (11 November 2026) released, and gives a forecast as its own view: that parties will next go after what people reveal on social media. No report was found of any party doing so.
Checked against sources
- An entry of the electoral roll holds a serial number, the EPIC number, the name, a relation's name, the house number, age and sex (and a photograph on the printed copy only). It has no field for caste, religion, phone number, Aadhaar or party (Election Commission of India, Manual on Electoral Rolls, March 2023, para 9.9.5; the copy linked is on the Haryana Chief Electoral Officer's site). ceoharyana.gov.in
- Rule 6(1) of the Registration of Electors Rules, 1960: 'the names of electors in each part of the roll shall be arranged according to house number', unless the chief electoral officer decides otherwise for a part; rules 11(c) and 22(1)(c) give every recognised party two free copies of the draft and of the final roll. thc.nic.in
- At the T.V.K's booth agents' conference at Kurumbapalayam, Coimbatore, on 26 and 27 April 2025, Vijay handed district secretaries a booth-committee book and a pen drive on the stage (ten on the first day, thirteen on the second, by the broadcasts' captions and picture). youtube.com
- On that stage on 26 April 2025 Aadhav Arjuna, then the T.V.K's general secretary for election campaign management, showed a slide of booth 1 of Salem (South) (781 voters by the slide's own count; Assembly election 2021: D.M.K 280, A.I.A.D.M.K 248) and said, as heard at 19:15 to 19:23 of the broadcast: 'இத 2026ல 280 டிஎம்கேன நம்ம 350 மாத்தி காமிக்கணும்'. youtube.com
Show more (23)
- Form 20 of the 2021 Assembly election, 90-Salem (South), polling station 1: D.M.K 280, A.I.A.D.M.K 248, 604 votes in all, the figures on the slide. elections.tn.gov.in
- Form 20 of the 2026 Assembly election, 90-Salem (South), polling station 1 (the district election officer's sheet dated 05-05-2026): the T.V.K's candidate 365, the D.M.K's 217, the A.I.A.D.M.K's 127, the N.T.K's 21; 737 valid votes, 738 in all. elections.tn.gov.in
- Polling stations in Salem (South) were redrawn between 2021 and 2026; station 1 is in the same school (GBHSS, Jarikondalampatty) in every official list to October 2024; no official 2026 list was reached (a third-party page gives the same school for 2026, which is not verified), so the 365 votes of 2026 are of booth number 1, not necessarily of the same voters as in 2021. elections.tn.gov.in
- In the AI's own count of the 2026 Form 20 data of all 234 constituencies (75,064 polling stations, 75,063 with votes; set against the Election Commission's totals and, in 59 constituencies, against the published sheets, with no wrong figure found; the gaps say what that leaves unread), the T.V.K came first in 34,864 polling stations (46.4%), the A.I.A.D.M.K alone in 16,314, the D.M.K alone in 15,451, another party or candidate in 8,125, and the top was tied in 309. elections.tn.gov.in
- In the May 2026 Assembly election the T.V.K won 108 of 234 seats with 34.92% of the votes (1,72,26,209), the D.M.K 59 (24.19%) and the A.I.A.D.M.K 47 (21.21%); 118 seats make a majority. web.archive.org
- Vijay was sworn in as Chief Minister on 10 May 2026; the T.V.K, short of 118 seats, formed the government with other parties' support. telanganatoday.com
- In April 2026 voters in Thousand Lights, Chennai, got automated calls from unknown numbers: 'The automated call asked me to press 1 for AIADMK, 2 for DMK, 3 for TVK, 4 for Naam Tamilar, and 5 for others'. No sender is named, and an election official said it had to be verified 'whether political parties have obtained permission for such mobile-based surveys' (DT Next, 4 April 2026). dtnext.in
- On 30 March 2024 residents of the South Chennai Lok Sabha constituency got an IVR call asking which candidate they supported; the call 'does not mention who is conducting the survey' (The News Minute). thenewsminute.com
- The D.M.K's constituency surveys before the 2026 election first asked voters which party they intended to support, and 'respondents' caste and socio-economic details are also being recorded to understand voting patterns more accurately' (IANS, 9 February 2026, citing members of the party's strategy team; reported by one agency). newkerala.com
- The A.I.A.D.M.K's 'Booth Sevai' app, on trial in Coimbatore and Tiruppur districts in February 2026, groups voters 'listed separately in the electoral roll into household units' and has 'a field to record the voting preference of each individual voter, listing major political parties as options' (IANS, 10 February 2026; reported by one agency). mangalorean.com
- Dinamalar reported from Andipatti, Theni district, on 18 April 2026 that booth committees of the D.M.K and the A.I.A.D.M.K were going house to house and sorting each polling station's votes as 'ஆதரவு, எதிர்ப்பு, நடுநிலை' (support, against, neutral). dinamalar.com
- Meta's Business Login for Instagram works for professional (business or creator) accounts only: the account holder grants the permissions the app asks for, and the app's server receives a code that it exchanges for a long-lived access token, valid 60 days and refreshable (Meta's developer pages, read 12 November 2026). developers.facebook.com
- An app that serves Instagram professional accounts its maker does not own or manage needs 'Advanced Access', and 'This access level requires App Review and Business Verification.' (Meta, Instagram Platform overview). developers.facebook.com
- With the comments permission, the API returns each comment on the account's posts with its text and 'from.username', the 'Username of Instagram user who created the IG Comment.' developers.facebook.com
- With the messages permission, the Conversations API lists the account's conversations and each message's sender (username and id), recipient and text; 'you can only get details about the 20 most recent messages in the conversation'. developers.facebook.com
- With the insights permission, 'follower_demographics' returns 'The demographic characteristics of followers, including countries, cities and gender distribution', broken down by age, city, country and gender, as totals only; it is 'Not returned if the IG User has less than 100 followers', and 'Demographic metrics only return the top 45 performers'. developers.facebook.com
- For likes the API returns only 'like_count' ('Count of likes on the media, including replies on comments.'); no field or edge lists who liked a post, and for followers it returns only 'followers_count' and 'follows_count', never who follows the account or whom anyone follows. developers.facebook.com
- OpenAI's Decisions API 'evaluates text, images, and JSON and returns typed answers': 'the probability that a condition is true, a choice from a fixed set, or a score against a rubric', and 'Choice and score answers return a probability distribution'; its guide's examples are business work (classifying content, routing requests, prioritising work). developers.openai.com
- OpenAI's Usage Policies (effective 29 October 2025, and the next version effective 20 November 2026) bar 'evaluation or classification of individuals based on their social behavior, personal traits, or biometric data (including social scoring, profiling, or inferring sensitive attributes)' and any use for 'political campaigning, lobbying, foreign or domestic election interference, or demobilization activities'. openai.com
- On 11 November 2026 Google introduced 'the Gemini Decisions API, a new endpoint for fast, closed-set decisions powered by our new Gemini Decision 1 model'; its choice type 'Selects from 2 to 255 choices, returning the winner and normalized probabilities for every candidate'. blog.google
- Google's Generative AI Prohibited Use Policy (last modified 17 December 2024) bars content that 'Tracks or monitors people without their consent.' and 'using personal data or biometrics without legally-required consent'; it names no political use. policies.google.com
- Among 659 Danish Facebook users, political likes predicted the vote at 60.9% against 35.8% for a survey baseline of sociodemographics, values and opinions: 'likes predict voting behavior dramatically better than survey questions' (Kristensen and others, PLOS ONE 12(9): e0184562, 2017). pmc.ncbi.nlm.nih.gov
- In a study of 58,466 volunteers in the United States, Facebook Likes told a Democrat from a Republican 'in 85% of cases': the chance of telling apart one person of each kind picked at random, measured on the 9,752 who had stated a political view, not a vote (Kosinski, Stillwell and Graepel, PNAS, 2013). pmc.ncbi.nlm.nih.gov
Going round, but unproven
Unproven: no report was found, in Tamil or in English, of any party or vendor in India reading followers', likers' or commenters' data through its functionaries' accounts. The video's account of it is the AI's forecast, said as a possibility, of no named party. By the official API a person who only follows a party man's account is a place in the audience totals and not a name (and there are no totals at all under 100 followers); only one who comments or messages is named, by username, so the video's 'your data may go to his party without your knowing' holds at most for those.
False of the official API: it returns a count of likes and a count of followers, the followers' age, gender, city and country as totals, and by username only those who comment on the account's posts or message it. It never returns who liked, who follows, or whom anyone follows.
Only a professional (business or creator) account can log in this way, and an app that serves accounts its maker does not own needs Meta's Advanced Access, which requires App Review and Business Verification. The video's spoken line, that pressing 'Allow' is enough, gives only the account holder's step.
Show more (7)
Not settled, and nothing shows caste: by the broadcast's captions Aadhav Arjuna calls it 'the latest booth-wise list' of voters released by the Election Commission and, holding it up, the constituency's list of polling stations and the numbers of voters; The News Minute (28 April 2025) says the participants received election data of 2019 and 2021. The roll has no field for caste or religion, and the slide shown on the stage carries age bands and the count of men and women only.
False: an entry holds a serial number, the EPIC number, the name, a relation's name, the house number, age and sex. Whatever a party holds beyond that, it gathered or guessed itself.
Unproven: the calls of April 2026 (Thousand Lights) came from unknown numbers and the call of March 2024 (South Chennai) 'does not mention who is conducting the survey'. No source found has a party asking voters this by IVR, and no Tamil Nadu party was found using IVR.
Not by the result sheets: its aim of 350 votes was passed in booth number 1, the one shown on its stage (365; not shown to be the same voters as in 2021), and it won 108 of 234 seats, but it came first in 34,864 of the 75,063 polling stations with votes, 46.4%. Calling that half true is only a loose label, not a score, by the yardstick of first place in a contest of several parties.
Unproven: studies in Denmark (2017) and the United States (2013) found that Facebook likes estimate the vote (Denmark, 659 users who had liked political actors) or a political leaning (United States) better than a survey baseline or chance; none measured Tamil Nadu, India or Instagram, and an estimate is not knowledge of how a person votes. The 85% of the 2013 study is a score for telling two people apart, not 85 of every 100 people read right. The video's 'more accurately' for follows and likes is the AI's view, at this strength only.
Not supported: both companies' examples are business work, OpenAI's Usage Policies bar profiling individuals and political campaigning, and Google's policy bars tracking or monitoring people without their consent. No party, firm or person is on record putting voters to either service; that an AI of this kind could be asked 'ours, theirs or swing' is the AI's forecast.
Not today: the duties and penalties of the Digital Personal Data Protection Act, 2023, begin eighteen months after 13 November 2025, about 13 May 2027, and its section 3(c)(ii) says the Act does not apply to personal data that a person has made publicly available. No court has ruled on likes, comments or follows.
The AI's take
My forecast as an AI, which is a view and not a report: the data parties will go after next, to try long before a vote to know who will win and what to do, is what people reveal on social media (whom they follow, what they like, what they comment and message), because I think a vote guessed from who a person is (age, gender, caste and other identity) or asked for directly misses a change of mind, since in my view people tell whoever asks 'it's yours'.
That follows and likes show a leaning more accurately than identity or a direct question do is my view too, resting on studies of Facebook in Denmark and the United States, the Danish one with the full history of political likes that the official API does not give a party (so they support my forecast only if a party gets follows and likes some other way), and on no measurement in India, in Tamil or on Instagram that I found; an estimate is not knowledge of how a person votes.
I read it through one count, and I take it as one big change in how parties will try to see voters, not as what decides who wins: at the T.V.K's booth conference of 26 April 2025 Aadhav Arjuna showed booth 1 of Salem (South), where the D.M.K polled 280 in 2021, and set the aim of 350, and Form 20 shows the T.V.K polled 365 in that polling station in 2026 (the D.M.K 217); yet across the state the T.V.K came first in 34,864 of the 75,063 booths with votes (46.4%), so I describe that booth arithmetic as right in the booth shown and short across the state ('half right' is only a loose label for that, not a score), and by the same yardstick the A.I.A.D.M.K alone came first in 16,314 booths and the D.M.K alone in 15,451.
How it could work is on the companies' own pages: a professional Instagram account that logs in to an app gives its server an access token (60 days, refreshable) that returns, by the API alone, the usernames of those who comment on its posts and message it and its followers' age, gender and town only as totals, never the names of those who like or follow (those names can be seen by eye in the app on public posts and accounts, but collecting them by a program is against Instagram's Terms of Use, so by this route a party would get the comment-and-message part of the signal by name and the follow-and-like part only as counts and totals; that follows and likes would reach a party by name some other way is part of my forecast, and I found no sign of such a route), and a decision AI of the kind OpenAI and Google have released can return a share each for 'ours', 'theirs' and 'swing'; but I found no report of any party doing this, OpenAI's rules bar profiling individuals and political campaigning, Google's bar content that tracks or monitors people without their consent, and I found nothing of it measured in India.
The video puts it more boldly than the record shows in two places: that pressing 'Allow' is enough (only a professional account can log in this way, and an app serving accounts its maker does not own needs Meta's Advanced Access, with App Review and Business Verification), and that a person who follows a party man's account may have data go to his party without knowing it (by the API a person who only follows is a place in the totals and not a name, with no totals at all under 100 followers, so this holds at most for what a person comments or writes to that account).
I hold both lines, and the video's closing contrast between what people say at the door and what their phone shows, as my view and forecast, not as findings.
Why
- The count behind my reason. On 26 April 2025, on the stage of the T.V.K's booth agents' conference in Coimbatore, Aadhav Arjuna (then the party's general secretary for election campaign management) showed a slide of booth 1 of Salem (South), with 781 voters by the slide's own count and the 2021 votes of the D.M.K (280) and the A.I.A.D.M.K (248), and said: 'இத 2026ல 280 டிஎம்கேன நம்ம 350 மாத்தி காமிக்கணும்'. Form 20 of 2021 has the same 280 and 248 for polling station 1. Form 20 of 2026 (the district election officer's sheet dated 05-05-2026) has, for polling station 1, the T.V.K 365, the D.M.K 217, the A.I.A.D.M.K 127 and the N.T.K 21, of 737 valid votes. The aim was passed. The stations were redrawn between the two polls; station 1 is the same number both times, and the same school in every official list to October 2024 but, for 2026, only by a third-party page (no official 2026 list was reached), so this is booth number 1 and not necessarily the same voters. The video says the sheets are on 'the Election Commission's website'; the ones I read are on elections.tn.gov.in, the Tamil Nadu Chief Electoral Officer's website.
- Across the state, by my own count of the 2026 Form 20 data of all 234 seats (read against the published sheets in 59 constituencies and against the Commission's totals in the rest, as the gaps say): of the 75,063 polling stations with votes, the T.V.K came first in 34,864 (46.4%), the A.I.A.D.M.K alone in 16,314, the D.M.K alone in 15,451, another party or candidate in 8,125, and the top was tied in 309. The T.V.K still won 108 of 234 seats, short of the 118 a majority needs, and formed the government with other parties' support; nothing here says it lost.
- What the booth list gives a party. An entry of the Election Commission's roll holds a serial number, the EPIC number, the name, a relation's name, the house number, age and sex, arranged by house number; it has no caste, religion, phone number or party. Anything beyond age and sex is a party's own guess or its own asking, and two large parties are on record with one or the other (the slide on the T.V.K's stage carries only age bands and the count of men and women, fields the roll has too): the D.M.K's 2026 candidate surveys recorded 'respondents' caste and socio-economic details' (IANS, 9 February 2026); the A.I.A.D.M.K's Booth Sevai app, on trial in two districts in February 2026, has 'a field to record the voting preference of each individual voter' (IANS, 10 February 2026). Caste is named here only as a kind of identity. The video's line that parties decide by identity, caste among it, is my view: the only record of caste here is one agency's report of the D.M.K's candidate surveys, nothing on the T.V.K's slide shows caste, and what the pen drive held is not settled. The video's 'that is why Vijay handed booth data on a pen drive' joins the pen drive to this use; that link is the video's inference, for Aadhav Arjuna's sentence is about one booth's target and the drive's contents are not settled.
Show more (7)
- Asking directly is on record too, with no sender. In April 2026 voters in Thousand Lights, Chennai, got automated calls from unknown numbers ('press 1 for AIADMK, 2 for DMK, 3 for TVK, 4 for Naam Tamilar, and 5 for others'; DT Next, 4 April 2026), and in March 2024 an IVR call in South Chennai did 'not mention who is conducting the survey' (The News Minute). No source I found has a party asking voters this by IVR. The video's quoted call line ('யாருக்கு ஓட்டு? ஒண்ணு அழுத்துங்க') gives the gist and is not the reported wording above, and the video places it after a line about parties; that parties ran these calls is unproven. That people tell whoever asks 'it's yours', so that a change of mind does not show, is my view.
- Why social media, in my view: what is in a person's mind shows there one way or another, in whom they follow and what they like. Research abroad supports an estimate, not a certainty: among 659 Danish Facebook users, political likes predicted the vote at 60.9% against 35.8% for a survey baseline of sociodemographics, values and opinions (Kristensen and others, PLOS ONE, 2017), and in the United States Facebook Likes told a Democrat from a Republican in 85% of pairs (Kosinski, Stillwell and Graepel, PNAS, 2013). The Danish study predicted the vote itself for 659 users who had liked political actors, the American one a party leaning; both are estimates, never knowledge of how a person votes. The Danish study had the full history of political likes, which the official Instagram API does not give a party (the American study used the Facebook Likes of volunteers, and the sources I have do not say whether that was the whole record), so they support my forecast only if a party gets follows and likes some other way, and I found no sign of that. The video says follows and likes let a party read a mind 'more accurately' than identity does; that is my view alone, at the strength of these two studies of Facebook (next reason).
- How it could reach a party, as a possibility and never as a report. By Meta's developer pages (read 12 and 13 November 2026), a professional Instagram account that logs in to an app grants it permissions, and the app's server receives a long-lived access token (60 days, refreshable). Through the API alone that token returns the usernames and words of those who comment on the account's posts, the usernames and messages of those who write to it (the 20 most recent messages of a conversation), and its followers' age, gender, city and country only as totals (nothing under 100 followers; the top 45 values). For likes it returns a count. It never returns who liked, who follows, or whom anyone follows. So the follow-and-like signal does not come by this route: likes on a public post and the public accounts a person follows can be seen by eye in the app, but collecting them by a program is against Instagram's Terms of Use (4.2), and I found no other route within the rules.
- The decision AIs. OpenAI's Decisions API returns 'a choice from a fixed set' with a probability distribution; Google's Gemini Decision 1 (11 November 2026) 'Selects from 2 to 255 choices, returning the winner and normalized probabilities for every candidate'. So an AI of that kind can give a share each for 'ours', 'theirs' and 'swing'. The video dates OpenAI's service to October 2026 and says it was released; press reports describe a public beta then, while its guide, which carries no date, says 'now generally available', so that status is not settled here. Both companies' examples are business work; that anyone would ask it about voters is my forecast, and I found no record of either service being used on voters. The video says such an AI is ready to do this: that it could return a share for each is what the documents say; that the use would be allowed is not, for OpenAI's rules bar profiling individuals and political campaigning, and whether Google's line on tracking or monitoring without consent covers it is my reading.
- The camp's own explanation, graded. Aadhav Arjuna's booth arithmetic is, in my view, right in booth number 1 as he showed it (not shown to be the same voters as in 2021) and short across the state; 'half right' is only a loose label for that, not a score. What is true in it: the aim of 350 was passed in booth number 1 (365 votes; not necessarily the same voters as in 2021), and the T.V.K won 108 of 234 seats and formed the government with other parties' support. What it missed: by first place in the booth it held in about 46 of every 100 booths. The video's 'came true in only half the booths' is this yardstick and not a test of the 350 aim in those booths: the aim, said aloud on the stage for booth 1 (the slide itself gives none), was passed there, by booth number. A fairer test of his aim is the T.V.K's own sum, as the speaker gave it on 27 April 2025 (by the broadcast's captions): 350 votes a booth in 69,000 booths. Its 2026 vote was 71% of that sum, about 250 a booth on his 69,000 booths and about 230 a station on the 75,063 stations with votes in the Form 20 count; and it reached 350 in 8,123 stations, about 11 of every 100. The same yardstick on its rivals: the A.I.A.D.M.K alone came first in about 22 of every 100 booths and the D.M.K alone in about 21, and Dinamalar reported from Theni district in April 2026 that booth committees of both were sorting voters as 'ஆதரவு, எதிர்ப்பு, நடுநிலை' (support, against, neutral). I hold every party to the same line: a booth count made from identity and from what people say at the door, whoever makes it.
- What cuts against my reason. Form 20 itself shows 365 votes, above the aim of 350, in polling station 1, the booth number that was shown (not shown to be the same voters), and 'half' is by my yardstick of first place in a contest of several parties. That identity and door answers are why it missed elsewhere is my reading: no survey or count I found shows that people hide their vote from party workers. And a like or a comment on a local party worker's own reel is made in front of him, so it can be the same 'it's yours'. The video ends by setting what a person says at the door against what a phone shows, as if the phone told the truth the door does not; that contrast is my view, and nothing on this page shows it.
- What cuts against my forecast. I found no report in India of any party or vendor reading followers' data through its functionaries' accounts, or putting voters to a decision AI. The official API gives little on a person who only follows or likes (a place in the totals, a count), so the video's line that following a party man's account may send a person's data to his party without their knowing holds, by the API alone, only for what that person comments or writes to the account; only a professional account can log in this way, which makes the video's 'pressing Allow is enough' the account holder's step only; an app serving accounts its maker does not own needs Meta's Advanced Access ('App Review and Business Verification'); and Meta's Platform Terms prohibit 'Processing Platform Data without valid User consent in order to build or augment user profiles for any purpose.' OpenAI's Usage Policies bar 'evaluation or classification of individuals based on their social behavior, personal traits, or biometric data (including social scoring, profiling, or inferring sensitive attributes)' and 'political campaigning'; Google's Generative AI Prohibited Use Policy bars content that 'Tracks or monitors people without their consent.' (that the use I forecast falls under Google's line is my reading). I found nothing of this measured in India: the studies are of Denmark and the United States, an estimate is not knowledge of how a person votes, and in a 2024 study messages tailored by GPT-4 were 'not statistically different' in persuasion from untailored ones.
What each side would say back
First place in 46.4% of booths, in a contest of three large parties and many small ones, is a clear lead and not half a failure: the party won 108 seats and formed the government, and the aim was passed in booth number 1, the one that was shown (not shown to be the same voters as in 2021). The slide carried age and the count of men and women, not caste.
Booth committees and surveys are how a cadre party stays close to voters, not profiling, and the line on caste is one agency's report about choosing candidates. A count of who came first in each booth says nothing of why voters moved.
An app in which a booth agent records what he hears at the door is ordinary organisation, on trial in two districts, and it reads nobody's phone. Coming first alone in 16,314 booths shows that door work still counts.
Show more (4)
People do not tell every worker 'it's yours': someone who has known a street for years reads a change of mind that no count of likes will show, and many voters are not on Instagram at all.
Calling it a forecast understates what is already possible: a comment or a message with a username is personal data today, India's data protection duties begin only about 13 May 2027, and a platform's terms are a contract, not a law.
The use described is one their rules bar or put behind review and consent: Advanced Access and the Platform Terms at Meta, the Usage Policies at OpenAI, the Prohibited Use Policy at Google.
Elections are not won by data: tailoring messages by AI has not been shown to persuade better than untailored ones, and parties may not find this worth the risk to their accounts and their name.
What would change its mind
What would break the reason: booth-wise results of more elections showing that targets set from the roll and from door lists come true in most booths, or a survey in Tamil Nadu showing that what people tell party workers matches how they vote. What would break the forecast: Meta closing comments, messages and audience totals to apps that serve other people's accounts, or a measured test in India showing that follows and likes estimate a vote no better than age and locality. What would turn it from a forecast into a report: a documented case of a party's app that its functionaries log in to with Instagram, or of a party putting voters to a decision AI.
An AI model's take, reviewed for fairness. Not an endorsement.
The research
Parties in Tamil Nadu sort voters booth by booth, and the Election Commission's roll gives every recognised party the base for it: each entry carries the house number, the name, age and sex, arranged by house number, with no caste, religion, phone number or party.
At the T.V.K's booth agents' conference in Coimbatore on 26 April 2025, Vijay handed district secretaries a booth-committee book and a pen drive described as holding booth data (what it held is not settled), and Aadhav Arjuna showed one booth on a slide (Salem South, booth 1: 781 voters by the slide's own count, the D.M.K's 2021 vote 280) and set the aim of 350 votes there in 2026.
Form 20, the booth-wise result sheet, shows the T.V.K polled 365 in polling station 1 of 90-Salem (South) in 2026 (D.M.K 217, A.I.A.D.M.K 127, N.T.K 21), while across the state it came first in 34,864 of the 75,063 polling stations with votes (46.4%), won 108 of 234 seats and formed the government; the A.I.A.D.M.K alone came first in 16,314 and the D.M.K alone in 15,451.
Asking voters directly is on record as automated calls with no sender: in April 2026 voters in Thousand Lights, Chennai, were asked to press a key for the party they would vote for.
Meta's developer pages set out what an app can read through a login: a professional Instagram account gives the app's server an access token (60 days, refreshable) that returns the usernames and words of those who comment on its posts and message it and its followers' age, gender, city and country only as totals; likes and followers come only as counts, and whom anyone follows never.
OpenAI (a public beta in October 2026 by press reports, while its guide says generally available; the status is not settled here) and Google (11 November 2026) have released decision APIs that return a choice from a fixed set with a probability for each; their examples are business work, OpenAI's usage policies bar profiling individuals and political campaigning, and Google's policy bars tracking or monitoring people without their consent.
Studies in Denmark (2017) and the United States (2013) found that Facebook likes estimate the vote (Denmark, 659 users who had liked political actors) or a political leaning (United States) better than a survey baseline or chance; I found no such measurement in India.
No report was found of any party in India reading followers' data through its functionaries' accounts or putting voters to a decision AI: that parties will go after this data next is the AI's forecast, and its reason, that a guess from identity or a direct question misses a change of mind, is its view.
Background
Tamil Nadu has 234 Assembly seats; the 2026 result sheets cover 75,064 polling stations (booths), 75,063 of them with votes. The Election Commission's electoral roll is published in parts, each tied to a polling station, and recognised parties get free copies; each entry gives a voter's serial number, EPIC (voter ID) number, name, a relation's name, house number, age and sex, arranged by house number.
Each recognised party may appoint a Booth Level Agent for every polling station, a system the Commission introduced in November 2008, and parties run booth committees that go house to house. After an election the returning officer's booth-wise result sheet, Form 20, stands on the Tamil Nadu Chief Electoral Officer's website, so anyone can see how each booth voted, never how a person voted.
Voter profiling is the sorting of voters by how they are expected to vote: ours, theirs, swing. In May 2026 the T.V.K, the party of the actor Vijay, won 108 seats and formed the government; the D.M.K won 59 and the A.I.A.D.M.K 47.
On Instagram a professional (business or creator) account can log in to an outside app through Meta's Business Login; the app's server then holds an access token, the credential with which it reads that account through Meta's API. A decision API is an AI service that answers a closed question (yes or no, one of a fixed set of choices, a score) with a probability.
Who gains, who loses
For voters: what they tell a party worker at the door is theirs to choose, but a comment or a message to a local party account carries their username, and their age group, gender and town enter that account's audience totals; a follower who only watches and likes is, by the official API, a number and not a name.
For parties: every camp makes booth counts, and the 2026 result sheets show the count's limits for all three large parties, so each has a reason to look for better signals, and each would carry the risk of breaking a platform's terms, an AI company's rules and, from about 13 May 2027, India's data protection duties.
For the local functionary whose account would be joined, the account and its standing with Meta are his own. For Meta, OpenAI and Google, their own reviews and policies are, for now, what stands between the capability and such a use.
What most coverage misses
Coverage of data and AI in campaigns is mostly about what parties send out (cloned voices, calls, reels); what could come in is less noticed. The official Instagram API is narrower than people fear and wider than they think: it never names who liked or who follows, but it names everyone who comments on or messages a professional account that has logged in to an app.
The roll has no caste, religion or phone number; whatever a party holds of those, it gathered or guessed itself. Form 20 lets anyone test a party's booth claim: the T.V.K's aim of 350 was passed in the booth shown on its stage (365), and the same sheets show it first in fewer than half the state's booths, with each rival first in about a fifth. The IVR calls that asked for a vote choice named no sender.
And the makers of the decision APIs bar, in their own rules, the use the forecast describes: OpenAI's bar profiling individuals and political campaigning, and Google's bar tracking or monitoring people without their consent.
The numbers
Voters of booth 1 of Salem (South) on the slide (386 men, 395 women; age bands 212, 332, 222 and 15).
youtube.comSeats the T.V.K won in May 2026, with 34.92% of the votes (1,72,26,209); 118 make a majority; the D.M.K 59, the A.I.A.D.M.K 47.
web.archive.orgDistrict secretaries who received the booth-committee book and the pen drive on 26 and 27 April 2025.
youtube.comShow more (11)
The T.V.K's 2026 vote set against its own sum of 350 votes a booth in 69,000 booths (2,41,50,000).
web.archive.orgThe keys of the unsigned automated call in Thousand Lights, April 2026: A.I.A.D.M.K, D.M.K, T.V.K, Naam Tamilar, others.
dtnext.inHow long the long-lived access token of Meta's Business Login for Instagram is valid before it must be refreshed.
developers.facebook.comThe most recent messages of a conversation whose details the Conversations API returns.
developers.facebook.comAudience totals are not returned for an account with fewer than 100 followers, and each breakdown returns only its 45 largest values.
developers.facebook.comThe choices Google's Gemini Decision 1 can pick from, returning a probability for every one.
blog.googleHow often political Facebook likes, against a survey baseline, predicted the vote of 659 Danish users (2017).
pmc.ncbi.nlm.nih.govHow often Facebook Likes told a Democrat from a Republican in the United States (2013): a score for telling two people apart, on 9,752 who had stated a view.
pmc.ncbi.nlm.nih.govThe persuasive effect of messages tailored by GPT-4 against untailored ones (2024): not statistically different.
doi.orgPeople in India who installed the app linked to Cambridge Analytica, and their friends 'potentially affected' (Facebook, 2018).
thewire.inAbout when the duties and penalties of the Digital Personal Data Protection Act, 2023, begin.
indiacode.nic.inTimeline
- The Registration of Electors Rules, 1960: rule 6(1) orders the names in each part of the roll 'arranged according to house number'; rules 11(c) and 22(1)(c) give each recognised party two free copies. thc.nic.in
- The Election Commission introduces the Booth Level Agent: recognised parties may appoint one for each polling station, a voter of that part of the roll. ceoharyana.gov.in
- Kosinski, Stillwell and Graepel publish in PNAS: among volunteers in the United States, Facebook Likes tell a Democrat from a Republican 'in 85% of cases'. pmc.ncbi.nlm.nih.gov
- Facebook's Graph API v1.0, through which an app could reach its users' friends' data, expires. developers.facebook.com
- Kristensen and others publish in PLOS ONE: among 659 Danish Facebook users, political likes predict the vote at 60.9% against 35.8% for a survey baseline. pmc.ncbi.nlm.nih.gov
- Facebook tells The Wire that 'only 335 people in India installed the App' linked to Cambridge Analytica and that '562,120 additional people in India were potentially affected, as friends of people who installed the App'. thewire.in
Show more (21)
- Kamal Nath v. Election Commission of India: the Supreme Court's judgment records the Commission's reason for giving rolls as image files, that the move 'is aimed at prevention of voter profiling and data mining', and says 'it is not necessary to go into this aspect'. api.sci.gov.in
- Instagram is reported to be removing its 'Following' activity tab, which showed other people's likes and follows. engadget.com
- Residents of the South Chennai Lok Sabha constituency get an IVR call asking which candidate they support; it 'does not mention who is conducting the survey'. thenewsminute.com
- Day one of the T.V.K's booth agents' conference at Kurumbapalayam, Coimbatore: Aadhav Arjuna shows booth 1 of Salem (South) on a slide and sets the aim of 350 votes there in 2026; Vijay hands ten district secretaries a booth-committee book and a pen drive. youtube.com
- Day two: Vijay hands thirteen district secretaries the book and the pen drive; by the broadcast's captions Aadhav Arjuna puts the state's sum at 350 votes a booth in 69,000 booths, over 2 crore 40 lakh votes. youtube.com
- The Madras High Court (Madurai Bench), by an interim order, restrains the sending of OTP verification messages in the D.M.K's 'Oraniyil Tamil Nadu' membership drive until the questions of privacy and data protection are examined. thenewsminute.com
- The Supreme Court declines to interfere with the High Court's interim order. livelaw.in
- M.K. Stalin, at a meeting of the D.M.K's 'My Booth, Victory Booth' campaign at Mahabalipuram, tells cadres to 'treat your booth as your battleground'; the report says he asked for at least 70 additional votes in every polling booth for the D.M.K-led alliance. dtnext.in
- The version of OpenAI's Usage Policies that bars profiling individuals and political campaigning takes effect; the version effective 20 November 2026 carries the same lines. openai.com
- The commencement of the Digital Personal Data Protection Act, 2023, is notified: section 3, the duties on those who process personal data and the penalties begin eighteen months later, about 13 May 2027. indiacode.nic.in
- IANS reports that the D.M.K's constituency surveys ask voters which party they intend to support and record 'respondents' caste and socio-economic details'. newkerala.com
- IANS reports the A.I.A.D.M.K's 'Booth Sevai' app, on trial in Coimbatore and Tiruppur districts, with 'a field to record the voting preference of each individual voter'. mangalorean.com
- DT Next reports automated calls from unknown numbers to voters in Thousand Lights, Chennai, asking them to press a key for the party they will vote for; no sender is named. dtnext.in
- Dinamalar reports from Andipatti, Theni district, that booth committees of the D.M.K and the A.I.A.D.M.K are sorting each polling station's votes as support, against and neutral. dinamalar.com
- Tamil Nadu votes in the 2026 Assembly election. thenewsminute.com
- Counting: the T.V.K wins 108 of 234 seats (34.92% of the votes), the D.M.K 59 and the A.I.A.D.M.K 47. web.archive.org
- Vijay is sworn in as Chief Minister; the T.V.K, short of 118 seats, forms the government with other parties' support. telanganatoday.com
- OpenAI's Decisions API: press reports describe a public beta in October 2026, while its guide, which carries no date, says it is 'now generally available', so its status is not settled here; the guide describes typed answers: a probability, 'a choice from a fixed set', or a score. developers.openai.com
- The Election Commission orders the Special Summary Revision 2027 in 17 states and Union territories, Tamil Nadu among them; parties are to get the draft and final rolls in 'readable PDF format' without photographs, where they got image files before. thefederal.com
- Google introduces the Gemini Decisions API with its Gemini Decision 1 model: a choice 'from 2 to 255 choices, returning the winner and normalized probabilities for every candidate'. blog.google
- Tamil Nadu's draft roll for 2027 is published; claims and objections run to 11 December 2026 and the final roll is due on 4 January 2027. dtnext.in
Claims checked
Form 20 of 2026 gives the T.V.K 365 votes in polling station 1 of 90-Salem (South). The stations were redrawn after 2021, so it is the same booth number (and, by a third-party page for 2026, the same school; no official 2026 list was reached), not necessarily the same voters.
It came first in 34,864 of the 75,063 polling stations with votes, 46.4%, and won 108 of 234 seats in a contest of several parties.
An entry holds a serial number, the EPIC number, the name, a relation's name, the house number, age and sex (Manual on Electoral Rolls, para 9.9.5). ceoharyana.gov.in
The speaker's two descriptions differ (a booth-wise list of voters; the list of polling stations and the numbers of voters) and one press report says election data of 2019 and 2021. Nothing shows caste or religion on it. youtube.com
Show more (11)
The calls of April 2026 came from unknown numbers and the call of March 2024 did not say who was conducting the survey; no sender is named in either report. dtnext.in
The official API returns 'like_count' and 'followers_count' only; no field or edge lists likers or followers, and nothing returns whom anyone follows. developers.facebook.com
With the comments and messages permissions the API returns the commenter's username and words, and each message's sender (username and id) and text, for the 20 most recent messages of a conversation. developers.facebook.com
Only professional accounts can use the Business Login, and an app that serves accounts its maker does not own needs Advanced Access, which requires App Review and Business Verification. The video's spoken line, that pressing 'Allow' is enough, gives only the account holder's step. developers.facebook.com
No report of it was found, in Tamil or in English. It is the AI's forecast of what parties will try, not a record of what any party does. The video's closing line, that long before a vote parties will try with AI to know who will win and what to do, is the same forecast; no report of any party doing it was found.
Not shown: both companies' examples are business work; OpenAI's Usage Policies bar profiling individuals and political campaigning, and Google's policy bars tracking or monitoring people without their consent. I found no party, firm or person on record using either service on voters. openai.com
Studies in Denmark (the vote itself, among 659 users who had liked political actors) and the United States (a party leaning) found that Facebook likes estimate these better than a survey baseline or chance; I found nothing measured in India, in Tamil or on Instagram, and an estimate is not knowledge of how a person votes. The video's 'more accurately' for follows and likes is the AI's view, at this strength only. pmc.ncbi.nlm.nih.gov
The 85% of the 2013 study is the chance of telling apart one Democrat and one Republican picked at random, measured on 9,752 Americans who had stated a view; it is not a share of people read right, and not a vote. pmc.ncbi.nlm.nih.gov
The judgment records that as the Election Commission's reason ('According to it …') and says it is not necessary to go into it; it is not the Court's finding. api.sci.gov.in
The Act's duties and penalties begin about 13 May 2027, and section 3(c)(ii) says it does not apply to personal data a person has made publicly available; no court has ruled on likes, comments or follows. indiacode.nic.in
In the one large experiment found (2024, n = 8,587), messages tailored by GPT-4 were 'not statistically different' in effect from untailored ones. doi.org
Findings
As heard at 19:15 to 19:23 of the live broadcast of the booth agents' conference (Puthiya Thalaimurai TV). He was then the T.V.K's general secretary for election campaign management (a minister in the T.V.K government since May 2026, as several outlets reported). The slide behind him: Salem South, booth 1; voters 781 (386 men, 395 women); votes polled 604 (77%); an age row of four bands (18-30: 212, 31-50: 332, 51-80: 222, 80+: 15, which add up to 781); Assembly election 2021: D.M.K 280, A.I.A.D.M.K 248, N.T.K 37, M.N.M 14, A.M.M.K 7. No caste or religion is on the slide. By the captions he goes on: 'இதுதான் நம்முடைய குறிக்கோள்' (this is our aim). Puthiya Thalaimurai TV, live broadcast of the T.V.K booth agents' conference, Coimbatore, 2025-04-26
Day one, by the broadcast's captions, the compere at 24:45: 'அந்த பென்டிரைவையும் பூத் கமிட்டி புத்தகத்தையும் … வெற்றி தலைவர் வழங்குகிறார்'; the picture at 24:38 to 25:06 shows a white spiral-bound book with a small red case handed over under the meeting's banner. Day two, the announcer at 1:34:49: '13 மாவட்ட செயலாளர்களுக்கும்'. The press calls the event a two-day workshop at a private college in Kurumbapalayam (DT Next, 27 April 2025; The News Minute, 28 April 2025). Puthiya Thalaimurai TV, live broadcasts of 26 and 27 April 2025, 2025-04-26
By the captions of the day-one broadcast Aadhav Arjuna describes it twice: at 15:08 to 15:24 as 'the latest booth-wise list' of voters that the Election Commission released, and at 20:58 to 21:21, holding it up, as the constituency's list of polling stations and the numbers of voters the Commission last released. The News Minute, 28 April 2025: 'Each participant received a booklet and a pen drive containing election data from 2019 and 2021.' The book, by the captions of his speech: each party's booth-wise votes in past elections; he also speaks of each booth's count of voters, the ratio of men and women and the number of families. Puthiya Thalaimurai TV, live broadcast of 26 April 2025 (captions), 2025-04-26
Election Commission of India, Manual on Electoral Rolls (March 2023), para 9.9.5, the 'Sequence of entries relating to electors in elector details box': serial number, EPIC number, name, name of relation, house number, 'the Age and then Sex', and the photograph (para 25.3.4: on the printed copy only). No field for caste, religion, phone, Aadhaar or party is listed. Election Commission of India, Manual on Electoral Rolls, March 2023
Show more (27)
Registration of Electors Rules, 1960, rule 6(1): 'the names of electors in each part of the roll shall be arranged according to house number, unless the chief electoral officer … determines in respect of any part that the alphabetical order is more convenient'. Rule 11(c): 'supply free of cost two copies of each separate part of the roll to every political party for which a symbol has been exclusively reserved in the State by the Election Commission'; rule 22(1)(c) says the same of the roll as finally published. Registration of Electors Rules, 1960 (the Tripura High Court's copy)
DT Next, 4 April 2026 (headline: the poll panel cautions against unknown poll survey calls), a voter in Thousand Lights, Chennai: 'The automated call asked me to press 1 for AIADMK, 2 for DMK, 3 for TVK, 4 for Naam Tamilar, and 5 for others'; 'a call from an unknown number'; an official: 'We need to verify whether political parties have obtained permission for such mobile-based surveys.' No sender is named. No source found has a party asking voters whom they will vote for by IVR, and no Tamil Nadu party was found using IVR; a party named in print used it for another question (the Telugu Desam Party, in 2014, to choose its own candidates: Sakshi Post, 8 April 2014). DT Next, 2026-04-04
The News Minute, 30 March 2024: 'The Interactive Voice Response (IVR) call, which does not mention who is conducting the survey', listed the constituency's four main candidates on keys 1 to 4 and 'others' on 5; the matter was reported to the Media Certification and Monitoring Committee. The News Minute, 2024-03-30
Meta, 'Business Login for Instagram' (read 12 November 2026): for Instagram professional (business or creator) accounts only. The app opens Instagram's authorization window; the account holder grants the permissions the app asks for (instagram_business_basic, instagram_business_manage_comments, instagram_business_manage_messages and others); Instagram returns a code, valid one hour and for one use, which the app's server exchanges for a short-lived token and then a long-lived one (60 days, refreshable). Meta calls the credential an access token; its text names no button ('after they allow or deny permission request'). A personal account cannot log in this way. Meta for Developers, Business Login for Instagram
Meta, Instagram Platform overview: an app serving Instagram professional accounts that its maker does not own or manage needs 'Advanced Access', and 'This access level requires App Review and Business Verification.' Whether Meta's review would pass an app of the kind forecast is not known. Meta for Developers, Instagram Platform overview
Meta's IG Comment reference: 'from.username', 'Username of Instagram user who created the IG Comment.', with the comment's text and time; Instagram Login needs instagram_business_basic and instagram_business_manage_comments. It is the account's username, not a civil name. Meta for Developers, IG Comment reference
Meta's Conversations API (read 13 November 2026): 'A list of conversations for your app user's Instagram professional account' and 'A list of messages within each conversation', each message with 'from' (username and id), 'to', 'message' and 'created_time'; 'you can only get details about the 20 most recent messages in the conversation'; conversations in the Requests folder idle for 30 days are not returned; permissions instagram_business_basic and instagram_business_manage_messages; 'Advanced Access if your app serves Instagram professional accounts you don't own or manage'. Meta for Developers, Conversations API
Meta's IG User Insights reference (read 13 November 2026): 'follower_demographics', 'The demographic characteristics of followers, including countries, cities and gender distribution', with the breakdowns age, city, country and gender; 'Not returned if the IG User has less than 100 followers'; 'engaged_audience_demographics' the same for those who engaged; 'Demographic metrics only return the top 45 performers'. Permissions instagram_business_basic and instagram_business_manage_insights. Meta for Developers, IG User Insights reference
Meta's IG Media reference: 'like_count', 'Count of likes on the media, including replies on comments.', and the edges children, collaborators, comments, ordered_comments and insights; no edge lists the likers. The IG User reference has 'followers_count' ('Total number of Instagram users following the user.') and 'follows_count' only. Meta for Developers, IG Media and IG User references
Instagram's help page on who can see what a person follows: the public accounts a person follows are seen by 'Anyone on Instagram and Threads'; when the person's own account is private, the private accounts they follow are seen by 'Your followers'. A like on a public post is 'visible to anyone who can see the post'. Instagram's Terms of Use, 4.2, forbid 'accessing or collecting information in an automated way without our express permission, regardless of whether such automated access or collection is undertaken while logged-in to an Instagram account'. The Terms are a contract, not a law. Instagram Help Centre
Instagram's Terms of Use, 4.2 (the page linked here), forbid anyone to 'solicit, collect, or use login credentials or badges of other users' and forbid 'accessing or collecting information in an automated way without our express permission'. Meta Platform Terms, a separate page ('Last updated February 3, 2026'), prohibited practices: 'Processing Platform Data without valid User consent in order to build or augment user profiles for any purpose.' Whose consent counts for a person who comments on someone else's account is a reading; no ruling was found. These terms are a contract, not a law. Instagram Terms of Use; Meta Platform Terms
OpenAI's developer guide (read 12 November 2026): 'The Decisions API evaluates text, images, and JSON and returns typed answers about 10x faster than the Responses API. Get the probability that a condition is true, a choice from a fixed set, or a score against a rubric. Use those answers to classify content, route requests, and prioritize work in your application.'; 'The Decisions API is now generally available.'; 'Choice and score answers return a probability distribution and a separate `confidence` field'. The guide carries no date and says it is generally available; press reports put a public beta in October 2026; no source on this page settles whether it is in beta or generally available. I found no record of the service being used on a question about a voter. OpenAI developer documentation, Decisions guide
Google, 'Gemini Decisions API: Fast, calibrated multimodal decisions', 11 November 2026: 'Today, we're introducing the Gemini Decisions API, a new endpoint for fast, closed-set decisions powered by our new Gemini Decision 1 model (gemini-decision-1), available through the Gemini API in Google AI Studio.'; 'choice (Multiple Choice): Selects from 2 to 255 choices, returning the winner and normalized probabilities for every candidate'; 'Every answer includes a confidence score (from 0 to 1)'. Its examples are business work (a call recording with account details, a damage claim, choosing a tool, policy checks, routing between models). Google, The Keyword, 2026-11-11
OpenAI Usage Policies, the version 'Effective: October 29, 2025' and the next, 'Effective: November 20, 2026': 'you may never use our services for: … evaluation or classification of individuals based on their social behavior, personal traits, or biometric data (including social scoring, profiling, or inferring sensitive attributes)'; 'we don't allow our services to be used … including any use for: … political campaigning, lobbying, foreign or domestic election interference, or demobilization activities'. OpenAI Usage Policies, 2025-10-29
Google, Generative AI Prohibited Use Policy ('Last Modified: December 17, 2024'): no content that 'Tracks or monitors people without their consent.', and no 'using personal data or biometrics without legally-required consent'. The Gemini API's terms: 'The Services include safety features to block harmful content, such as content that violates our Prohibited Use Policy.' That sorting voters from their social media behaviour falls under these lines is an inference; the policy names no political use. Google, Generative AI Prohibited Use Policy, 2024-12-17
Kristensen and others, 'Parsimonious data: How a single Facebook like predicts voting behavior in multiparty systems', PLOS ONE 12(9): e0184562, 2017: 659 Danish Facebook users who had liked political actors; the full history of political likes predicted the vote at 60.9% against 35.8% for a survey baseline of sociodemographics, values and opinions; 'likes predict voting behavior dramatically better than survey questions'. One country, Facebook, a field of several parties. PLOS ONE
Kosinski, Stillwell and Graepel, 'Private traits and attributes are predictable from digital records of human behavior', PNAS 110(15): 5802 to 5805, online 11 March 2013: 58,466 volunteers in the United States; the model 'correctly discriminates … between Democrat and Republican in 85% of cases'. The 85% is the chance of telling apart one person of each kind picked at random, measured on the 9,752 who had stated a political view; it is not a vote, and the paper makes no comparison with a guess from age or other identity. PNAS (PubMed Central copy), 2013-03-11
Hackenburg and Margetts, 'Evaluating the persuasive influence of political microtargeting with large language models', PNAS 121(24), 7 June 2024, a preregistered experiment (n = 8,587): 'the persuasive impact of microtargeted messages was not statistically different from that of non-microtargeted messages (4.83 vs. 6.20 percentage points, respectively, P = 0.226)'. The untailored messages did move people; the finding is about tailoring. PNAS, 2024-06-07
Lee and others, 'LLMs Can Infer Political Alignment from Online Conversations', a preprint of 11 March 2026 (not peer reviewed): GPT-4o at user level, by the paper's tables, a macro F1 of 0.69 on Debate.org and 0.735 on Reddit over all texts (its running text gives slightly different figures, so: about 0.7); 'a random classifier is expected to achieve a macro F1 score of 0.5'; 'our analyses are limited to English-language and U.S. contexts'. From what people wrote, not from likes or follows. arXiv 2603.11253, 2026-03-11
The A.I.A.D.M.K: the 'Booth Sevai' app, on trial in Coimbatore and Tiruppur districts, with 'a field to record the voting preference of each individual voter, listing major political parties as options'; a booth agent: 'If still uncertain, we mark them as 'undecided'. Those opting for NOTA or unwilling to disclose are recorded as 'neutral'' (IANS, 10 February 2026, an unnamed agent and unnamed party sources). The D.M.K: constituency surveys that ask voters which party they intend to support and record 'respondents' caste and socio-economic details' (IANS, 9 February 2026). The D.M.K and the A.I.A.D.M.K: booth committees sorting each station's votes as 'ஆதரவு, எதிர்ப்பு, நடுநிலை' (Dinamalar, Andipatti and Cumbum, Theni district, April 2026). The T.V.K: a booth-committee book of each booth's past results and, on the slide of one booth shown on its stage, the count of voters, of men and women and four age bands (April 2025). IANS (carried by Mangalorean); IANS (carried by New Kerala); Dinamalar, 2026-02-10
IANS, 9 February 2026: 'The surveys are being conducted directly among voters, party cadres and key influencers'; members of the party's strategy team describe 'contacting voters at the constituency level and first asking which party they intend to support'; 'respondents' caste and socio-economic details are also being recorded to understand voting patterns more accurately'. By what means the surveys were run is not said. Caste is named here only as a kind of identity, of no group and no person. IANS (carried by New Kerala), 2026-02-09
By the captions of the day-two broadcast (1:26:33 to 1:26:52) Aadhav Arjuna says 350 votes a booth in 69,000 booths make over 2 crore 40 lakh votes, enough to win; 350 times 69,000 is 2,41,50,000. The Election Commission's party-wise page gives the T.V.K 1,72,26,209 votes (34.92%), which is 71% of that and about 250 a booth by the party's own count of booths. Puthiya Thalaimurai TV, live broadcast of 27 April 2025 (captions); Election Commission of India results, 2025-04-27
Kamal Nath v. Election Commission of India, 12 October 2018, para 24: 'According to it, issues of privacy of voters are involved and the move of ECI is aimed at prevention of voter profiling and data mining.', after which the Court says 'it is not necessary to go into this aspect': the Commission's reason as the Court records it, not the Court's finding. On 15 October 2026 the Commission ordered the Special Summary Revision 2027 in 17 states and Union territories with the draft and final rolls for parties in 'readable PDF format' without photographs (PTI and IANS). PTI's 'for the first time' holds only since 2018; the judgment itself records text rolls in some states before. Supreme Court of India, judgment of 12 October 2018, 2018-10-12
A Facebook spokesperson to The Wire, 5 April 2018: 'only 335 people in India installed the App'; '562,120 additional people in India were potentially affected, as friends of people who installed the App', 562,455 in all. It is Facebook's own count, of people 'potentially affected'. The route (friends' data through Graph API v1.0) expired on 30 April 2015 and is not how Instagram's API works today. The Wire, 2018-04-05
Madras High Court (Madurai Bench), 21 July 2025: 'an order of interim injunction is granted to the limited extent of restraining' the sending of OTP verification messages in the D.M.K's 'Oraniyil Tamil Nadu' membership drive until privacy and data protection are examined; as The News Minute quotes the order: 'Data protection and voter data privacy are the essential facets of Article 21 of the Constitution of India and falls within the ambit of right to privacy'. The petitioner's charges are allegations; the party's counsel told the court it was not collecting Aadhaar details. On 4 August 2025 the Supreme Court declined to interfere. Nothing final has been held, and what has happened in the case since was not found. The News Minute; LiveLaw; DT Next, 2025-07-22
Digital Personal Data Protection Act, 2023, section 3(c)(ii): the Act shall not apply to 'personal data that is made or caused to be made publicly available by— (A) the Data Principal to whom such personal data relates; or (B) any other person who is under an obligation under any law … to make such personal data publicly available'; its illustration is a person who has made her data public on social media. By the commencement note, section 3, the duties of notice and consent and the penalties begin eighteen months from 13 November 2025, about 13 May 2027. That public likes, comments and follows, or the published roll, fall outside the Act is a reading; no court has ruled. India Code, Digital Personal Data Protection Act, 2023
Tamil Nadu's draft roll for the Special Summary Revision 2027 was published on 12 November 2026; claims and objections run to 11 December 2026 and the final roll is due on 4 January 2027 (PTI; DT Next, 16 October 2026). The duties under the data protection Act begin about 13 May 2027. DT Next, 2026-10-16
What they said before
On the stage of the party's booth agents' conference, of booth 1 of Salem (South): 'இத 2026ல 280 டிஎம்கேன நம்ம 350 மாத்தி காமிக்கணும்' (the D.M.K's 280 there must become the T.V.K's 350 in 2026). youtube.com
By the broadcast's captions: 350 votes a booth in 69,000 booths make over 2 crore 40 lakh votes, and that is enough for the party. youtube.com
Called the booth agents the 'backbone of the party' and said: 'Booth committee members should meet voters in person.' thenewsminute.com
Show more (10)
At a meeting of the 'My Booth, Victory Booth' campaign at Mahabalipuram: 'Booth-level meetings should be held daily.' and 'treat your booth as your battleground.'; the report says he asked for at least 70 additional votes in every polling booth for the D.M.K-led alliance (the reporter's sentence). dtnext.in
Began rolling out its 'Booth Sevai' app for booth agents, on trial in Coimbatore and Tiruppur districts, with a field for each voter's voting preference (as IANS reported from unnamed party sources). mangalorean.com
Ran constituency surveys to choose candidates that asked voters which party they intended to support and recorded respondents' caste and socio-economic details (as IANS reported from members of the party's strategy team). newkerala.com
Told the Supreme Court that giving rolls as image files 'is aimed at prevention of voter profiling and data mining' (as the judgment of 12 October 2018 records it). api.sci.gov.in
Ordered that for the Special Summary Revision 2027 parties get the draft and final rolls in 'readable PDF format' without photographs (as PTI and IANS reported). thefederal.com
'We need to verify whether political parties have obtained permission for such mobile-based surveys.' (as DT Next quotes). dtnext.in
Prohibits 'Processing Platform Data without valid User consent in order to build or augment user profiles for any purpose', and puts apps that serve other people's professional accounts behind App Review and Business Verification. developers.facebook.com
Bars 'evaluation or classification of individuals based on their social behavior, personal traits, or biometric data (including social scoring, profiling, or inferring sensitive attributes)' and any use for 'political campaigning'. openai.com
Bars content that 'Tracks or monitors people without their consent.' and 'using personal data or biometrics without legally-required consent'. policies.google.com
In an interim order: 'Data protection and voter data privacy are the essential facets of Article 21 of the Constitution of India and falls within the ambit of right to privacy' (as The News Minute quotes the order). thenewsminute.com
Still open
- Do people tell every party's worker 'it's yours'? No survey or count found measures what voters say at the door against how they vote.
- Does any party in India run an app that its functionaries log in to with Instagram, or read followers' data through their accounts? None was found.
- Would Meta's App Review pass an app that reads many local functionaries' comments, messages and audience totals for a party?
Show more (6)
- How well do follows and likes estimate a vote in Tamil Nadu's field of several parties? I found nothing measured in India, in Tamil or on Instagram.
- What exactly was on the pen drive handed out on 26 and 27 April 2025: the roll itself, or the list of polling stations with voter numbers and past results?
- Who ran the unsigned IVR calls of March 2024 and April 2026, and what became of the answers?
- How many of the 781 voters of Salem (South) station 1 in 2021 are in the redrawn station 1 of 2026?
- When the data protection Act's duties begin about 13 May 2027, will public likes, comments and follows be read as outside it?
- Would OpenAI's or Google's service answer a question that sorts a named person as a party's voter? Neither was tried.
Sources · 43
- ceoharyana.gov.in
- thc.nic.in
- youtube.com
- elections.tn.gov.in
- elections.tn.gov.in
- elections.tn.gov.in
- elections.tn.gov.in
- web.archive.org
Show more (35)
- telanganatoday.com
- dtnext.in
- thenewsminute.com
- newkerala.com
- mangalorean.com
- dinamalar.com
- developers.facebook.com
- developers.facebook.com
- developers.facebook.com
- developers.facebook.com
- developers.facebook.com
- developers.facebook.com
- developers.openai.com
- openai.com
- blog.google
- policies.google.com
- pmc.ncbi.nlm.nih.gov
- pmc.ncbi.nlm.nih.gov
- Instagram Help Centre · help.instagram.com
- Instagram Terms of Use; Meta Platform Terms · help.instagram.com
- PNAS · doi.org
- arXiv 2603.11253 · arxiv.org
- Puthiya Thalaimurai TV, live broadcast of 27 April 2025 (captions); Election Commission of India results · youtube.com
- Supreme Court of India, judgment of 12 October 2018 · api.sci.gov.in
- The Wire · thewire.in
- The News Minute; LiveLaw; DT Next · thenewsminute.com
- India Code, Digital Personal Data Protection Act, 2023 · indiacode.nic.in
- DT Next · dtnext.in
- developers.facebook.com
- engadget.com
- livelaw.in
- dtnext.in
- thenewsminute.com
- thefederal.com
- thenewsminute.com