Caste in the Machine: What AI's Nostalgia for the 1980s Leaves Out of the Frame

Source: Original editorial ↗ The Hindu

Resources for this Article

Download PDF Print-ready A4 notes · Free

The Lift Line

“The machine did not invent this prejudice; it inherited it and now industrialises it.”

Why This Editorial Matters for Your Exam

Rejimon Kuttappan, a workers’ rights advocate and writer, connects a viral AI trend to the social history of caste. The piece gives a ready GS1 example of how technology interacts with social stratification, a GS3 hook on AI governance, and a GS4 case on fairness and accountability in algorithms. It also works for an Essay on technology and society.

GS Paper 1: Salient features of Indian society; social empowerment; caste. GS Paper 3: Science and technology developments (AI). GS Paper 4: Ethics in public administration; fairness, accountability.

Background and Context

Karamchedu, 1985. On 17 July 1985, at Karamchedu in coastal Andhra Pradesh, an argument at a drinking-water tank ended with six Madiga men killed and three Dalit women raped by dominant-caste men, the author writes. The police called it a riot; a civil liberties team called it a one-sided massacre. It led to the Andhra Pradesh Dalit Mahasabha.

The constitutional and legal frame.

Provision	What it does
Article 15(2)	No citizen may be denied access to shops, public restaurants, hotels, places of entertainment, or the use of wells, tanks, bathing ghats, roads and places of public resort, on grounds including caste
Article 17	Abolishes untouchability; its practice is an offence
Protection of Civil Rights Act, 1955	Punishes practice of untouchability (originally the Untouchability (Offences) Act)
SC and ST (Prevention of Atrocities) Act, 1989	Defines atrocities and sets up special courts; enacted 1989, four years after Karamchedu

India’s AI governance, as the author describes it.

Instrument	Status
India AI Governance Guidelines (MeitY, November 2025)	Name bias and discrimination as risks; rely on voluntary measures and self-certification
Horizontal AI law	The Centre told the Rajya Sabha none is needed “at this stage”, the author writes
IndiaAI Mission (approved March 2024)	Outlay about Rs 10,371 crore; compute, datasets, foundation models
IndiaAI Safety Institute	Set up under the Mission to study AI risks and testing

The Analysis

1. Nostalgia is an archive, and archives have owners. Film stills, magazine spreads, studio portraits and family albums from the 1980s belonged overwhelmingly to urban, landed or savarna households. In 1983, 44.5 per cent of Indians lived below the poverty line by the Planning Commission’s estimate, and in 1980 only about a quarter of households had electricity.

2. Who was photographed, and by whom. Dalit and Adivasi lives were photographed, but by others: the state for welfare files, activists after atrocities, anthropologists. The family photograph, taken “simply to be seen”, was the one the poor could least afford.

3. Models reproduce and harden the silence. The author cites MIT Technology Review tests in which GPT-5 chose the stereotypical answer in 80 of 105 sentences, and a study at the ACM FAccT conference of 1,536 images from Gemini’s image model, prompted with Indian names only, in which caste surfaced through food, neighbourhood, work and worship.

4. The labelling layer is invisible. Nobody outside the companies knows what is in the training sets or who labelled them, often low-paid South Asian workers. “Can a labeller who has never seen a colony flag a model that has never rendered one?”

5. Governance is voluntary. The guidelines acknowledge bias but set no test for caste, while ministers promise that Indian-trained models will be free of bias without saying how that is measured.

6. Four questions, and a positive remedy. The author wants MeitY and the IndiaAI Safety Institute to answer publicly what is in the training data, who labelled it, whether a caste-bias evaluation has been done and whether the results will be published. And instead of banning filters, “go upstream”: fund community photo archives of Dalit, Adivasi, Muslim and working-class families as seriously as film restoration.

Data and Institutions Vault

Prelims-grade facts:

Constitution and law:

• Article 15(2): equal access to wells, tanks and bathing ghats, among others.
• Article 17: untouchability abolished; enforced through the Protection of Civil Rights Act, 1955.
• SC and ST (Prevention of Atrocities) Act, 1989: special courts; amended in 2015 and 2018.

History:

• Karamchedu massacre: 17 July 1985, Andhra Pradesh; six Madiga men killed.
• Mahad Satyagraha (1927): B.R. Ambedkar led Dalits to drink from the Chavdar tank.

AI governance:

• India AI Governance Guidelines: MeitY, November 2025; principle-based, voluntary.
• IndiaAI Mission: approved March 2024, about Rs 10,371 crore.
• ACM FAccT: the Conference on Fairness, Accountability and Transparency.

⚠️ Watch the trap: access to public wells and tanks is protected by Article 15(2); Article 17 abolishes untouchability itself.

The Debate

For the author’s view. Bias in systems used by millions scales discrimination, and voluntary codes leave no one accountable. Public answers on data and testing are a low bar.

The complications. Full disclosure of training data may clash with trade secrets and copyright; caste-bias benchmarks are still being developed; heavy compliance could slow Indian start-ups while foreign models face no equivalent scrutiny.

The balanced verdict. Start with systems used in public functions (welfare, policing, education, recruitment): mandatory bias testing with published results, audits by the IndiaAI Safety Institute and a grievance channel. Build open Indian datasets with community participation so that better data, not just rules, fixes the archive.

How to Think About This

Ask who made the record. When a question involves data, AI or history, ask who produced the source, for whom, and who was left out. The answer usually explains the bias before any algorithm is involved.

Diagram-in-Words

Takeaway Box

• Hook: AI-generated “1980s” images show a savarna India.
• Cause: an archive owned by the privileged; opaque training data and labelling.
• Evidence: caste-stereotyped outputs in independent tests.
• Gap: MeitY’s guidelines are voluntary.
• Remedy: disclosure, caste-bias audits, community photo archives.

Sources: The Hindu, IndiaAI Mission, Ministry of Electronics and Information Technology

Source: Caste in the Machine: What AI's Nostalgia for the 1980s Leaves Out of the Frame — Ujiyari.com | Free UPSC & State PCS Editorial Analysis

Key Arguments at a Glance

Rejimon Kuttappan argues that AI-generated nostalgia for 1980s India reproduces an upper-caste archive, because the photographs, films and magazines models learn from belonged to privileged households, and that India’s voluntary AI governance must be replaced by mandatory, public answers on training data, labelling and caste-bias testing.

✓ Supporting

• The Karamchedu massacre of 17 July 1985, which began at a drinking-water tank, is the kind of 1980s reality absent from the retro images now filling social media feeds.
• In 1983, 44.5 per cent of Indians lived below the poverty line, and in 1980 only about a quarter of households had electricity; family photography was a privilege.
• Tests reported by MIT Technology Review found GPT-5 choosing the stereotypical answer in 80 of 105 sentences, and a study at the ACM FAccT conference found caste surfacing in Gemini images through food, neighbourhood, work and worship.
• MeitY’s AI Governance Guidelines name bias as a risk but rely on voluntary codes and self-certification, while the IndiaAI Mission funds sovereign models.

⚠ Counter

Mandatory disclosure of training data may clash with trade secrets and slow Indian start-ups; bias audits need agreed benchmarks that do not yet exist for caste; and over-regulating a fast-moving field could leave India dependent on foreign models that are no better.

→ Way Forward

Write caste explicitly into forthcoming AI rules; require developers and the IndiaAI Safety Institute to disclose training data sources, labelling practices and caste-bias evaluations; and fund community photo archives of Dalit, Adivasi, Muslim and working-class families.