The government has asked teams building large IndiaAI foundation models to tighten and standardise the way they test for social bias, after internal reviews found that some systems were reproducing harmful stereotypes on caste, gender, religion and region. Developers backed by the IndiaAI Mission have been told to run more rigorous pre‑deployment bias audits, use evaluation datasets that reflect India’s diversity, and document how they will monitor and correct biased outputs over the life cycle of the model.
This push is linked by officials to the new India AI Governance Guidelines that place demands on high‑risk and general‑purpose models to have safeguards against discrimination, transparency about limitations, and clear accountability for harms. Funded projects on bias detection and mitigation under the “Safe & Trusted AI” programme of IndiaAI can be expected to ensure common tools and benchmarks so that scaling indigenous LLMs and generative models does not entrench social inequalities.
