Govt should consider national database for AI incidents: White paper

The PSA has proposed a national database to record AI incidents such as safety failures, bias and security breaches, with reports from public bodies, private firms, researchers and civil society

The government should consider having a national database to record, classify, and analyse safety failures, biased outcomes, security breaches and misuse of artificial intelligence (AI), the office of the Principal Scientific Advisor (PSA) has suggested.

 

In a white paper on the techno-legal measures for AI governance in the country, the PSA’s office has suggested that the national database should have reports on such incidents submitted by public bodies, private entities, researchers, and civil society organisations.

 

“Within techno-legal framework, this database will enable post-deployment accountability through measures such as India-specific risk taxonomy, detection of systemic trends and emerging threats, data-driven audits and targeted regulatory interventions, and evidence-based refinement of technical and legal controls,” the white paper has suggested.

 

Apart from this, the industry should publish transparency reports, conduct regular fairness and robustness testing, perform security reviews, and have red-teaming exercises, the white paper has said.

 

“These steps allow organisations to build familiarity with compliance processes and documentation before requirements become mandatory. Voluntary measures also help identify risks early, improve preparedness, and develop sector-wide capacity for responsible AI deployment,” the white paper has suggested.

 

To support the functioning of the AI governance group, which is chaired by the PSA, the Ministry of Electronics and Information Technology should set up a dedicated Technology and Policy Expert Committee (TPEC), which could pool multidisciplinary experts from law, public policy, machine learning, AI safety, cybersecurity, and public administration, among other things, the PSA white-paper has suggested.

 

A techno-legal framework that is likely to be adopted by the government must also carefully balance privacy safeguards with inclusion, equity, and model utility considerations, the white paper has said.

 

In practice, privacy, fairness and the performance of AI models often exist in “structural tension”, which means that strengthening one aspect could lead to automatic weakening of the others. For a country that is as linguistically and demographically diverse as India, these decisions should not be left to implicit engineering decisions, the white paper has suggested.

 

“To mitigate these risks, impact-aware data withdrawal mechanisms are needed rather than unconditional erasure. Large-scale unlearning requests should be subject to fairness and representativeness impact assessments, with safeguards triggered where data removal risks degrading performance for underrepresented groups,” the white paper has said.