A few hours ago, I found my name, address and linked to my father’s name, on a website that has digitised India’s Election Commission rolls. In the email, the person who wrote in, offering to partner with MediaNama for “a custom site, a data extract/service, or an API” said, “We index the published Election Commission rolls (~57 crore electors) so you can query people, households and booth-to-constituency composition.”
What the website revealed from my name
From a search of the website for my name, I found that, including me, there are 20 voters in India with the same name as mine, 12 in my state, and 2 in my district. The website, clearly made with AI, also has an extensive relationship graph. It identifies:
- Relatives, including generational relationships between people
- Who else has the same address (ours is a neighbourhood, and there are 26 people in the same neighborhood)
- Who lives next door: it identified houses/blocks on either side, and who lives there.
Essentially, it creates a relationship graph for a person, including their immediate family and their relationships with their neighbours and the voting booth.
Privacy Issues that SIR and this mapping create
Mapping political data is probably as old as organised politics. In 2013, at the IAMAI Mobile Marketing Summit, Netcore founder Rajesh Jain, then a key member of Narendra Modi’s team, had explained how political mapping worked then. He said at the summit:
“What if you get them to SMS the voter ID card number on a long code, on a number, where you can use the voter ID to pinpoint the person on the electoral rolls with are available with the party workers locally, and also available on the election commission website.
The key idea here is that the voter ID plus the mobile number is the double KYC (Know Your Customer). The Voter ID has very good KYC, and the mobile number does lead to a genuine person. You get a unique geo-identity. You know who the person is. You have name, address, location, age, gender, all of that, from the voter rolls. Because of the SMS, you have a confirmation, which you can then use for continuing engagement.” This not only allows political parties to know who has attended, but more importantly, “this can lead to micro targeted campaigns.”
If you think about it, this will now only end up enriching political party databases for targeting by mapping relationships.
The following privacy issues arise:
- Mass search: Finding an entry in a constituency requires some knowledge of a person, and it is a manual process. Digitisation means that it becomes a nationwide name search, fuzzy mapping of individuals, and instant reports on individuals, making mass identification and surveillance possible, especially when layered with other data.
- Household and social-graph reconstruction: the ability to use AI to create embeddings and relationship graphs means that it exposes relatives, surnames, house numbers and adjacent houses and links them together: it can be used to reveal multi-generational households and relationships in areas, and probabilistically, across regions. Information about neighbours and adjoining households was disclosed without their consent.
- Political and community profiling: Layer historical voting percentage data on top of this over a period of time, and you can start using polling station, geography, surname and household metrics as proxies, however unreliable, for caste, religion, income or political preferences. This enables microtargeting, discrimination and communal targeting.
- Fraud enabling: Indian government is doing a lot to address fraud, including pushing through stupid ad ignorant proposals like SIM Binding, trying to force private entities to share their spam detection mechanisms, blocking accounts, but they’re doing little to address the rampant disclosure of personal information that is enabling this fraud, especially social engineering. This SIR exercise and the relationship mapping, done for the political purpose of disenfranchising voters, is only going to make fraud worse because of relationship mapping.
- Enabling reuse and AI scraping: the system generates a JSON file for each individual, helpfully enabling anyone to make copies of it for processing by AI. I mean, the guy wrote in to pitch enabling API access to this data, so it gets worse. This means someone making copies won’t need to do OCR, dedupe info, map relationships. While the page has a robots.txt file with a noindex clause, to prevent search engine scraping, that itself is insufficient: we know that AI engines, and especially the Chinese AI bots, can ignore noindex clauses. In 2023, MediaNama had server issues because we moved to a new server and were inundated with Chinese AI bots. It’s not just the Chinese. Last month, despite a no-index file, Amazon’s scraping bots took our site down. This is why you have Cloudflare bot verification when you try to access our archives.
How this was possibly made
The cost of each output, according to my JSON file:
- one Claude Haiku 4.5 call;
- 152 tokens in, 27 out;
- cost ₹0.13;
Based on an analysis of my information, which reveals the processing to achieve a confidence score for individuals, I asked ChatGPT to reconstruct the process of creating this analysis.
Here’s what I got:
Published electoral-roll PDFs
↓
Machine reading/OCR of multilingual pages
↓
Transliteration + field extraction
(name, relative, relation, age, sex, house, EPIC, district, constituency, part, serial, roll language/type)
↓
Search index with exact, token, phonetic and geographic facets
↓
Subject fields parsed into structured search inputs
↓
Multiple candidate searches and frequency measurements
↓
Bayes-like multiplicative score for each candidate
↓
Requester manually selects an elector
↓
Selected row expands into:
households → family graph → roll-order neighbours
→ polling-station statistics → namesake maps
↓
Server-rendered HTML + public JSON + technical log
For this run, the system made 50 searches and says it screened 62 records. It queried:
- all name matches nationwide;
- exact-phrase versus “sounds like” matches;
- state and district subsets;
- name plus relative-name pairs;
- candidates within a three-year age tolerance;
- reordered name tokens;
- shortened and full versions of the relative’s name;
- addresses associated with possible relative matches; and
- geographical and roll-type facets.
The search modes visible in the JSON include all, phrase and any. “PAHWA NIKHIL” shows token reordering; fuzzy candidates scored at 67% and 33% show token/string similarity. The precise phonetic algorithm is not disclosed.
The engine begins with a name-frequency prior:
[ P_0 = 1/20,\qquad O_0 = P_0/(1-P_0)=1/19 \]It then multiplies the prior odds by hard-coded “likelihood ratio” weights:
| Evidence | Multiplier |
|---|---|
| Age within one year | ×25 |
| Same sex | ×1.6 |
| Same house and locality | ×200 |
| Exact name spelling | ×3 |
| Same locality | ×25 |
| Same district | ×5.01 |
| Exact relative-name pair | ×20 |
| Correct relation type | ×2.16 |
For household and neighbour relationships:
the system queries everyone sharing:
rollId + partNumber + normalized house field
For family relationships:
elector ──father/husband/mother/etc.──> named relative
The engine:
- creates a node for the elector;
- searches for the named relative as another elector;
- creates a “ghost” node if no matching electoral entry is confidently found;
- assigns generation from relation type;
- attempts to merge nodes using names, sex, age gaps and co-residence; and
- looks for shared named relatives across nearby households.
While we have a fundamental right to privacy, unfortunately, there appears to be no requirement for the state bodies, including the Election Commission of India, to avoid doxxing of citizens. Meanwhile, the Government of India refuses to bring in surveillance reform, continues to protect the bureaucracy by weakening the RTI, and exempts government bodies from laws that should apply to them, and our courts rarely address these concerns.
