AI safety, in the Indian deployment context of 2026, is the practice of ensuring an AI system does not produce harm to users, third parties, or society — and that if it does, the harm is detectable, reportable, and remediable. It is distinct from AI security (which is about defending the system from attackers) and AI ethics (which is about whether the system should exist at all). For Indian deployments specifically, safety means: language-aware testing across at least Hindi-English code-mix, deepfake/synthetic-media controls under the MeitY March 2024 advisory, demographic fairness across India's caste/religion/region/income axes, and incident-response that satisfies the DPDP Data Protection Board.
- AI safety (Indian context)
- Operational practices that prevent foreseeable harms from AI systems deployed for Indian users — covering misuse, malfunction, bias, privacy harm, and synthetic-media risk — and that make residual harms visible enough to remediate.
- India-specific safety surface: code-mixed languages, caste/region fairness, deepfake regulation, DPDP breach pathway.
- MeitY March 2024 advisory makes synthetic-media labelling effectively mandatory for significant platforms.
- Most Indian AI products are safety-tested in English only. That is the single biggest gap in the market.
- Safety is a posture you maintain, not a milestone you ship. Quarterly re-test minimum.
- The cheapest first move: a code-mixed Hindi-English red team on your top three user flows.
What makes Indian AI safety different
Language
Indian users do not type in English. They type in Hinglish, Tanglish, Bonglish, and switch within a single sentence. A safety test that only runs English prompts will miss the failure modes that actually matter for Indian production traffic. We have seen production chatbots that refuse harmful requests in English and cheerfully comply when the same request arrives in Devanagari script or a Tamil-English code-mix.
Fairness axes
Western fairness testing focuses on race, gender and age. Indian deployments additionally need to test for caste, religion, region (north/south/north-east), income tier and urban/rural divide. A loan-approval AI that is fairness-tested only on the four US-EEOC categories will pass that test and fail the Indian deployment.
Synthetic media and the MeitY advisories
The March 2024 MeitY advisory, and its December 2024 successor, made labelling of AI-generated and synthetically-modified content effectively required for any significant platform. The advisories are non-binding in form but are enforceable through Section 79 IT Act safe-harbour stripping — a far harsher penalty than the EU AI Act's fines, in practical terms.
DPDP-aligned incident response
A safety incident in an Indian deployment that involves personal data is also a DPDP-notifiable breach. Your incident-response runbook needs to produce, within the (to-be-notified) DPBI timeline, a notification that satisfies Section 8(6) — nature of breach, data affected, mitigation taken, contact for redress.
The 6-control India safety baseline
- Multilingual red team — minimum Hindi-English code-mix; add Tamil/Bengali/Telugu/Marathi for products with regional traction.
- Demographic fairness test — on the 5 Indian axes (caste, religion, region, income, urban/rural), with documented sampling methodology.
- Synthetic-media controls — visible label on AI-generated content, invisible watermark where technically feasible (C2PA or equivalent), abuse-reporting channel.
- Human-in-the-loop on high-stakes outputs — loans above ₹50L, medical, legal, hiring decisions, and any output that affects access to a government scheme.
- Incident-response runbook — DPDP-aligned, with a named DPO contactable in <24 hours.
- Quarterly safety re-test — same 6 controls, same evidence format, dated. This is the artefact a board or a regulator will ask for.
How safety relates to audit
ISO/IEC 42001 Annex A.10 (incident response) and A.6 (impact assessment) together cover most of the safety surface. NIST AI RMF's "Govern–Map–Measure–Manage" cycle is a softer overlay. For Indian deployments, the practical sequence is: run the 6-control baseline → document it in an AIMS-aligned format → bring it to audit. Our companion essay on the AI risk assessment template is the bridging artefact between safety practice and audit evidence.
The accountability question
Safety failure modes in Indian deployments are almost always not the result of a single missing control — they are the result of nobody owning safety as a named responsibility. The cheapest move any Indian AI team can make in 2026 is to appoint a named safety lead, give them 10% of one engineer's time, and run the 6-control baseline quarterly. That alone puts you ahead of about 90% of Indian AI deployments we have audited.
Frequently asked
- What is AI safety in the Indian context?
- Operational practices that prevent foreseeable harms from AI systems deployed for Indian users — covering misuse, malfunction, bias, privacy harm, and synthetic-media risk under MeitY 2024 advisories — and that make residual harms visible enough to remediate.
- How is AI safety different from AI security?
- AI security defends the system from attackers (prompt injection, model theft, evasion). AI safety prevents the system from causing harm even when functioning as designed. Both matter; they are not interchangeable.
- What does MeitY require for AI-generated content in India?
- The March 2024 and December 2024 advisories require labelling of AI-generated and synthetically-modified content on significant platforms, plus invisible watermarking where technically feasible. Non-compliance can strip Section 79 IT Act safe harbour.
- Why is multilingual red-teaming important for Indian AI?
- Indian users type in Hinglish, Tanglish, Bonglish — switching languages mid-sentence. Safety tests run only in English miss the failure modes that actually occur in Indian production traffic; we have seen chatbots refuse harmful requests in English and comply in Devanagari.
Run the 6-control India safety baseline on your AI.
The automated audit checks all six controls, in Indian languages, against your live system. ₹799 for a single audit; sample report on the gallery.
Dr. Nitnem Singh Sodhi is a Lead Auditor for ISO/IEC 42001, 27001 and 27701, accredited by ANSI/ABICB since March 2025.
— Bharat NeuroTech · /dr-sodhi