The 80s nostalgia holds caste bias

1. At a Glance

• Core argument: India's viral AI-made "1980s nostalgia" images come from savarna, urban or landed sources: film stills, magazine spreads, studio portraits and family albums. The trend therefore reproduces caste bias and leaves out the Dalit experience of that decade [1].
• Algorithmic bias: Tests by MIT Technology Review found that GPT-5 picked the stereotypical answer in 80 of 105 sentences. In those answers the "clever man" was upper caste and the "sewage cleaner" was Dalit [1].
• Historical contrast: The same decade saw the Karamchedu massacre (1985) and led to the SC/ST (Prevention of Atrocities) Act, 1989 [1][4].
• UPSC value: The topic links GS-I (caste, social change), GS-II (vulnerable sections, protective laws, SC-Parliament relations), GS-III (AI governance) and GS-IV (algorithmic ethics).

2. Why in the News

• An op-ed by Rejimon Kuttappan, a workers' rights advocate, appeared in The Hindu on 29 September 2026 (Chennai print edition, p. 9) [1].
• India became the No. 1 country for Google's Nano Banana image-generation model. Its signature trend in India was retro-Bollywood portraits of the 80s: chiffon sarees, Ambassadors and Contessas [1].
• The article places this trend beside evidence of caste stereotyping in large language models (GPT-5, per MIT Technology Review tests) [1].

3. Background & Evolution

• 1980: Only about a quarter of households lit their homes with electricity [1].
• 1983: The Planning Commission estimated that 44.5% of Indians, about 323 million people, lived below the poverty line [1].
• 17 July 1985: Karamchedu massacre (coastal Andhra Pradesh) [1]:
• A Madiga woman objected to a Kamma youth soiling the public drinking-water tank her family used.
• By nightfall six Madiga men had been killed and three Dalit women raped.
• The police called it a "riot". A civil liberties fact-finding team called it a "one-sided massacre".

• The killings led to the Andhra Pradesh Dalit Mahasabha.

• 1989: Parliament enacted the Scheduled Castes and the Scheduled Tribes (Prevention of Atrocities) Act, 1989 (Act No. 33 of 1989) [1][4].

• 20 March 2018: In Dr. Subhash Kashinath Mahajan vs State of Maharashtra (Criminal Appeal No. 416 of 2018), the Supreme Court issued directions that diluted the Act [2][3].
• 2018: Parliament passed the SC/ST (PoA) Amendment Bill, 2018. The President assented on 14 August 2018 [2][3].
• 2025–26: Generative-AI image trends such as Nano Banana retro portraits went viral, and caste bias in LLMs was documented [1].

4. Core Static Facts

Item	Fact
Principal law	SC and ST (Prevention of Atrocities) Act, 1989, Act No. 33 of 1989 [4]
Key amendment	SC/ST (PoA) Amendment Act, 2018; assent 14 Aug 2018 [2][3]
Section inserted in 2018	Section 18A [3]
What S.18A(1) says	No preliminary enquiry is needed to register an FIR. No approval is needed to arrest an accused. No procedure other than the Act or CrPC applies [3]
Section restored	Section 18 of the Act (bar on anticipatory bail provisions) [3]
Judgment overturned	Subhash Kashinath Mahajan v. State of Maharashtra, 20 Mar 2018, Crl. Appeal 416/2018 [2][3]
Poverty benchmark cited	44.5%, about 323 million BPL in 1983 (Planning Commission) [1]
AI bias metric	GPT-5: 80/105 stereotypical choices (MIT Technology Review) [1]
AI trend	India is No. 1 country for Google's Nano Banana [1]

5. Multi-Dimensional Analysis

Social

• Nostalgia is selective. The visual record of the 1980s (films, magazines, studio photos, albums) belonged to households with money and standing, mostly savarna, urban or landed [1].
• When about 44.5% of people were below the poverty line and only about 25% of homes had electricity, most people were not in the archive. The Dalit poor are therefore missing from the "aesthetic" [1].
• Social media turns this exclusion into a mass-produced "memory".

Scientific / Technological

• AI outputs reflect their training corpora. Skewed source material produces skewed images and text [1].
• The GPT-5 test (80/105 stereotyped completions) shows representational harm: caste hierarchy is encoded in the associations between jobs and intelligence [1].
• India's scale of adoption (No. 1 for Nano Banana) magnifies the effect of this bias [1].

Legal / Constitutional

• The PoA Act, 1989 is India's special criminal law against caste atrocities [4].
• After the SC's 2018 Mahajan dilution, Parliament restored the Act's strict procedure through Section 18A [2][3].
• Karamchedu raises a four-decade question: does the Constitution reach inside the village? [1]

Ethical / Governance

• Karamchedu was labelled a "riot" rather than a massacre. This shows how state narratives can hide caste violence [1].
• Algorithmic accountability matters: bias audits and representative datasets decide whose past gets "remembered".

Historical

• Karamchedu, the AP Dalit Mahasabha and the PoA Act, 1989 form a sequence in which atrocity led to mobilisation, which led to legislation [1][4].

6. Recent Developments (last 12–18 months)

• 2025–26: India became the top country for Google's Nano Banana model, and retro-Bollywood 80s portraits became its signature trend [1].
• 2025–26: MIT Technology Review tests documented caste stereotyping in GPT-5: 80 of 105 sentences [1].
• 29 Sep 2026: The Hindu op-ed "The 80s nostalgia holds caste bias" was published [1].

7. Prelims Hooks

• The SC/ST (Prevention of Atrocities) Act is Act No. 33 of 1989 [4].
• The Karamchedu massacre took place on 17 July 1985 in coastal Andhra Pradesh [1].
• Karamchedu victims were Madiga Dalits. The attackers belonged to the Kamma community [1].
• The trigger at Karamchedu was a dispute at a public drinking-water tank [1].
• Karamchedu led to the Andhra Pradesh Dalit Mahasabha [1].
• Section 18A was inserted into the PoA Act by the 2018 Amendment [3].
• Section 18A: no preliminary enquiry is needed before an FIR, and no approval is needed before arrest [3].
• The 2018 Amendment answered Subhash Kashinath Mahajan v. State of Maharashtra (20 March 2018) [2][3].
• Presidential assent to the PoA Amendment Act, 2018 came on 14 August 2018 [2][3].
• Planning Commission, 1983: 44.5% (about 323 million) of Indians were below the poverty line [1].
• In 1980, only about one-quarter of Indian households used electricity for lighting [1].
• Nano Banana is Google's image-generation model, and India is its No. 1 country [1].
• In MIT Technology Review tests, GPT-5 gave stereotyped caste answers in 80 of 105 sentences [1].

8. Why a Strong Law Still Loses Most Cases in Court

• The number of cases waiting for trial keeps growing
• Pendency means cases that are registered but whose trial has not finished.
• Under the PoA Act, 1,77,379 cases were waiting for trial at the end of 2020 [5].
• By 2024 this had grown to 2,87,694, the highest in five years [5].

• So courts finish fewer cases each year than the number of new ones that arrive. Every year the queue gets longer.

• About two out of every three finished trials end without punishment

• The conviction rate is the share of finished trials in which the accused is found guilty.
• For cases involving Scheduled Castes it was 34.1% in 2022, 32.4% in 2023 and 33.9% in 2024 [5].

• The number of convictions did go up, to 5,192 in 2024 [5]. But the rate stayed almost flat, so the law is not getting better at finishing cases.

• Too few special courts for the size of the backlog

• The Centre gives money to states under a Centrally Sponsored Scheme (a scheme where the Centre and states share the cost) to set up Exclusive Special Courts for these cases [6].
• Only 189 such courts have been set up [6].

• If every pending case went to these courts, each court would carry about 1,500 cases [5][6]. A fast trial is not possible with that load.

• Section 18A fixed the first step, not the later ones

• Section 18A made FIR registration and arrest easier [3].
• But a victim's case can still fail later: when the police describe it, during investigation, or during a trial that takes years.

• At Karamchedu, the police called a massacre a "riot" [1]. The 2018 amendment does not stop that kind of labelling.

• What should follow: the Ministry of Social Justice and the states should link special-court funds to how many cases each court finishes

• Today the scheme counts how many courts are set up [6].
• The real test is whether the 2,87,694-case backlog gets smaller [5].

9. The Case for the Mahajan Judgment, and Why Parliament Rejected It

• The strongest argument on the other side
• In Mahajan, the Supreme Court said there should be a preliminary enquiry before an FIR and approval before an arrest [2][3].
• Supporters of the judgment point to the low conviction rate. About 66% of finished trials end without punishment [5].

• Their argument is that if most accused are acquitted, arresting people without any check can hurt innocent people. For them, the Act can be used as a weapon in personal quarrels.

• What is right about this argument

• An arrest made without an enquiry cannot be undone, even if the person is later acquitted.

• Any criminal law with strong powers can be misused in some cases. Admitting this makes your answer stronger, not weaker.

• Why the argument still does not hold

• An acquittal (the court setting the accused free) does not prove that the complaint was false.
• A case can fail because trials take years, because the investigation is weak, or because the victim gives up. In 2024, 2,87,694 cases were still waiting [5].
• The conviction rate cannot tell a false complaint apart from a failed prosecution. So it cannot justify putting a check at the FIR stage.

• Karamchedu shows the opposite danger. The police themselves played down the violence by calling it a "riot" [1]. A preliminary enquiry would give that same police force one more chance to stop a case early.

• A fair way to write the conclusion

• Parliament chose to protect the victim at the entry point by inserting Section 18A [3].
• The honest answer to the fear of misuse is faster trials, not a harder path to filing an FIR.

10. Is the Machine Biased, or Only the Old Photos?

• The defence: "the AI only copies what exists"
• This argument says the tool invents no bias. It only reflects the old photos and texts it learned from.

• That part is true. Training data (the material an AI learns from) decides what the AI produces [1].

• Why "only copying" still causes harm

• In 1980 only about a quarter of homes had electric lighting [1]. In 1983, 44.5% of Indians were below the poverty line [1].
• Cameras, studio portraits and magazine photos belonged to a small, better-off group. Most of them were savarna, urban or landed [1].

• When millions of people make "80s" images from that narrow record, the missing majority goes missing a second time. India is Nano Banana's No. 1 country [1], so this happens at a very large scale.

• The same pattern appears in text AI

• GPT-5 chose the caste stereotype in 80 of 105 test sentences [1].
• This is called representational harm: an AI shows a group in a narrow or insulting role again and again, so that role starts to look normal.

• The link "sewage cleaner = Dalit" is not random. It repeats a real pattern of caste-based work, and the AI makes it look natural.

• The takeaway for an answer

• AI did not create caste bias. But it multiplies bias that was already there.
• It also removes the context. A photo filter carries no note about who was left out of the pictures.

11. What India's AI Rules Cover, and What They Leave Out

• The rules name the problem
• MeitY released the India AI Governance Guidelines on 5 November 2025 [7].
• One of their seven guiding principles is "Fairness & Equity". They say AI must avoid discrimination, especially against marginalised communities [7].

• They also say global risk models may not fit India's social situation. So India needs its own risk assessment (a check of what could go wrong, and for whom) [7].

• But the rules mainly target high-stakes uses

• The guidelines are risk-based (stricter rules for more dangerous uses) and proportional [7]. They limit the use of high-risk AI systems [8].
• Their examples of risk focus on areas like welfare delivery, where an AI tool decides who gets benefits [7].
• A retro-portrait filter or a chatbot's word choices are unlikely to count as high-risk. So the kind of bias this article describes mostly falls outside the strict part of the rules.

• One principle is "Innovation over Restraint" [7]. This tells regulators to avoid heavy rules where the harm is not clear.

• Caste is not part of global bias tests

• Most global fairness tests look at gender and race.

• The India-specific risk assessment in the guidelines [7] is the right place to add caste. Tests like the MIT Technology Review exercise (80/105) [1] show this can be measured.

• What should follow: MeitY and the IndiaAI Mission should make caste-bias testing a standard check

• Large AI models used widely in India would need to publish results of caste-association tests, like the GPT-5 sentence test [1].
• This would put the guidelines' own "Fairness & Equity" principle [7] into practice. It would not need a new law.
• UNESCO's 2021 Recommendation on the Ethics of AI, already listed in the note, sets the same non-discrimination standard at the global level.

12. Anchors for Answers

• Data: 2,87,694 PoA Act cases waiting for trial in 2024, up from 1,77,379 in 2020 [5]
• Data: Conviction rate for SC-related PoA cases was 33.9% in 2024, so about two in three finished trials end in acquittal [5]
• Data: Only 189 Exclusive Special Courts set up for PoA cases [6]
• Data: GPT-5 chose caste stereotypes in 80 of 105 test sentences (MIT Technology Review) [1]
• Report/Committee: India AI Governance Guidelines, MeitY, 5 November 2025. Principle: "Fairness & Equity"; calls for India-specific risk assessment [7]
• Law/Case: SC/ST (PoA) Act, 1989; Subhash Kashinath Mahajan v. State of Maharashtra (2018); Section 18A inserted by the 2018 Amendment [3][4]
• Scheme: Centrally Sponsored Scheme funding Exclusive Special Courts under the PoA Act. Its money should be linked to how many cases each court finishes [6]

13. Mains Relevance

• GS-I: Salient features of Indian society; caste; social empowerment; effects of globalisation and media on society.
• GS-II: Mechanisms, laws and institutions for the protection of vulnerable sections; separation of powers (SC ruling vs. Parliament's amendment).
• GS-III: Science and technology in everyday life; AI and emerging technologies.
• GS-IV: Ethics of technology; fairness, accountability and transparency in algorithmic systems.
• Possible question stems: 1. "Generative AI does not create bias; it inherits it." Examine this with reference to caste representation in Indian digital culture. (GS-I/III, 250 words) 2. Trace the evolution of the SC/ST (Prevention of Atrocities) Act, 1989, including the 2018 legislative response to judicial dilution. How far has it achieved its objectives? (GS-II) 3. What is "selective nostalgia"? Discuss how the portrayal of the past in mass media shapes present-day social hierarchies. (GS-I)

14. Related Topics to Study Next

• Article 17 and untouchability laws (Protection of Civil Rights Act, 1955): the constitutional basis of anti-caste law.
• Manual scavenging and sewer deaths: the "sewage cleaner = Dalit" stereotype reflects a real occupational pattern.
• Anti-caste movements in Andhra Pradesh (Karamchedu, Chundur 1991): regional history of Dalit mobilisation.
• Poverty estimation in India (Planning Commission, Lakdawala/Tendulkar/Rangarajan methods): context for the 1983 BPL figure.
• AI governance in India (IndiaAI Mission, Responsible AI principles): regulation of algorithmic bias.
• UNESCO Recommendation on the Ethics of AI (2021): the global standard on fairness and non-discrimination.
• NCRB "Crime in India": data on crimes against SCs and STs.
• Judicial–legislative relations: the Mahajan ruling and the 2018 amendment as a case study of Parliament overriding a judgment.

15. Common Errors / Trap Areas

• Year of the Act: The PoA Act is from 1989, not 1985 (the year of Karamchedu) and not 1955 (the Protection of Civil Rights Act) [1][4].
• Section 18 vs 18A: Section 18A was inserted in 2018. Section 18 is the original provision that it restored [3].
• Karamchedu vs Chundur: Both happened in Andhra Pradesh. Karamchedu is 1985 (Madigas vs Kammas). Chundur is 1991 and a different incident [1].
• Mahajan case date: The judgment was 20 March 2018. Assent to the amendment came on 14 August 2018. Don't mix them up [2][3].
• Source of the AI finding: The 80/105 GPT-5 result is from MIT Technology Review tests, not a government audit [1].