Artificial intelligence

The first principle

A conventional program follows rules a person writes, while a machine learning program finds its own rules from examples. Given enough examples and enough computing power, a system that adjusts millions or billions of internal numbers to reduce its errors can learn to recognise faces, translate speech or write text without anyone spelling out how. That is the source of both its power and its limits: it learns whatever patterns are in its data, including errors and prejudices, and it produces what is likely given that data, not what is checked to be true.

The core

Learning from data

• Artificial intelligence (AI) is the broad aim of making machines perform tasks that need intelligence in people, such as perceiving, reasoning, using language and deciding. Early AI wrote rules by hand, which worked in narrow settings such as chess but broke down in the messy real world.
• Machine learning is the approach that now dominates: the system is shown data and improves its performance on a task by adjusting itself. In supervised learning it learns from labelled examples, such as X ray images marked healthy or diseased; in unsupervised learning it finds groupings in unlabelled data; and in reinforcement learning it learns by trial, guided by a reward, which is how game playing systems are trained.
• A neural network is a machine learning model built from layers of simple units loosely inspired by neurons. Each unit multiplies its inputs by weights, adds them and passes the result through a simple function; training compares the output with the right answer and nudges every weight slightly in the direction that reduces the error, a method called backpropagation. Deep learning means networks with many layers, which became practical around 2012 when graphics processors and large labelled datasets became available.

How a large language model works

• Tokens and next token prediction. A large language model (LLM) cuts text into tokens, which are words or parts of words, turns each into a list of numbers, and is trained on a very large body of text with one task repeated billions of times: predict the next token, compare with the real one, and adjust its parameters. When it answers a prompt it does the same thing, adding one likely token at a time.
• The transformer. LLMs use the transformer design, introduced by Google researchers in 2017, whose attention mechanism lets every token weigh its relation to every other token in the passage. This captures context, so that "bank" is read differently near "river" and near "loan".
• Parameters and the context window. Parameters are the learnt weights, now running to hundreds of billions in the largest models, and the context window is how much text a model can consider at once. Both drive the cost of computing, which is why AI depends on thousands of GPUs in large data centres.
• Alignment and grounding. After pretraining, models are further trained on human feedback so that they follow instructions and decline harmful requests. Retrieval augmented generation attaches a search engine or document store and puts what it finds into the prompt, which brings in current facts.
• Hallucination is built in. Because a model produces likely text rather than checked facts, it can state false things fluently, so its output must be grounded in sources or checked where accuracy matters.
• Open weight and closed models. An open weight model releases its trained parameters for anyone to download and adapt, while a closed model is used only through its maker's service; a released model cannot be withdrawn.

Dimensions

• Economic. AI raises productivity in software, customer service, design and research, but it can also replace routine office work, so the gains depend on reskilling; India's large services workforce is both exposed and well placed to use it.
• Scientific. AI now does science: AlphaFold predicts the shape of proteins from their sequence, work recognised by the Nobel Prize in Chemistry 2024, and the Nobel Prize in Physics 2024 went to John Hopfield and Geoffrey Hinton for foundations of neural networks.
• Social. A model trained on biased data repeats the bias, for example in hiring or lending, and AI that works mainly in English excludes most Indians, which is why models in Indian languages matter.
• Security. Deepfakes, synthetic audio and video that imitate real people, enable fraud and disinformation, and AI assisted cyber attacks are covered in the internal security note on cybersecurity; military uses are in the note on drones and emerging warfare.
• Environmental. Training and running large models consumes much electricity and water for cooling data centres, so the growth of AI adds to energy demand.
• Legal and ethical. Questions of who is liable for an AI decision, whether training on copyrighted text is lawful, and how personal data in training sets is protected are still being settled by courts and lawmakers.

In practice

• National strategy. NITI Aayog's National Strategy for Artificial Intelligence of 2018, under the theme AI for All, named health, agriculture, education, smart cities and mobility as priority sectors.
• IndiaAI Mission. Approved by the Union Cabinet in March 2024 with about Rs 10,372 crore over five years under the Ministry of Electronics and Information Technology (MeitY), it has seven pillars: compute, foundation models, AIKosh (a platform for datasets), application development, FutureSkills, start up financing, and safe and trusted AI. More than 38,000 GPUs have been made available through a common compute portal at subsidised rates for start ups and researchers, and firms and consortia such as Sarvam AI and BharatGen, led by the Indian Institute of Technology Bombay, have been selected to build foundation models trained for Indian languages.
• Language and public services. Bhashini, the national language translation platform, provides speech and text translation across Indian languages as a public good, so that AI can reach citizens through digital public infrastructure.
• Governance. MeitY released the India AI Governance Guidelines in November 2025, which rest on seven principles, including trust, people first, fairness and accountability, and prefer applying existing laws such as the Information Technology Act 2000 and the Digital Personal Data Protection Act 2023 over a separate AI law for now. They propose an AI Governance Group to coordinate policy, a Technology and Policy Expert Committee and an AI Safety Institute. In February 2026 amendments to the Information Technology Rules required synthetically generated content to be labelled and carry provenance markers, and cut the time to act on lawful takedown orders to three hours.
• Global. The European Union's Artificial Intelligence Act, in force from August 2024, bans some uses outright and puts strict duties on high risk uses, and India hosted the AI Impact Summit in New Delhi in February 2026, after earlier summits at Bletchley Park, Seoul and Paris.

Issues and way forward

• Compute and chips. Training frontier models needs advanced GPUs that are made abroad and subject to export controls, so India depends on imports; shared national compute, data centres powered by clean energy and the semiconductor mission are the long term answers.
• Data. Indian language text and good public datasets are scarce, and personal data must be protected; AIKosh, Bhashini's data collection and consent based sharing under the data protection law can supply data lawfully.
• Trust. Hallucination, bias and opaque decisions make AI risky in health, justice and welfare, so high stakes uses need human oversight, testing before deployment and a way for citizens to appeal.
• Jobs. Routine tasks in services are exposed, so skilling, including through FutureSkills, and social protection for displaced workers must keep pace.
• Regulation without stifling. Too little regulation invites harm, while too much drives work abroad; India's approach of guidelines, sector regulators and targeted rules such as deepfake labelling is meant to balance the two, and will need review as the technology changes.

Common confusions

• AI, machine learning and deep learning. AI is the goal, machine learning is the main method of reaching it, and deep learning is machine learning with many layered neural networks.
• Training and inference. Training builds the model by adjusting its parameters on data, which is costly and done once, while inference is using the trained model to answer, which is done millions of times.
• Generative and predictive AI. Predictive AI classifies or forecasts, such as flagging a fraudulent payment, while generative AI produces new text, images or sound.
• Open source and open weight. Open weight means the trained parameters are released, while open source in the full sense also releases the training code and data, which few large models do.
• Narrow and general AI. Every system in use today is narrow, good at the tasks it was trained for; artificial general intelligence, a system as capable as a person across tasks, does not yet exist.

For the exam

• Keyword bank: machine learning, neural network, deep learning, transformer, large language model, parameters, hallucination, retrieval augmented generation, algorithmic bias, deepfake, compute, sovereign AI, IndiaAI Mission, AIKosh, risk based regulation.
• Previous year questions: Practise the previous year questions on this topic →
• Use in an answer:
  • Introduction with data: The IndiaAI Mission of March 2024, with about Rs 10,372 crore, has made more than 38,000 GPUs available to Indian start ups and researchers at subsidised rates.
  • Introduction with a source: NITI Aayog's National Strategy for Artificial Intelligence of 2018 set India's aim as AI for All, using AI for inclusion in health, agriculture and education.
  • Conclusion: AI turns data and computing power into decisions, so India's advantage lies in its languages, its digital public infrastructure and its talent, provided it builds compute and trust alongside.
  • Conclusion: Because a model produces what is likely rather than what is true, AI should assist human judgement in high stakes decisions, with accountability resting with people and institutions.