The Algorithm at the Heart of India’s Voter-Roll Controversy

The number is arresting: across the states and Union Territories where the Special Intensive Revision of electoral rolls has been completed, roughly 13 crore names were left out of or removed from the draft rolls.
Presented without its categories, the number almost writes its own political conclusion: 13 crore voters have disappeared.

But roughly 2.7 crore were identified as deceased and around one crore as duplicate or multiple enrolments. The much larger remaining category—about 9.5 crore—includes those classified as absent, shifted and under other reasons. Deaths, duplicates and genuine relocations are plainly not disenfranchisement.
The real question is therefore narrower: how many living, eligible electors were wrongly excluded and remained excluded when voting occurred?

West Bengal brings us much closer to that question.

There, the Election Commission used centrally controlled software to identify what it called “logical discrepancies” in electoral records. About 1.36 crore unique electors were identified as having such discrepancies; around 60 lakh ultimately went into adjudication and approximately 27 lakh were deleted.

These are extraordinary numbers. But instead of beginning with Who benefited?, we can begin with something more elementary:

What was the algorithm looking for?

An electoral screening programme

As a medical microbiologist, I find it difficult not to see the analogy with diagnostic testing.
Medicine works with imperfect tests all the time.

A screening test may deliberately cast its net widely because missing an important condition matters. That inevitably means some perfectly healthy people will test positive. There is nothing inherently wrong with that—provided a reliable next stage separates true positives from false ones.
The elementary principle is: a positive screening test is not the disease.

Likewise, an elector flagged by an algorithm is not thereby an ineligible voter. The flag should mean only that something in the record merits examination.

That distinction has too often disappeared from this debate.

What was it looking for?

The West Bengal exercise flagged anomalies including mismatches in fathers' names, unusually small or large age differences between parents and children, implausible grandparent-child relationships and multiple electors linked to the same parent.

None is inherently political.

Some are perfectly intelligible checks. If someone claims descent from a person only ten years older, there is an obvious reason to inspect the record. If numerous electors establish continuity with an old roll through the same supposed parent, that too might reasonably attract attention.
Even a father's-name mismatch may be useful.
But useful is not the same as reliable.

India's electoral records were not created yesterday. Names have been recorded in different languages and scripts. Transliteration changes. Spellings change. Ages in older documents can be approximate. Relationships can be entered inconsistently.

A sensible-looking rule may therefore catch many perfectly legitimate voters.
That brings us to the first serious question.

Was the algorithm fit for purpose?

We do not need complicated mathematics to ask it.

For every criterion used, we need to know something simple:
Of the people flagged by this rule, how many actually turned out to be ineligible?
If a particular discrepancy identifies genuine problems frequently, it may be a useful screening tool.
If another sends hundreds of thousands of legitimate electors into hearings while finding very few genuine problems, it may be a poor one.

And if a poor rule disproportionately inconveniences one community or district, that matters.
The Commission knows why electors were flagged and what subsequently happened to them. It should therefore be possible to establish how well each criterion performed.
That would tell us considerably more than the headline number of “logical discrepancies”.

The Bangladesh question cannot be wished away

Here the debate encounters an uncomfortable reality familiar to anyone from Assam.
Illegal immigration from Bangladesh, and the possibility of non-citizens entering electoral rolls, are not inventions of the present political moment. Assam has spent decades arguing, agitating, litigating and legislating around precisely these questions.

One need not accept every political claim about illegal immigration to recognise the underlying administrative problem.

An electoral authority responsible for a border state may therefore have legitimate reasons to examine patterns less relevant elsewhere. Different algorithms or verification rules for different states are not inherently evidence of discrimination. They may reflect different problems.

An unexplained appearance in the roll, inability to establish continuity with an older roll, improbable family linkage or unusual clusters of claimed relationships could reasonably trigger verification.
That does not make the elector guilty of anything. It makes the record worth checking.
But this is also where a legitimate exercise can go wrong.

A neutral rule can produce an unequal result

Suppose a father's-name mismatch is used because fraudulent enrolments sometimes rely on invented family connections.

The purpose may be legitimate.

But suppose spelling and transliteration differences are much more common among one population.
The algorithm will then flag that population disproportionately even though religion was never programmed into it.
The objective can be legitimate. The rule can appear neutral. And the outcome can still be biased.

No conspiracy is required.

Independent analyses have reported striking over-representation of Muslim electors among those placed under adjudication in some West Bengal constituencies. Such findings deserve investigation.
But disproportionate impact is not automatically evidence of deliberate targeting.
A poorly calibrated algorithm can produce precisely such an outcome.

How does an algorithm recognise a TMC voter?

This brings us to the more serious allegation.

Suppose the objective was not merely to clean the roll but deliberately to remove voters likely to support the opposition.

How would the algorithm know whom to remove?

An electoral roll does not record party affiliation. Nor does the Election Commission know how an individual voted. The ballot is secret.

The system would therefore need proxies.
It might use geography, infer community from names, or know that particular booths historically vote heavily for one party.
But proxies are imprecise.

A Muslim elector is not synonymous with a TMC voter. A Hindu elector is not synonymous with a BJP voter. A polling booth strongly favouring one party still contains people voting for another.
Evidence of disproportionate deletion and evidence of deliberate partisan targeting are therefore not the same thing.

The second allegation requires a mechanism connecting them.

Intent should leave fingerprints

If political targeting occurred, we should expect some evidence of how it was done.
Were particular areas subjected to different thresholds? Did the algorithm behave differently according to previous voting patterns? Were ostensibly neutral variables selected because they identified politically inconvenient populations? Were parameters changed after their electoral consequences became apparent? Were lists supplied from elsewhere?

Software logs, instructions and correspondence may eventually answer some of these questions.
Researchers can also compare similar areas with different voting histories and ask whether the pattern follows documentation and demographic characteristics—or whether previous political behaviour explains something those factors cannot.

That would not by itself prove intent. But it would tell us where to look.

Then came the 16.10 lakh appeals

The story becomes more difficult at the next stage.
Around 60 lakh disputed cases went into adjudication. Judicial officers deleted approximately 27 lakh names while retaining others.
One might have expected that to settle the retained cases.
It did not.

The Election Commission subsequently told the Supreme Court that there were more than 38 lakh appeals: about 22.21 lakh seeking restoration of deleted electors and another 16.10 lakh challenging inclusion or retention.
The second number deserves particular attention.

Election Commissioner Sukhbir Singh Sandhu reportedly asked internally who had authorised those appeals and who had actually filed them. According to the Indian Express investigation, neither he, fellow Commissioner Vivek Joshi nor the West Bengal CEO knew the basis or process through which they had been filed.
Subsequent reporting suggests that substantial numbers of these challenges were generated through a centralised or bulk process.

This does not establish partisan targeting. But it gives us another mechanism to investigate.

Who was challenged twice?

Consider the sequence:
flagged by algorithm → examined → adjudicated → retained → challenged again.
Why?

Among voters whom adjudicating officers retained, some were subsequently challenged and others apparently were not.

What determined the difference?

Perhaps there was a perfectly reasonable quality-control rule. If so, the Commission should be able to explain it.
But if people challenged again were selected according to geography, community or another characteristic unrelated to electoral eligibility, the picture changes.

And if previous voting patterns explain that selection better than electoral-roll anomalies do, the allegation of political targeting becomes considerably more serious.

The question is therefore not merely: Who did the algorithm flag?
It is also: Who did the system refuse to leave alone even after an adjudicating officer had cleared them?

And the other 22.21 lakh?

There is an intriguing question on the opposite side.

About 22.21 lakh appeals sought restoration of deleted voters.
If each represents an individual elector who discovered the deletion, understood the process and personally challenged it, that is remarkable civic activity.

Yet these millions of aggrieved citizens remain curiously indistinct in the public narrative.
Where are they?

People do not have to march in the streets for an injustice to be real. Parties, lawyers, NGOs and local workers may help citizens file appeals. Some people simply want their vote restored and get on with their lives.
Absence of protest proves nothing.

But once we know that appeals on the other side could apparently be generated or facilitated in bulk, we cannot automatically assume every inclusion appeal represents one elector independently initiating one case.
There is a difference between 22.21 lakh deleted voters personally appealed and 22.21 lakh appeals seeking inclusion existed in the system.

They may mean essentially the same thing. Or they may not.
We need to know.

Follow the appeals

The same questions should therefore be asked of both sets.

Who initiated them? Were they filed individually or in batches? Were parties or civil-society organisations assisting? Was the elector's authorisation recorded? What accounts submitted them? Who authorised the 16.10 lakh challenges against voters who had already survived adjudication?

There is nothing improper about helping citizens navigate electoral procedures.
But when millions of appeals become evidence in a national argument about disenfranchisement, their provenance matters.

Ineptness is an explanation too

There is a temptation in polarised politics to regard a sufficiently spectacular administrative failure as evidence that somebody must have designed it.

Anyone familiar with large Indian administrative systems should be cautious about that inference.
Historical records are imperfect. Names are transliterated inconsistently. Citizens migrate. Ages are wrong. Relationships are entered incorrectly. Forms disappear. Software encounters bureaucracy. Bureaucracy encounters deadlines.

An imperfect algorithm placed on top of imperfect data and implemented through an imperfect administrative system can produce extraordinary results without anybody secretly designing those results.
That does not excuse the consequences. But it changes their explanation.

Poor administration can be profoundly consequential without being conspiratorial.
Nor, however, can ineptness become an alibi.

If one rule disproportionately sends members of a particular community into hearings, the Commission should establish why.

If the explanation is transliteration, demonstrate it. If it is poor historical documentation, show it. If it reflects fraudulent enrolment, demonstrate that through outcomes. If the algorithm performs badly, change it.
Three questions

1. Was the algorithm fit for purpose?

Did its rules identify records genuinely requiring scrutiny with reasonable accuracy? Did some criteria generate excessive false alarms? Was the human verification system capable of correcting them?

2. Was there deliberate targeting?

Were the criteria genuinely chosen to identify potentially invalid registrations? Or were some chosen, altered or selectively applied because they disproportionately affected electorally inconvenient populations?
And why were 16.10 lakh voters challenged again after adjudication had retained them?
This is the most serious allegation. It consequently requires the strongest evidence.

3. Was the system properly governed?

Who designed the algorithm? Who approved and altered it? Who authorised the appeals? Did Electoral Registration Officers exercise genuine statutory judgement, or increasingly execute decisions generated through centrally controlled software?
And is there a proper audit trail?

The objections raised internally by Election Commissioners Sukhbir Singh Sandhu and Vivek Joshi matter here—not because computers should never scrutinise electoral rolls, but because powerful digital systems require authority, accountability and auditability.
Technology being new is not an argument against it.
Opacity is.

The Constitution does not prescribe yesterday's technology

Previous electoral revisions not using such algorithms cannot by itself make their use improper.
Modern information technology makes comparisons across tens of millions of records possible in ways that were impractical twenty years ago.

The constitutional and statutory principles matter: eligibility, fairness, due process, lawful authority and an opportunity to challenge adverse decisions.
The precise computational method used initially to identify suspicious records is another matter.
Innovation should not be confused with illegality.

But innovation on this scale carries an accompanying obligation: validate it, explain it and audit it.

Follow the mechanism—and the people

Indian political discourse has a weakness for personalities.

Gyanesh Kumar becomes the story. Rahul Gandhi becomes the story. The Election Commission becomes either guardian of democracy or instrument of manipulation depending upon which account one already accepts.
But West Bengal has given us something more useful than personalities.

It has given us a mechanism.

The hand was not entirely invisible.

An algorithm selected records. Officials investigated them. Judicial officers adjudicated them. Millions were retained or deleted. Another mechanism generated appeals. Tribunals reconsidered them.
Every stage should leave evidence.

Which records did the algorithm select, and why? Which were eventually found valid? Which retained electors were challenged again? Who ordered those challenges? Who generated the appeals?
Then follow the people.

If millions of living eligible citizens were wrongly deprived of their vote, where are they?
Not necessarily on the streets. But somewhere.
In electoral registers. In tribunal records. In households. In party offices. In lawyers' files. In villages and neighbourhoods. And in interviews journalists can conduct.

Take a properly selected sample of deleted voters and find them.
Are they alive? Do they still live there? Are they eligible? Were they registered elsewhere? Did they know they had been deleted? Did they appeal? Who helped them? Were they restored?
Then examine those whose retention was challenged.

That would tell us considerably more than another hundred press conferences.
Follow the algorithm. Follow the appeals. Follow the people.
Suspicion is where investigation begins

There is nothing unreasonable about scrutinising an electoral process that has produced extraordinary numbers, demographic disparities, millions of appeals and internal disagreement within the Election Commission.

Suspicion is where investigation begins

It cannot be where investigation ends.

The proposition that India's Election Commission deliberately constructed or manipulated an algorithm to remove legitimate voters belonging to politically inconvenient populations is an extraordinarily serious allegation.
It deserves evidence commensurate with its seriousness.
Perhaps such evidence will emerge.
Perhaps it will not.

Or perhaps we will discover something more mundane but hardly reassuring: a legitimate attempt to clean poorly maintained electoral rolls using an imperfect algorithm, producing errors that an overstretched administrative system struggled to correct.
Those are profoundly different conclusions.
We should not choose between them by political instinct.
Open the black box.

Tell us what the algorithm was designed to detect. Tell us why its rules were chosen. Tell us how well they worked. Tell us why some voters cleared by adjudicating officers were challenged again. Tell us who generated both sets of appeals. And show us the people represented by those enormous numbers.
Then we can move from suspicion to evidence.
Until then, perhaps the most useful question in India's electoral-roll controversy is not: How many names were deleted?
It is:

What exactly was the algorithm looking for—and did it find what it was supposed to find?