A sourced record of the labs, organisations, programmes and researchers building AI safety, interpretability, evaluation and governance capacity in India, and of the diaspora and Global South networks they connect to. Every entity carries what it is known for, where it sits, and who is affiliated with it.
States and union territories coloured by the sector leading their trustworthy-AI work, with every mapped entity plotted on top. Hover a state to pin its record.
Hover any state to pin its entities here; the panel updates as you move across states. Click a marker for the full record — once clicked, the panel holds that entity until you hit back or click elsewhere on the map. Entities that share a campus fan into rings, and a dashed line runs back to the true coordinate where a cluster had to be nudged clear of its neighbours.
SRC · compiled by Secure AI Futures Lab from institutional pages, funder announcements and event rosters · boundaries from Survey of India district data, dissolved and simplified
Pick one to filter the directory below.
Search indexes names, institutions, topic tags and every affiliated person, so a name that appears only inside a roster still returns the entity that lists it.
Moved IIT Delhi → IIT Kharagpur in 2024 (previously MPI-SWS).
Fairness in algorithmic decision making, social computing, AI for social good. Algorithmic auditing of Amazon under the EU Digital Markets Act; few-shot fairness in LLMs; whether LLMs can predict elections in India.
Hallucination, alignment and watermarking; developing a 'Civilizing AI' agenda that trades off LLM/VLM usability against vulnerability to adversarial behaviour, and treating alignment as instruction-driven and runtime-controllable — a natural-language behavioural contract rather than a fixed training-time property. AAAI Senior Member, class of 2026. Formerly Research Associate Professor at the AI Institute of South Carolina, which makes him a live edge between the AIISC cluster on this map and Indian academia. AI Alignment core track at the Agyeya workshop.
HateXplain; guardrail vulnerability studies; multilingual safety (Soteria) across high, mid and low-resource languages.
Human–AI interaction, socially responsible recommendation, robustness and incentives in federated learning. Ex-FAIR; AI2050 Early Career Fellow 2025.
CeRAI core member; co-teaches the NPTEL Responsible & Safe AI Systems course; co-organiser of the Deployable AI workshop at AAAI-26.
Co-author on the 'Towards Measuring and Modeling Culture in LLMs' survey with Monojit Choudhury; legal NLP.
Moved from IIT Patna to IIT Jodhpur; verify current affiliation.
MemeDetoxNet — balancing toxicity reduction against context preservation (ACL 2025); multimodal harm.
India's most connected trustworthy-AI node. India co-chair of the Safe & Trusted AI Working Group for the India AI Impact Summit 2026 (with Brazil and Japan); serves on the Government of India's advisory group on AI; on the committee authoring the FREE-AI report for the RBI; appointed to the UN Independent International Scientific Panel on AI (Feb 2026).
Co-organiser of the Deployable AI workshop at AAAI-26 with Rajkumar and Ramaswamy.
Leads the IIT Delhi computer vision group. Moved from MRF-MAP optimisation into adversarial attacks and the reliability of deep networks, with a specific line on calibration: deep networks are uncalibrated and make overconfident mistakes, which is the binding constraint on deploying them for critical tasks like medical imaging. AI4GS participant.
Detecting AI-generated content, curbing unhealthy reliance on AI, geo-cultural representation, holistic LLM evaluation. AI2050 Early Career Fellow. Runs the AI Alignment core track at the Agyeya workshop at IISc; AI4GS participant.
Adversarial ML, side-channel recovery of transformer architectures in shared-GPU MLaaS, encrypted computation.
CeRAI core member; theory of learning with structured losses, explanation methods.
Deepfake detection and responsible datasets; lead on the IndiaAI Saakshya deepfake governance project and the Itisaar verification portal.
Multimodal harm detection and safety in Indian-language settings; LCS2 co-lead.
Indic model and evaluation infrastructure underpinning most Indian-language safety benchmarks.
Frequent co-author across the Kharagpur fairness and hate-speech cluster.
Co-author on SurakshaEval, an Indic safety benchmark spanning ten Indian languages.
NLP; frequent co-author on the Kharagpur fairness and multilingual safety line.
Responsible AI, interpretability, unlearning and alignment; associate faculty fellow at CeRAI and adjunct at IIIT-Delhi. Becomes President of ACM India for the 2026–28 term. Co-teaches the NPTEL Responsible & Safe AI Systems course and sits on the programme committee of the Agyeya interpretability-and-safety workshop at IISc.
Federated and collaborative learning with differential privacy.
Fairness in speech recognition (Amazon Research Award); multilingual and code-switched speech.
One of very few India-based researchers working the agentic-safety framing directly rather than the fairness/audit one. Leads Human-Aligned Intelligent Learning Systems: what is the minimal set of requirements in an AI system that lets a user assess and understand the limits of its safe operability — autonomous capability assessment of black-box sequential decision-making systems (NeurIPS 2023), differential assessment of black-box AI agents (AAAI 2022), action-model learning and analysis of abstractions. Postdoc with Julie Shah in MIT's Interactive Robotics Group before IIT Madras. Invited speaker at the Agyeya workshop.
The adversarial-robustness anchor of the Bengaluru cluster. Chairs the Department of Computational and Data Sciences and leads VAL (formerly the Video Analytics Lab): adversarial attacks and adversarial training, guided adversarial attack as a way of evaluating rather than just defeating adversarial defences, and certified robustness within multiple perturbation bounds. Complements the CeRAI/IISc evaluation agenda from the vision side, where most Indian trustworthy-AI work is NLP-first. On the Agyeya programme committee.