Artificial Intelligence (AI): Introduction
1. Definition
Artificial Intelligence (AI) is the branch of computer science that builds machines and software capable of performing tasks that normally require human intelligence — learning, reasoning, problem-solving, perception, understanding language and decision-making. John McCarthy, who coined the term, defined it as "the science and engineering of making intelligent machines". AI does not have to imitate a human brain; it only has to achieve a comparable outcome (acting rationally and effectively). Alan Turing's 1950 test — can a machine's conversation be indistinguishable from a human's? — gave the field its first operational benchmark.
2. A Brief History
1950 Turing test → 1956 Dartmouth Conference (the term "Artificial Intelligence" is born) → 1960s-70s symbolic AI, search, game playing → 1970s-80s expert systems (MYCIN, DENDRAL) and the first AI winter → 1997 Deep Blue beats Kasparov → 2011 IBM Watson wins Jeopardy, Siri launches → 2012 AlexNet ignites deep learning (thanks to GPUs and big data) → 2016 AlphaGo beats Lee Sedol → 2017 the Transformer architecture → 2022+ generative AI (ChatGPT, diffusion image models) → today, multimodal and agentic AI.
3. Levels and Types of AI
| By capability | Meaning | Status |
|---|---|---|
| Narrow / Weak AI | Performs one specific task as well as or better than a human | Exists today (search, translation, recommendations, chess, face unlock) |
| General AI (AGI) | Human-level ability across all cognitive tasks, transferring learning between domains | Hypothetical — not achieved |
| Super AI | Far exceeds the best human minds in every field | Speculative |
| By functionality | Meaning |
|---|---|
| Reactive machines | Respond to the current input only, no memory (Deep Blue, simple game AI) |
| Limited memory | Use recent past data (self-driving cars, chatbots) |
| Theory of mind | Model the beliefs and emotions of others — research stage |
| Self-aware | Conscious of itself — conceptual only |
4. Subfields and Techniques
| Subfield | What it does | Typical use |
|---|---|---|
| Machine Learning (ML) | Learns patterns from data instead of being explicitly programmed | Spam filters, credit scoring, recommendations |
| Deep Learning (neural networks) | Multi-layer networks learn features automatically | Image/voice recognition, LLMs |
| Natural Language Processing (NLP) | Understand and generate human language | Translation, chatbots, summarisation, sentiment |
| Computer Vision | Interpret images and video | Face recognition, medical imaging, ANPR, defect detection |
| Speech processing | Convert speech ↔ text, identify speakers | Voice assistants, dictation |
| Robotics | Combine perception, planning and control to act physically | Industrial arms, drones, surgical robots |
| Expert systems | Rule-based reasoning with a knowledge base and inference engine | Medical/fault diagnosis (legacy, still used) |
| Search, planning and optimisation | Find solutions in large spaces | Route planning, scheduling, game AI |
| Knowledge representation | Encode facts and relationships for reasoning | Knowledge graphs, semantic web |
| Generative AI | Produce new text, image, audio, video or code | Chatbots, image generators, code assistants |
| Multi-agent systems | Many autonomous agents cooperating or competing | Traffic simulation, market modelling |
5. How AI Roughly Works
Data is collected and cleaned → features are engineered or learned → a model (statistical, neural or rule-based) is trained by minimising error on examples → the trained model performs inference on new inputs → feedback and retraining improve it. Traditional "symbolic" AI instead writes explicit rules and uses logic/search; modern AI is overwhelmingly data-driven and statistical, with neural networks whose millions of parameters are adjusted during training.
AI ⊃ ML ⊃ Deep Learning ⊃ Generative AI — each is a subset of the one before.
6. AI vs Human Intelligence
| Aspect | AI | Human |
|---|---|---|
| Speed of computation | Very high, parallel on GPUs | Slower, but efficient |
| Learning | Needs large labelled data; no common sense transfer | Learns from few examples, transfers knowledge |
| Memory | Huge but literal | Associative, reconstructive |
| Emotion, ethics, creativity | Absent; can imitate style | Central to judgement and meaning |
| Energy | Very high for training | ~20 W for a brain |
| Consistency | Tireless, repeatable | Varies with fatigue and mood |
| Explainability | Often a "black box" | Usually can explain reasoning |
7. Benefits, Applications and Risks
Benefits/applications: automation of repetitive and high-volume work, faster and more accurate diagnosis, fraud detection, personalisation, language access, scientific discovery, accessibility (speech/vision for the disabled), and new products (assistants, generative tools).
Limitations and risks: bias inherited from data, hallucination (confidently wrong output), lack of explainability, privacy and data-consent issues, high compute cost and energy use, job displacement, dependency and over-trust, misuse (deepfakes, automated disinformation, cyberattacks), security of models and prompts (prompt injection), and legal/regulatory questions (EU AI Act, NITI Aayog's responsible-AI principles, data-protection law). AI ethics therefore emphasises fairness, transparency, accountability, safety, privacy and human oversight.
8. Exam Points
- Define AI and state its goals — 5 marks, almost certain.
- Differentiate narrow AI, general AI and super AI with examples — 5 marks, frequent.
- List any six subfields of AI with one application each — 5 marks, very frequent.
- Explain the relationship AI ⊃ ML ⊃ DL with a diagram — 3-5 marks.
- What was the Turing test? — 3 marks.
- Compare AI and human intelligence — 5 marks.
- Explain any four risks/ethical concerns of AI (bias, hallucination, privacy, job loss) — 5 marks.
- Write short notes on generative AI — 3-5 marks.