Generative AI

1. Meaning and essential distinctions

Generative artificial intelligence is a class of AI systems capable of producing new content resembling patterns learned from training data. Outputs include essays, computer code, images, speech, music, videos and candidate molecular structures. Unlike traditional rule-based software, these systems learn statistical relationships through training. The output is generated in response to an input, commonly called a prompt, which may contain instructions, examples, documents or images.

Artificial intelligence is the broad field of building systems that perform tasks associated with human intelligence. Machine learning is a subset in which systems learn from data, while deep learning uses multilayer neural networks. Much contemporary generative AI is based on deep learning. However, AI also includes non-generative applications such as credit-risk prediction, object detection and route optimisation.

A large language model, or LLM, is trained to process and generate language. A multimodal model handles more than one modality, such as text and images. A foundation model is trained on broad data and can be adapted to multiple tasks. Artificial general intelligence refers to a proposed broadly capable form of machine intelligence; generative ability alone does not establish general intelligence, consciousness or human-like understanding.

• Generative task: draft a description of a satellite image.
• Discriminative task: classify the image as forest or non-forest.
• Predictive task: estimate the probability of forest loss.

2. How generative models work

Training adjusts a model’s numerical parameters to reduce errors against a learning objective. Language-model training commonly uses self-supervision: the data itself supplies targets, such as the next token in a passage. Tokens may be words, word fragments or characters. During inference, the trained model processes a prompt and generates output. Training and inference both require computing resources, but only training necessarily changes the model’s parameters.

Transformers, introduced in 2017, use attention mechanisms to capture relationships among tokens. Many LLMs generate text autoregressively, predicting one token after another. Their context window limits the material they can process in a single interaction. A temperature setting influences sampling randomness: lower values generally make output more predictable, but do not guarantee correctness.

Other important architectures include generative adversarial networks, or GANs, in which a generator and discriminator compete during training. Diffusion models learn to reverse a progressive noising process and are widely used for image generation. Variational autoencoders learn probabilistic latent representations from which new samples can be generated. These architectures are not interchangeable, and modern systems may combine techniques.

• Fine-tuning further trains a model on selected data for a task or domain.
• Instruction tuning and reinforcement learning from human feedback can improve instruction-following and alignment with preferred behaviour.
• Retrieval-augmented generation, or RAG, retrieves external documents and supplies them as context. It does not inherently update model weights.
• Prompt engineering changes instructions or context without necessarily retraining the model.

3. Applications and enabling infrastructure

In education, generative AI can create practice questions, translate lessons and provide explanations adapted to different learning levels. In public administration, it can assist with document summarisation, multilingual citizen communication and searching departmental records. Such applications need authoritative source material, accessibility testing and escalation to accountable officials, particularly where welfare entitlements or legal rights are involved.

Healthcare applications include drafting clinical notes, assisting medical-image interpretation and proposing molecules for drug discovery. Generated molecules remain candidates requiring laboratory validation, toxicity assessment and clinical trials. In agriculture, locally adapted assistants can communicate crop advisories, but recommendations must reflect location, season and verified agronomic knowledge. Software development, design and scientific research are other major application areas.

The enabling infrastructure includes training data, GPUs and other AI accelerators, high-bandwidth networking, reliable electricity and data centres. Cloud computing improves access to powerful systems, while smaller models can run closer to users on edge devices. Quantisation reduces numerical precision to lower memory and computational requirements, sometimes at a performance cost. Environmental impacts depend on model size, usage, hardware efficiency, electricity sources and cooling arrangements.

• Indian priorities include regional-language coverage, affordable access and representation of low-resource languages.
• Synthetic data can supplement datasets, but is not automatically unbiased, accurate or privacy-preserving.

4. Limitations and societal risks

A hallucination is an incorrect or unsupported output presented as plausible information. It arises because generating statistically likely content is not equivalent to verifying truth. A model may fabricate references, misstate laws or make arithmetic errors. RAG, external tools and expert review can reduce such failures, but retrieved material may itself be unreliable and the model may misinterpret it.

Bias in training data can reproduce stereotypes or create unequal performance across languages and social groups. Privacy risks arise from collection of personal information, memorisation of training examples and disclosure through prompts. Copyright questions concern both training material and generated outputs; legal treatment varies across jurisdictions. Open-weight availability does not automatically confer unrestricted rights over training data or satisfy every definition of open-source software.

Deepfakes can impersonate people through synthetic audio or video, facilitating fraud and misinformation. Models can also be manipulated through prompt injection, including instructions hidden in retrieved documents. AI agents that invoke tools or execute workflows introduce additional risks if granted excessive permissions. Employment effects may involve task transformation and displacement, making reskilling important. High-stakes decisions require human accountability rather than uncritical acceptance of automated recommendations.

• Watermarking and provenance records can support authenticity checks, but are not foolproof.
• Fluency, model size and benchmark scores are not substitutes for domain-specific testing.

5. Indian policy and international governance

The Union Cabinet approved the IndiaAI Mission on 7 March 2024. Its seven pillars cover compute capacity, an innovation centre, datasets, application development, future skills, startup financing, and safe and trusted AI. The mission is implemented through the IndiaAI Independent Business Division under Digital India Corporation, within the Ministry of Electronics and Information Technology. Its objectives include broadening access to AI infrastructure and strengthening indigenous capabilities.

NITI Aayog’s National Strategy for Artificial Intelligence, published in 2018, advanced an AI for All approach. Its responsible-AI work emphasises safety, equality, inclusion, privacy, transparency and accountability. Relevant legal frameworks include the Information Technology Act, 2000, the Copyright Act, 1957, and the Digital Personal Data Protection Act, 2023. Applicability depends on the activity, statutory scope and commencement of relevant provisions; IndiaAI Mission is not itself an AI regulation.

Internationally, UNESCO adopted its Recommendation on the Ethics of Artificial Intelligence in 2021. The European Union’s AI Act entered into force in August 2024 with phased application, combining risk-based obligations with rules for general-purpose AI models. For India, an appropriate approach balances innovation with safeguards: representative datasets, security testing, independent evaluation, grievance redress and meaningful human oversight. Regulation should distinguish a low-risk creative assistant from systems affecting health, employment or access to essential services.

Generative AI within the broader AI landscape

The supplied metadata describes a Venn diagram relating artificial intelligence to generative models, making it directly relevant to this microtopic. AI is the broad field of building systems that perform tasks associated with intelligence; machine learning develops capabilities through learning from data. Generative models learn statistical patterns that enable them to produce new samples, such as text or images. Most prominent contemporary generative AI systems use deep learning, but generative modelling is not confined to deep neural networks.

A subset diagram is a conceptual aid, not proof of a rigid technological hierarchy. Generative and predictive functions can coexist in one system: a language model commonly generates text by repeatedly predicting the next token. This does not establish factual accuracy, human-like understanding or artificial general intelligence. Since only metadata is available, the diagram’s precise labels and nesting cannot be verified here.

For India, applications include multilingual public-service assistance, educational resources and administrative drafting. Deployment requires attention to uneven language representation, personal-data protection, copyright and accountability. Retrieval from trusted sources can improve grounding, but neither retrieval nor fluent output guarantees correctness. Consequential uses need proportionate safeguards, human review and clear responsibility for errors.

• Prelims distinction: generative AI creates content; a search engine primarily retrieves information, though products may combine both functions.
• Evidence limit: plausible generated statements require independent verification, especially in legal, scientific and policy contexts.
• Mains approach: assess productivity and inclusion alongside misinformation, bias, labour impacts and institutional oversight.

Previous year questions

No source-linked PYQ is available here yet. Use the practice questions below.

Practice questions