India’s AI Revolution Must Speak Every Indian Language

The technological feasibility about localising LLMs (Large Language Models) isn’t just restricted to the pure technical question of if a model can speak a language well, but also whether Indian public institutions are equipped to procure, regulate and sustain it. The Indian states cannot accept such a structure in the first place, it does go through rules and procedures such as: General Finance Rules, GeM-based procurement procedures and tender conditions made for specifying qualification criteria, pricing logic, data hosting and compliance obligations. The rules are a direct determinant about defining the government’s capacity to buy an AI system as a commodity, commission it as a service, or insist on local hosting, security controls, and vendor accountability.

Currently, India doesn’t have one single dedicated AI-model statute governing its procurement, but procurement is already being used as an instrument of regulation. So, the recent AI empanelment efforts also point to an emerging localisation logic: AI services can’t be separated from Indian data centres and are tied to compliance with domestic data-security and IT rules.

What the Rules Actually Say: Technological Capacity vs. Administrative Feasibility

According to the General Financial Rules’ Rule 149, central government ministries and departments must use the GeM platform to procure goods and services with the use of a graded system based on value thresholds and competitive pricing. Along with this, Rule 144 read with procurement principles expects efficiency, fairness and competition in public buying. In principle, AI can be produced as a service, but only with fitting in procurement arrangements that can define performance, hosting, security, and accountability without vendor lock-in permission. The IndianAI tender reveals that the government has been trying to localise AI using cloud hosting, compute etc like conditions. But a more grave challenge is whether this procurement can endure open, contestable and governable over time. The understanding leads this judgement to another policy risk, meaning even if the government may technically obey procurement rules, it can still become dependent on a single vendor’s model, which makes later correction costly as well as difficult.

To encapsulate, the institutionalization of administration has become an unequivocal imperative because the result of not doing that won’t be inefficiency, but a governance gap in which complex systems are incorporated into public

administration more rapidly than the state can effectively regulate them. There are various other operational bottlenecks like even if a department wants to introduce a localised model, it frequently encounters difficulties in accessing interoperable datasets, harmonising technical standards and establishing algorithmic accountability for errors. Consequently, inter-departmental coordination transcends administrative concern.

Also Read – Why NEET Leaks Perpetually, But IIT-JEE & CA Tests Don’t (Part 1)

Language Choice in AI: Never Neutral

The localisation has never become the truly neutral technical fix because of the ingrained language politics continuing from the past. In multilingual societies, the languages aren’t merely supported by the state but rather hierarchised by formal chains of bureaucratic command, public administration tends to prioritise procedural compliance making it more cumbersome. Scholarship on indigenous AI and data sovereignty argues that, in absence of community governance, development of language technologies risks perpetuating data colonialism. Apparently, linguistic resources are extracted from indigenous communities, while concentrating authority and equitable ownership disproportionate. This raises a crucial question of “accountability” because language policy in the context of artificial intelligence transcends it and also becomes a fundamental contest over epistemic sovereignty, power and autonomy, not just accessibility.

A second challenge is about integral bias in the propensity of AI systems to disproportionately privilege standardized, codified and officially sanctioned languages over their local ones. In the highly multilingual context of India, characterised by profound lingual diversity, it can systematically reproduce hierarchies in language making indigenous languages and oral traditions subservient. Ultimately, mere expansion of language coverage can’t be called inclusion, the critical inquiry would be whether AI systems genuinely redistribute power or not.

Data Sovereignty, Community Participation and Accountability

The collection of linguistic data from communities cannot be considered as a neutral exercise as it becomes a governance question about ownership, usage, controls correction, and responsibility bearing when an AI model misinterprets cultural or linguistic identities? While scholarship on Indigenous data, as argued above, has consistently highlighted potential misuses surrounding the collection, access and governance of the data. The CARE Principles in this regard, standing for Collective, Benefit, Responsibility, and Ethics present a normative configuration that re-orients AI governance away from extractive data practices and towards community-led

stewardship. This perspective has been further reinforced by contemporary work of UNESCO on artificial intelligence and cultural heritage, underscores that culturally responsible AI must privilege community agency to whom that knowledge belongs, it will ensure that the cultural representation remains accountable. Ensuring community ownership, ethical data governance and institutional accountability establishes the legitimacy of multilingual AI systems. Yet legitimacy alone is insufficient unless these principles translate into measurable improvements in public service delivery and administrative outcomes. Accordingly, the next challenge is determining how the success of multilingual AI should be evaluated beyond purely technical benchmarks.

Measuring Success Beyond Translation Accuracy

Integration of AI in public administration was meant to ease the flow of public service delivery to the citizens. So, the success of multilingual AI in public administration cannot be judged solely by technical metrics such as translation accuracy or model performance as these do not capture whether an AI system succeeds in real administrative settings. A multilingual chatbot that translates perfectly but fails to help citizens access welfare schemes cannot be considered a successful public-sector AI system. Therefore, the evaluation metrics need to revolve around the capability of the system to enhance administrative accessibility, improve service delivery, strengthen citizen trust and promote equitable participation of each linguistic community.

Conventional approaches to determine success are confined to computational performances only such as translation accuracy, response fluency and the pace of translation through AI models. These criteria tell little about the administrative proficiency of these models. Public administration pursues fundamentally different objectives which include equity, accessibility, accountability and the scale of citizen welfare. Consequently, an AI system may excel on technical benchmarks while failing to improve governance outcomes.

Thus, rather than treating AI evaluation as an isolated technical exercise, the assessment should emphasise the following five parameters:

This framework moves evaluation from technology-centric benchmarks to tangible governance outcomes. It aligns with global thinking from the OECD and UNESCO, which stress that trustworthy AI in the public sector must demonstrate real improvements in service delivery, equity and public confidence thereby producing transparent, accountable and sustainable administration.

For instance, India’s BHASHINI (Bhasha Interface for India) initiative offers early signals: multilingual tools in panchayat platforms and state partnerships have enabled voice-based access in rural areas but their ultimate value will be proven only through these governance metrics and not model accuracy alone. Policymakers must ask whether BHASHINI has improved citizens’ access to government services, narrowed linguistic exclusion and strengthened trust in digital governance.

This shift from technical to administrative evaluation paves the way for examining how different countries have built institutional frameworks to implement and govern multilingual AI in practice.

Comparative Governance Models: Lessons for India

While a governance-centric evaluation framework defines what success looks like, comparative institutional experience depicts the process of achieving it in practice. By examining how leading digital governments have structured their systems for excellent deployment of multilingual AI in public services, India can draw actionable lessons to strengthen its own administrative capacity, regulatory coherence and inclusive implementation.

Among these, Estonia offers one of the most standout models. Rather than building solely on technological innovation, Estonia embedded AI within an already mature digital governance ecosystem through robust institutional coordination, interoperable digital infrastructure and citizen-orientated framework.

Key institutional practices include:

• ●  Improving existing system rather than replacing it: Estonia integrated Bürokratt, its government-wide AI virtual assistant with the existing X-Road interoperability platform and digital public services, thereby avoiding presence of any parallel systems. This made the system interoperable, scalable and seamless.
• ●  Central coordination network: A dedicated AI programme manager tasked with handling AI projects across ministries, maintaining visibility over ongoing initiatives, preventing duplication of efforts and promoting institutional learning through regular inter agency engagement.
• ●  Break administrative silos through knowledge sharing: Monthly webinars, cross agency consultations and collaborative forums enabled ministries to exchange implementation experiences and avoid reinventing existing solutions.
• ●  No compromise with privacy and public trust: Estonia retained government regulation over core digital infrastructure while inculcating private sector innovation, ensuring that citizen data remained shielded and public confidence in AI enabled services was maintained. While Estonia showcases how important interoperability and institutional coordination are for AI integration in administration, Singapore offers valuable lessons in adaptive and risk based AI governance. With the deployment of the Infocomm Media Development Authority’s (IMDA) Model AI Governance Framework, Singapore adopts a comprehensive approach focused upon four pillars of upfront risk assessment, meaningful human accountability, technical safeguards and end-user responsibility. Rather than relying solely on rigid regulation, it follows a pilot-first approach, thereby scrutinising the system on a smaller scale at the initial level; a continuously monitored and sector-specific approach, enabling innovation while ensuring transparency, accountability and public trust. Such a governance model is particularly relevant for multilingual AI systems deployed in high-impact public services where errors can have significant administrative ramifications. Collectively, these experiences reflect that successful multilingual AI is fundamentally an institutional project instead of merely a technological one. For India, these models suggest that the breakthrough of multilingual AI will ultimately rest on giving weight to administrative coordination, ensuring regulatory coherence, safeguarding citizen trust and inculcating AI within existing digital public infrastructure rather than treating it as a standalone technological pillar. Building on these comparative lessons, it is high time for India to translate institutional best practices into targeted administrative

reforms that can enable multilingual AI to become a reliable instrument of inclusive public service delivery.

Policy Recommendations: Building a Multilingual AI Governance Framework for India

India’s BHASHINI initiative has laid an essential foundation for multilingual digital governance by enabling AI powered language translation across public services. However, technological progress alone cannot guarantee inclusive and effective service delivery. To realise the full potential of multilingual AI, India must complement BHASHINI with robust institutional, regulatory and administrative reforms.

Despite its promise, significant technical, linguistic and operational challenges remain. While languages such as Hindi and Tamil tend to perform well due to the availability of their massive volume of digital data, tribal and regional languages (e.g., Bodo, Santali, Kodava) lag behind with higher translation error rates due to a lack of digitised parallel text corpora. Moreover, concerns persist regarding long-term data privacy frameworks, downstream commercial consent, and explicit language community rights

To address these challenges, India should prioritise following administrative reforms:

1. Strengtheningcoordinationatthenationalscale:

ANationalMultilingual AI Coordination Cell can be established under the ambit of the IndiaAI Mission or a body under MeitY. Modelled on NITI Aayog’s coordinating role, the cell can function as a nodal institution linking ministries, state governments and technical agencies. It ensures non-duplication of efforts and sharing of regional or state level practices that can be implemented on a national scale.

2. IntegratingAIwithexistingdigitalpublicinfrastructure:

Drawing from the success of interoperability in Estonia’s governance model, Indian administration should move towards integrating AI with existing digital public infrastructure with BHASHINI, UMANG, DigiLocker, Aadhar, eDistrict and CPGRAMS. This ensures that AI becomes an additional administrative layer and not just another isolated platform.

3. AdoptingariskbasedAIgovernanceframework:

Building on the model of Singapore, India should institutionalise a risk based AI governance framework with mandatory human oversight for high-impact administrative decisions. Definite steps and measures should be codified with

tiered risk classification. This should be layered with continuous monitoring over the mechanisms and impact assessment, constructed through audit trails.

4. BuildingAdministrativeCapacity:

No new initiative would ever be successful until the operators are familiar with the know-how of the system. Therefore, capacity building and AI literacy programmes for civil servants become imperative. Not only this, but there should also be readily available assistance from the respective language experts and interdisciplinary implementation teams to smoothen the flow of service delivery. Additionally, as the mechanisms tend to evolve over time, periodic training of the administrators should be ensured. This enables the timely implementation of integrated technology.

5. Institutionalisingcommunityparticipationinlanguagedevelopment:

Since linguistic knowledge resides within local communities, governments should partner with universities, tribal councils, language academies and civil society organisations to create, validate and continuously update datasets for low-resource languages. Such a participatory approach would enhance model quality while ensuring that linguistic communities retain agency over their cultural and linguistic knowledge. Additionally, funding in a dedicated pool for the digitisation and development of low resource languages through collaborations with universities, language academies and community organisations can enable more communities to cherish the success of this remarkable initiative.

6. Institutionalisingoutcomebasedevaluation&independentaudits:

Citizen engagement and feedback loops should form the core of the implementation process. There should be an annual evaluation considering accessibility, grievance resolution, citizen satisfaction, inclusivity, trust and explainability of the system. Existing grievance platforms such as CPGRAMS can be leveraged to institutionalise citizen feedback, enabling continuous evaluation of multilingual AI systems while keeping implementation costs low.

In contemporary times, India’s linguistic diversity cannot be viewed as a barrier to AI-enabled governance but as an opportunity to redefine digital inclusion. If bolstered by coordinated administration, interoperable digital infrastructure, accountable governance structures and meaningful citizen participation, multilingual AI initiatives can transform India’s multilingual AI from just a translation tool into a game-changing instrument of equitable, accessible and citizen-centric public administration. Ultimately, India’s AI revolution will succeed not when it speaks every language perfectly but when every citizen can access the state with equal ease, dignity and trust.

About The Authors – (Ms. Ritika Swami is a final-year Political Science student at Lady Shri Ram College for Women. While, Nitish Narang is a final-year Political Science student at Atma Ram Sanatan Dharma College, University of Delhi.) 

Disclaimer—(The views and opinions expressed in this article are solely those of the author and do not necessarily reflect the views of Indian Masterminds. For feedback or queries, please write to [email protected].)

Also Read – Possible Solutions To Paper Leaks (Part 2)