ABSTRACT
This paper looks at the challenges that come with state-funded foundational artificial intelligence models in India. It does this by looking at the country laws and regulations. The paper starts by examining the role of the Information Technology Act, 2000 the Digital Personal Data Protection Act, 2023 the Indian Penal Code, 2023 and the Copyright Act, 1957 in shaping the environment for artificial intelligence. The study finds some challenges, including unclear liability and accountability tension between data protection principles and training data needs, algorithmic bias and violations of constitutional rights, copyright and ownership conflicts and cross-border data flows and compute sovereignty. All these challenges come from the fact that Indias laws were made before artificial intelligence became big. The paper says that not having a law for artificial intelligence leaves state-funded models open to unclear rules and gaps in accountability. It ends with suggestions for an approach to artificial intelligence governance that aligns with Indias constitutional commitments to equality, dignity and due process.
INTRODUCTION
The world is seeing a change in the digital landscape moving from software services to artificial intelligence. Foundational models, which are scale artificial intelligence systems trained on huge datasets have become the new general-purpose technology of the 21st century. For a country like India that is developing its economy these models offer a big chance for growth and a challenge to national digital sovereignty. The Government of India started the IndiaAI Mission in 2024 with a big budget to develop foundational models that meet the country unique needs. This initiative includes projects like BharatGen and the IndiaAI Innovation Centre, which aim to reduce the country reliance on models. However the fast deployment of state funded intelligence models raises complex legal and ethical questions. As the state becomes a developer and funder of intelligence traditional frameworks of accountability are being tested. This research paper looks at the evolution, funding and regulation of state-funded artificial intelligence models in India. It examines the touch regulatory approach adopted by the Ministry of Electronics and Information Technology and assesses whether current regulations are enough to manage the risks of artificial intelligence. The paper argues that India needs a legal framework to ensure that these models remain a public good while protecting against the risks of automated decision-making.
STATE FUNDED FOUNDATIONAL AI MODELS
Foundational artificial intelligence models are large-scale machine learning systems trained on diverse datasets. They are designed to serve as a general-purpose base for downstream applications. Unlike narrow artificial intelligence systems foundation models are built to exhibit broad capabilities that can be adapted across multiple domains.
Core technical characteristics
Foundational models are big both in terms of their training datasets and internal parameters. They are trained on corpora drawn from sources like the open web, digitized books and public-domain archives. This allows them to learn patterns, structures and semantic relationships across a spectrum of human knowledge. The resulting models are usually neural networks, which enable them to attend to long-range dependencies in text and other sequential data. A key feature of foundation models is their two-stage development pattern: first they undergo a pre-training phase on large unstructured datasets, where they learn general linguistic, visual or logical patterns. After pre-training the model can be fine-tuned on task-specific datasets to become more accurate and reliable for specialized applications.
Relationship to Generative AI and GPAI
Foundational models are closely linked to generative intelligence, which refers to systems capable of producing new content like text, images or code. Many foundation models, large language models fall within the generative artificial intelligence family. However not all foundation models are purely generative; some are used for prediction, summarization or other forms of latent representation learning. The concept of purpose artificial intelligence is sometimes used to describe foundation models that can be adapted to many different tasks and sectors. Regulators and governance frameworks treat GPAI as a category because a single model can influence a broad ecosystem of products and services. If such a model is deployed in sector or state-funded contexts any systemic flaws can propagate across multiple government applications creating systemic risk.