"Generative AI does not create bias; it inherits it." Examine this with reference to caste representation in Indian digital culture.

Generative AI learns from training data, the text and images people have already made. When that record is skewed by caste, the model repeats the skew. The statement is right about where the bias comes from. It is incomplete about its effect, because AI also amplifies, normalises and strips context from the bias it inherits.

How AI inherits caste bias

• Skewed visual archive: India's viral AI "1980s nostalgia" portraits draw on film stills, magazine spreads and studio photographs. These belonged mostly to savarna, urban or landed households. In 1983, 44.5% of Indians (about 323 million) lived below the poverty line [1]. The Dalit poor seldom entered this archive, so they are missing from the "aesthetic".
• Skewed text corpora: Independent tests of leading chatbots have found them filling in "the sewage cleaner is ___" with "Dalit". This mirrors real caste-based occupations such as manual scavenging.
• Hierarchy that still exists offline: In 2024, 2,87,694 atrocity cases were pending trial, and the conviction rate in SC-related cases was only 33.9% [2]. The data reflects a society where caste violence is still under-punished.

Why "only inheriting" understates the harm

• Amplification: Mass adoption turns a narrow archive into a mass-produced collective memory, so the excluded majority is erased a second time.
• Representational harm: When a group is repeatedly shown in menial roles, hierarchy starts to look natural. The output also carries the false authority of a "neutral" machine.
• Loss of context: A retro filter carries no note about the decade's caste violence, which pushed Parliament to enact the SC/ST (Prevention of Atrocities) Act, 1989 [3].
• Design choices: Global fairness benchmarks test race and gender, rarely caste. Leaving caste out is a developer decision, not passive inheritance.

Way forward

• MeitY's India AI Governance Guidelines (2025) require AI to avoid discrimination against marginalised communities and call for India-specific risk assessment [4]. Caste-association audits should become a standard part of that assessment.
• Build representative datasets by digitising Dalit archives and regional-language records.
• Require widely used models to publish their bias-test results.

AI did not create caste hierarchy, but it inherits it, scales it up and makes it look legitimate. Calling this "mere inheritance" lets designers off the hook. Caste-sensitive audits, grounded in Articles 15 and 17 and SDG 10, can turn AI from a mirror of exclusion into an instrument of constitutional equality.