Who Gets to Be the Past? Why Neutral Machines Still Remember Unequally

The Machine Remembers What the Album Kept

Every society argues about its past. The quieter question is who supplied the evidence in the first place. That question, once a matter for historians, is now being settled by software.

A tank, a massacre and a fight over the record On July 17, 1985, in Karamchedu in coastal Andhra Pradesh, a Madiga woman objected to a Kamma youth fouling the public water tank her family used.

By nightfall, six Madiga men had been killed. The official account called it a riot; a civil liberties fact-finding team called it a one-sided massacre. That gap between two words was itself a struggle over memory: whether the event would be recorded as a clash between equals or as violence by the powerful against the powerless.

The struggle had consequences. Karamchedu gave rise to the Andhra Pradesh Dalit Mahasabha and within four years Parliament enacted the SC/ST (Prevention of Atrocities) Act, 1989, creating a legal category for exactly the kind of violence the police had declined to name.

The archive nobody chose, but everybody inherited The India of the 1980s that survives in pictures comes mostly from film stills, magazine spreads, studio portraits and family albums.

In 1983, by the Planning Commission's estimate, 44.5% of Indians lived below the poverty line. The family photograph, the one taken simply to be looked at, was a luxury. Dalit and Adivasi lives were photographed, but usually by others: the state for its files, activists after an atrocity, anthropologists for their studies. The album belonged to those who could afford it.

Enter the

model Generative AI learns from whatever record exists. Fed an archive of the comfortable, it returns comfort as the national past. Asked to reason about people, it can return prejudice: MIT Technology Review found GPT-5 choosing the caste-stereotyped answer in 80 of 105 test sentences.

No engineer typed that bias in. It arrived with the data. The pattern is not uniquely Indian. For decades, colour film was calibrated against reference photos of light-skinned models, so darker skin rendered poorly. The chemistry was "neutral"; its assumptions were not.

What follows India's response so far is thin.

MeitY's AI Governance Guidelines name bias as a risk yet depend on voluntary codes, even as the ₹10,371 crore IndiaAI Mission funds sovereign models promised to be bias-free. A promise is not a test. At minimum, developers should disclose what their training data contains, who labelled it and whether a caste-bias evaluation was done.

But audits only catch errors after the fact. The deeper fix is authorship. Films like Fandry, Kammattipaadam, Pariyerum Perumal and Vaazhai show what changes when the person remembering has lived on the other side of the water tank.

Community photo archives from Dalit, Adivasi, Muslim and working-class families deserve funding as serious as film restoration. A nation's memory is only as fair as the hands that made its record. Machines have simply made that old truth faster.