The 1985 Karamchedu massacre in Andhra Pradesh exposed deep caste violence. Today, AI models that generate 1980s‑style images repeat the same upper‑caste bias because their training data largely ignore Dalit and Adivasi lives. This raises urgent questions about data transparency, labeling practices, and the adequacy of existing government guidelines.
Key Developments
- Google’s Gemini produces 1980s images that show only affluent, upper‑caste settings.
- Tests by MIT Technology Review found GPT‑5 repeats caste stereotypes in 80 of 105 sentences.
- The AI Governance Guidelines acknowledge bias but rely on voluntary codes.
- The government’s IndiaAI Mission funds “sovereign” models without clear bias‑testing mandates.
- Calls are growing for a regulatory committee to make caste a mandatory audit parameter.
Important Facts
- In 1983, 44.5% of Indians (≈323 million) lived below the poverty line.
- Only about 25% of households had electricity in 1980.
- The SC/ST (Prevention of Atrocities) Act was passed four years after the Karamchedu incident.
- Dalit and Adivasi photographs exist mainly in state welfare files, activist archives, or anthropological studies, not in commercial image pools.
- Labelers for AI datasets are often low‑paid South Asian workers who may never have seen a Dalit colony.
Exam Relevance
Understanding AI bias links to multiple GS papers. MeitY’s guidelines illustrate how technology policy is framed within constitutional values of equality (Article 14). The caste‑based data gap highlights challenges in implementing the SC/ST (Prevention of Atrocities) Act in the digital age. Aspirants should analyse how public‑sector funding (IndiaAI Mission) can be aligned with ethical AI, social justice, and inclusive development.
Way Forward
Instead of merely filtering outputs, policymakers must address the data source. A national audit should answer: what images are in training sets, who labels them, and has a caste‑bias evaluation been conducted? Funding community photo archives of Dalit, Adivasi, Muslim, and working‑class families (1975‑1995) can diversify datasets. Mandatory public disclosure of bias‑testing results will ensure accountability and help AI reflect India’s true social fabric.