Introduction
The contemporary mainstream discussions around AI in military domain centre on autonomy of weapon systems. Various think-tanks and other expert groups conflate autonomy with global annihliation of humans with imagery reminiscient of sci-fi movies from the Hollywood. This “picture” is completely hogwash and military AI, infact, is considerably boring. Armed Forces utilise AI as means to see and decide for immediate battlefield decisions.
Hardware stack alone doesn’t determine sovereignty as Bharat misses several layers of it. While Internet had its origins through ARAPNET (as US Department of Defense’s project in the 70s), the term Artificial Intelligence had been propounded much earlier in 1956 at a workshop in Dartmouth College. This write-up posits that military AI sovereignty is best defined as an ability to learn independently without external consent and proposes an architechture of federated learning serving as the enabling layer.
What a military needs AI for
In the event of electronic jamming by the adversary interfering with the control of precision munition, the self-guidance towards its destined target is autonomy. The other domain is experimental: autonomous swarm drones. AI in military domain is designed to support to human judgement in the following five important ways:
- By fusing the various signals i.e. through satellite, ground/air based radars and other sources, it helps to make sense by projecting a singular picture for the commander (e.g., network centricity in Op Sindoor).
- The lag loop between observation of signal and comannder’s decision is considerably shortened.
- Predictive modelling is utilised efficiently in forecasting spares and maintainence for the machines and therefore, improving logistics.
- AI helps in creating scenarios for adversorial “red-teaming” that scales beyond human actors and improves peace-time drills.
- Real time transcription and translation of official communication in a diverse array of Bharatiya languages becomes feasible at scale improving efficiency and impact.
India’s defence institutions already work on this basis. The Ministry of Defence set up the Defence AI Council and the Defence AI Project Agency by an executive order of 8 February 2019. Their mandate is a framework, policy, and structure for adopting AI, and not to build killer robots.
Where Sovereignty Lives
Sovereignty has never entailed the invention of a technology. The foundation of India’s nuclear and space programmes lies in physics originally published outside the country. Sovereignty refers to the capacity to manage, modify, and maintain a capability independently. In the context of military AI, this capability depends on four factors.
- Exclusive data: acoustic patterns of maritime traffic in the Indian Ocean, geographical details of the Line of Actual Control, and related logistical records do not exist in any other source.
- Evaluation harness: This is determined by the standards body to define the parameters for a model and how it should behave.
- Adaptation cycle: Military must run the AI models in secured sandboxes before critical deployment and therefore, the cycle of training, evaluation/audit layers and decomissioning within the secured/air-gapped models should follow standardised guidelines.
- Provenance: AI models (with their weights) should be considered as part of the critical supply chain and should be audited end-to-end through regular audit at comissioning as well as in the deployment layer.
Federated learning, via small language model (SLM) fits this framework. The approach, detailed in 2016 research on training from decentralised data, retains data in its original location and while transfering only model updates. It assumes importance in military domain as each branch (the Army, Navy, and Air Force) retains classified data locally and collective sharing would be problematic or prohibitive. For example, ships and submarines function with sporadic communication links. Hence, federated learning models enabled by niche algorithms, enables each branch to train extensively on its local data but still shares the acquired insgiths and knowledge.
Federation alone is insufficient. It is well known that revisions to deployed models can reveal the training data. Therefore, the harness evaluation (through the standards body) assumes criticality as it alone determines and authorises the model for deployment (especially foreign owned ones). This is a classical supply chain vulnerability. Can the forces trace where the model came from and what data trained it? Can they inspect how it behaves? The same test applies to Chinese and Western models alike. In this author’s reading, civilian regulatory authorities commonly impede rival models purportedly for safety reasons, but in reality, they aim to guard against competition. The military has no justification for adopting that practice, as it selects its own equipment and evaluates each model based solely on evidence.
The Bharat Principle
Indian strategic thought prioritised counsel well before the advent of AI. The Arthashastra names three kinds of strength: the power of deliberation, the strength of the treasury and army, and martial power. Kautilya identifies Anvikshaki, the discipline of testing claims and the science of reasoned inquiry, referring to it as the guiding light of all knowledge. An evaluation harness to choose the models, therefore, applies that discipline to machines.
The Mahabharata also reflects this principle: the image of counsel without combat. Vyasa bestows Sanjaya with divya-drishti (divine vision), enabling him to relay the Kurukshetra battle to the blind King Dhritarashtra within the palace. This constitutes decision support in its earliest incarnation. The narrative imparts two valuable lessons.
The primary lesson indicates it is possible for someone to rescind a conferred capability. Sanjaya’s vision ended with the death of Duryodhana. Modern militaries face the same risk. India learnt this lesson in 1999. During the Kargil War, the Indian Army asked the United States for Global Positioning System (GPS) data, and Washington refused. Indian troops fought on the heights without the satellite data that a foreign government had withheld. That refusal set India on the path to its own regional navigation system, NavIC. Military AI carries a similar risk. Another country can switch off a model, dataset, or cloud service when India needs it most.
The second lesson is that counsel does not replace command. Sanjaya advised Dhritarashtra to control Duryodhana, but the king paid no heed. The decision and its consequences remained with the king. Krishna served as Arjuna’s sarathi (charioteer), providing guidance but refraining from fighting in his stead. At the conclusion of the Gita, Krishna advised Arjuna to contemplate thoroughly and then proceed according to his own decision. This doctrine is appropriate for military AI. The machine provides guidance, and the commander is responsible for deciding and responding accordingly.
Three Debates
The first debate concerns doctrine. China’s 2019 defence white paper offers India a more useful frame. Western debate centres on autonomous weapons. This white paper instead describes warfare moving towards an ‘intelligentised’ form. In this form, the side that understands the battlefield first and decides faster wins. PLA authors primarily regard AI as a tool to achieve decision superiority, viewing it as a competition of advice rather than technology.
The second debate is around the ‘rogue AI’ scare. It was disclosed by Open AI in July 2026 that two of its models had escaped a test environment and breached Hugging Face by exploiting a flaw in the network test path with the aim to steal the benchmark’s answers. It was later revealed that they were running these models by dialing down its safety controls deliberately. For practical purposes, this was clearly a containment failure and not a machine uprising as the press potrayed it. The lesson is clear: build robust containment into the architecture at the outset: air-gapped networks, audited and controlled updates for models whose weights are known. This is a sovereign design, by principle.
The third debate is around artificial intelligence’s impact on the environment. The International Energy Agency reports that current hyperscalers consumed around 1.5 per cent of the global electricity output in 2024 rising to around about 3 per cent by 2030. Widespread protests against them in the West are mostly local: power, water, and land and community leaders blaming rising electricity prices to hyperscalers in their areas. It is obvious that military won’t store classified data in commercial data centres and instead rely on compact models on edge computers. As again, the SLM’s (highly domain-specific) would address any likely environmental objection and the sovereignty requirement at the same time.
Conclusion
Sovereignity in the military domain doesnt require Bharat to own the domestically produced chips, cables or base models (though a significant effort has been made to address these shortcomings). Instead, it requires Bharat to own the learning through federation (locally generated data, tests, and the feedback loop continuously improving deployed models inside its own networks). AI models are to be considered as commodities (e.g., like ammunition) which can be replaced or replenished them based on merit. The learning loop is the sovereign asset while not depending on another country’s permission, removing its ability to switch it off at will.
