Artificial Intelligence is the broad field of building computer systems that do things which normally require human intelligence: recognising a face, understanding a sentence, planning a route, spotting a fraudulent transaction, writing a paragraph.
The term covers a huge range of techniques with very different capabilities, which is why "AI" in a news headline and "AI" in a product brochure often mean entirely different things.
The layers, from widest to narrowest
- Artificial Intelligence is the umbrella: any technique that makes a machine behave in a way we would call intelligent.
- [Machine learning](/glossary/machine-learning) is the subset where the system learns patterns from examples instead of being given explicit rules.
- Deep learning is machine learning built on neural networks with many layers, which is what made modern image and speech recognition work.
- [Generative AI](/glossary/generative-ai) is deep learning applied to producing new content rather than only classifying existing content.
- [Large language models](/glossary/large-language-model) are the specific kind of generative AI behind ChatGPT, Claude and Gemini.
When someone says "we use AI", the useful follow-up question is which of these layers they actually mean.
Rule-based AI and learned AI
Early AI was written by hand. An expert wrote out thousands of if-then rules, and the system followed them. This worked for narrow problems like chess openings and tax calculations, and failed completely at anything fuzzy, because nobody can write down the rules that distinguish a cat from a dog.
Modern AI mostly learns. You show it a million labelled photos and it works out the distinguishing patterns itself. Nobody programmed the definition of a cat. The system inferred it from examples, which is why it can also be confidently wrong in ways nobody anticipated.
Narrow AI is all that currently exists
Every AI system in use today is narrow: excellent at the specific task it was trained for, and useless outside it. A model that beats every human at Go cannot order a taxi. A model that writes fluent essays cannot reliably add up a column of numbers.
Artificial general intelligence, a system with broad human-level ability across arbitrary tasks, does not exist. It is a research goal and a subject of serious disagreement about timelines. Marketing that implies otherwise is marketing.
This matters practically. The right question about an AI tool is never "is it intelligent" but "what exactly was it trained to do, and is my task inside that boundary".
Where you already meet it in India
Most people use AI dozens of times a day without noticing:
- UPI and card fraud detection flagging an unusual transaction in milliseconds
- Loan underwriting at banks and NBFCs scoring applications
- Voice assistants and speech recognition in Hindi, Tamil, Bengali and other Indian languages
- Recommendations on YouTube, Instagram, Amazon and streaming apps
- Maps predicting traffic and arrival times
- Spam and phishing filters in Gmail
- Chatbots on banking, telecom and government service portals
- Document processing for KYC, insurance claims and cheque reading
Practical implications if you are using AI tools
It does not understand, it predicts. A model producing a confident paragraph about Indian tax law is producing statistically likely text, not consulting the Income Tax Act. See hallucination.
It can be wrong in an authoritative voice. This is the single most dangerous property for anyone using AI for finance, health, legal or tax questions. Verify anything that affects a real decision.
Training data has a cutoff. Unless a tool is explicitly searching the web, it does not know about rules, rates or events after its training data ended, which matters a great deal for Indian tax and regulatory questions that change every Budget.
It reflects the biases in its data. Models trained mainly on Western English text often handle Indian names, addresses, contexts and languages less reliably.
Your inputs may be retained. Free consumer AI tools frequently use conversations to improve their models. Do not paste Aadhaar numbers, PAN details, bank statements, salary slips, client contracts or anything covered by confidentiality obligations. India's Digital Personal Data Protection Act places real obligations on organisations handling personal data, and pasting customer information into a public chatbot can breach them.
It is a drafting tool, not an authority. The reliable pattern is to use AI to produce a first version quickly and then apply your own judgement, rather than to outsource the judgement.
Explore practical tools in our AI tools directory and see prompt engineering for how to get better output.