AI — UPSC Science & Technology Notes

1. Meaning, capabilities and technological foundations

Artificial intelligence involves machines performing tasks normally associated with human cognitive abilities. Earlier systems relied heavily on explicitly programmed rules and expert knowledge. Contemporary machine learning identifies statistical patterns from data and uses them to classify inputs, predict outcomes or support decisions. Deep learning employs multilayer neural networks and has significantly improved image recognition, speech processing and natural language processing.

In supervised learning, models learn from labelled examples, such as medical images linked to diagnoses. Unsupervised learning identifies patterns in unlabelled data, including customer clusters or unusual transactions. Reinforcement learning uses rewards and penalties to learn action strategies. Training adjusts model parameters; inference applies the trained model to new inputs. Their computing requirements, costs and risks can differ substantially.

Generative AI models learn patterns that enable them to produce new content. Large language models process text through tokens and generate responses using learned relationships. Foundation models are trained on broad datasets and can be adapted to multiple tasks. Multimodal systems process combinations of text, images, audio or video. AI agents combine models with tools to execute multistep tasks, introducing additional risks when allowed to act autonomously.

Present-day systems remain fallible: they may generate fabricated information, fail on unfamiliar inputs or reproduce bias. Artificial general intelligence denotes a proposed broadly capable system; it should not be confused with demonstrated, dependable human-level competence across all domains. AI capability depends on data quality, algorithms, computing infrastructure, electricity, connectivity and domain expertise.

Automation follows a predefined workflow; AI can infer patterns, but not every automated system is AI.
Retrieval-augmented generation supplies external reference material to a model; it can improve grounding but does not eliminate errors.

2. Societal applications and development opportunities

In healthcare, AI can assist radiologists, screen retinal images, predict patient deterioration and support drug discovery. Its strongest public-health contribution may be extending scarce specialist capacity. However, performance on a laboratory dataset cannot substitute for prospective clinical validation. Differences in disease prevalence, equipment and patient populations can reduce reliability. Clinical responsibility, informed consent and referral pathways must remain clear.

In agriculture, AI combines satellite imagery, weather records, soil information and farm observations to support pest detection, irrigation scheduling and yield estimation. Benefits depend on local-language interfaces and advice suited to small and marginal farmers. In education, adaptive learning, translation and assistive technologies can support differentiated instruction, but teacher judgement and safeguards for children remain essential.

For governance, AI can help classify grievances, translate documents, detect anomalies in expenditure and prioritise inspections. Judicial applications include transcription, translation and legal-information retrieval, rather than replacing judicial reasoning. In financial services, fraud detection and credit-risk assessment can improve efficiency, although opaque models may exclude applicants with limited formal financial histories.

AI can also improve flood forecasting, electricity-demand prediction, wildlife monitoring and industrial maintenance. Accessibility tools such as speech-to-text and image descriptions can expand participation for persons with disabilities. Yet digital exclusion may worsen if services require smartphones, reliable connectivity or dominant-language proficiency. Public agencies should therefore retain assisted and non-digital channels and measure outcomes such as improved access, reduced error and citizen satisfaction—not merely the number of models deployed.

Responsible adoption requires domain-specific validation, user training and mechanisms to contest consequential decisions.
Public-interest datasets should be representative, lawfully collected and protected against unauthorised disclosure.

Responsible AI deployment life cycle

1. Define the public problem and assess whether AI is necessary
2. Establish lawful data use and assess likely social impacts
3. Train or procure a model with documented limitations
4. Test accuracy, subgroup performance, privacy and security
5. Pilot with meaningful human oversight and grievance redress
6. Monitor outcomes and update, suspend or withdraw when required

3. Strategic significance, economic transformation and security

AI is a general-purpose technology with applications across manufacturing, services, research and defence. It can accelerate materials discovery, optimise supply chains and improve software development. However, productivity gains require complementary investment in skills, organisational redesign and infrastructure. Control over advanced semiconductors, cloud services, foundation models and critical datasets creates geopolitical leverage and technological dependencies.

For India, strategic autonomy means developing domestic capabilities while retaining international research and commercial partnerships. Priorities include affordable computing access, semiconductor design, Indian-language resources, secure cloud infrastructure and research talent. Complete technological self-sufficiency is neither easy nor necessarily efficient; diversified suppliers, interoperability and resilience are more practical objectives.

Defence applications include intelligence analysis, reconnaissance, logistics, predictive maintenance and decision support. Autonomous weapons raise difficult questions about distinction, proportionality, accountability and meaningful human control. Faster machine-assisted decisions can also create escalation risks, particularly where uncertain information is treated as authoritative. AI does not displace obligations under international humanitarian law.

Cybersecurity illustrates AI’s dual-use character: it can detect suspicious activity but also assist phishing, malware development and social engineering. Deepfakes can facilitate impersonation, financial fraud and electoral manipulation. Watermarks and provenance credentials help establish origin, but neither offers a complete solution. On employment, AI is likely to reorganise tasks as well as displace some roles. Policy should support reskilling, worker transitions and broader access to productivity gains rather than assume either universal unemployment or automatic prosperity.

Strategic bottlenecks include advanced chips, high-bandwidth memory, energy supply and concentration of cloud infrastructure.
AI-related national security measures should remain proportionate and preserve legitimate research and civil liberties.

Major AI approaches and their examination relevance
Approach	Core feature	Example	Principal caution
Rule-based AI	Uses explicitly defined rules	Expert decision-support system	Brittle outside encoded situations
Supervised learning	Learns from labelled examples	Disease classification	Biased labels and poor generalisation
Unsupervised learning	Finds patterns in unlabelled data	Transaction clustering	Patterns may lack meaningful interpretation
Reinforcement learning	Learns through reward signals	Robotic control	Poorly specified rewards can induce unsafe behaviour
Generative AI	Produces new content from learned patterns	Text and image generation	Fabrication, privacy and intellectual-property risks

4. India’s AI policy and institutional framework

NITI Aayog’s National Strategy for Artificial Intelligence, released in 2018, identified healthcare, agriculture, education, smart cities and infrastructure, and smart mobility as priority sectors. Its AI for All approach connected technological development with inclusion. Subsequent Responsible AI documents emphasised safety, equality, privacy, transparency and accountability.

The IndiaAI Mission, approved in March 2024, seeks to strengthen the domestic AI ecosystem. Its seven pillars address computing infrastructure, indigenous AI innovation, datasets, application development, future skills, startup financing, and safe and trusted AI. The mission’s approved design envisaged access to an ecosystem of 10,000 or more graphics processing units through public-private collaboration; this is a programme-design figure, not a fixed statement of current installed capacity.

BHASHINI, under the National Language Translation Mission, supports language technology and multilingual digital access. Initiatives such as these are important because commercial training data often underrepresent Indian languages, dialects and cultural contexts. Government procurement can stimulate demand, but contracts should specify performance standards, data rights, security requirements, audit access and exit arrangements to prevent vendor lock-in.

AI governance in India involves existing constitutional, statutory and sectoral obligations. The Digital Personal Data Protection Act, 2023 establishes a framework for digital personal data, subject to its commencement and implementation arrangements. The Information Technology Act, 2000, applicable rules, consumer-protection law and sectoral regulation may also be relevant. Following the Supreme Court’s 2017 Puttaswamy judgement, state uses affecting privacy must satisfy constitutional safeguards; efficiency alone cannot justify unrestricted surveillance.

Distinguish an announced policy, a funded mission, a statutory obligation and an operational achievement.
Institutional capacity in ministries, states and local bodies is as important as model-building capacity.

5. Risks, ethical principles and a responsible pathway

AI bias may arise from unrepresentative datasets, historical discrimination, unsuitable labels or deployment conditions. A model can be statistically accurate overall yet perform poorly for a vulnerable group. Explainability concerns understanding how an output was produced; accountability concerns who must justify it and provide a remedy. Neither can be achieved merely by publishing a general ethical statement.

Privacy risks include excessive data collection, re-identification and leakage of sensitive training information. Security threats include data poisoning, adversarial inputs and prompt injection. In generative AI, confidential information can also leak through careless user prompts. Copyright disputes concern training material and generated outputs, while misinformation undermines trust in authentic evidence.

A risk-based approach applies stronger requirements where failure threatens life, liberty or access to essential services. Appropriate safeguards include impact assessments, subgroup testing, independent audits, documented limitations, incident reporting and accessible appeals. Human oversight must be meaningful: an official needs the time, competence and authority to challenge the system rather than mechanically approve its recommendation.

Internationally, UNESCO’s 2021 Recommendation on the Ethics of Artificial Intelligence provides a normative framework. The European Union’s AI Act entered into force in August 2024 with phased application and a risk-based structure. India can learn from these approaches without copying them mechanically. Its governance should combine innovation-friendly testing with enforceable safeguards, public participation and competition. Environmental assessment should also consider electricity demand, water used in cooling and hardware-related emissions alongside any efficiency gains.

Evaluate necessity first: a simpler statistical method or administrative reform may solve the problem more reliably.
Monitor model drift after deployment and establish clear thresholds for correction, suspension or withdrawal.

Real-world case studies

AI-supported tuberculosis screening in India

Mumbai-based Qure.ai’s qXR has been used in tuberculosis-screening programmes to flag suspicious chest X-rays for further investigation. WHO’s 2021 guidance allowed computer-aided detection to substitute for human interpretation of digital chest radiographs for TB screening and triage among people aged 15 years or older. Such screening does not itself confirm TB: bacteriological testing, locally appropriate thresholds and referral arrangements remain essential.

AlphaFold and scientific discovery

DeepMind’s AlphaFold demonstrated major advances in predicting protein structures from amino-acid sequences. In 2022, the AlphaFold Protein Structure Database, developed with EMBL-EBI, expanded to more than 200 million predictions. It illustrates AI’s potential to accelerate basic science. Nevertheless, predicted structures have varying confidence and do not automatically reveal protein function, molecular interactions or the clinical effectiveness of a drug.

Previous year questions

UPSC Mains 2023 · GS-III

Introduce artificial intelligence. How does it help clinical diagnosis? Do you perceive any threat to individual privacy when AI is used in healthcare?

Define AI and briefly distinguish machine learning from rule-based systems.
Explain imaging support, risk prediction, clinical decision support and early detection.
Discuss sensitive health data, re-identification, unauthorised sharing and security breaches.
Recommend clinical validation, data minimisation, access controls, consent safeguards and professional accountability.

Practice questions

Practice MCQ 1

Consider the following statements: 1. Deep learning is a subset of machine learning. 2. Generative AI systems necessarily retrieve an existing document before answering. 3. Reinforcement learning can use rewards to shape an agent’s behaviour. Which statements are correct?

A. 1 and 2 only
B. 1 and 3 only
C. 2 and 3 only
D. 1, 2 and 3

Practice MCQ 2

Which of the following best describes model drift?

A. Physical movement of a computing device
B. Mandatory replacement of all training data after inference
C. A change in real-world data or relationships that can reduce a deployed model’s reliability
D. Conversion of a neural network into an explicit legal rule

Practice MCQ 3

A public authority uses AI to identify welfare applications for additional scrutiny. Which arrangement best ensures meaningful human oversight?

A. An officer approves every flagged decision without accessing supporting information
B. The vendor’s accuracy claim replaces independent testing
C. Applicants are denied any explanation to protect the algorithm
D. Trained officers can examine evidence, override recommendations and provide an appeal mechanism

Mains practice · Artificial intelligence is simultaneously a development opportunity and a strategic vulnerability for India. Discuss and suggest a framework for its responsible adoption. Answer in 250 words.

Introduce AI as a general-purpose, dual-use technology.
Illustrate opportunities in healthcare, agriculture, public administration and scientific research.
Examine dependence on chips, cloud infrastructure, foreign models and concentrated datasets.
Address privacy, bias, employment transitions, deepfakes and autonomous decision-making.
Link the IndiaAI Mission with domestic capabilities, multilingual access and skills.
Recommend risk-based regulation, lifecycle audits, human oversight, grievance redress and international cooperation.