How should we define AI?

A working definition

AI is not a single technology. It is a loose family of techniques for getting computers to do things that, until recently, required a human mind. Recognising a face in a photograph. Translating Nepali into English. Deciding whether a Khalti transaction looks like fraud. Drafting an email. Playing chess.

What unites these techniques is not magic. It is pattern — vast amounts of it — and the mathematical machinery to find that pattern in data.

Why “intelligence” is a misleading word

The word intelligence is doing a lot of work in “artificial intelligence” — and most of it is misleading. The systems we are about to study do not think. They do not understand. They do not want anything. They are mathematical functions: you give them input, they produce output, and that is all.

What is genuinely new is how good these functions have become at certain narrow tasks. A modern translation model trained on millions of Nepali-English sentence pairs can produce translations that, ten years ago, would have required a human. That is remarkable. It is also not intelligence in any sense your aama would recognise.

Three examples close to home

To make this concrete, consider three systems you may already use every day in Nepal:

1. Khalti fraud detection. When you send Rs. 5,000 to a vendor, a model decides — in milliseconds — whether the transaction “looks like” you, based on patterns from millions of previous transactions. If it doesn’t, you get an OTP challenge. The model has never met you. It has only ever seen patterns.

2. Pathao route suggestions. When a rider opens the app in Naxal at 6pm and needs to reach Patan, the system suggests a route. It is balancing traffic patterns it has learned from past trips against the real-time positions of other vehicles. No human is choosing your route. A function is.

3. Devanagari OCR on your phone. Point your phone camera at a sign in Asan and your phone can read the Nepali text. Twenty years ago this required a research lab. Today it is a free feature, running locally on your handset, because the underlying pattern — what does the letter क look like across millions of handwriting samples? — has been learned.

None of these systems is intelligent. Each is, in its own narrow way, very useful.

Why this matters for Nepal

There is a temptation, when reading a course like this, to assume AI is something that happens to Nepal — built in California, deployed in Kathmandu, with us as passive users. This is wrong on two counts.

First, the techniques themselves are not secret. The mathematics of a neural network is in textbooks. The code is on GitHub. The constraint is not knowledge; it is data, computation, and the imagination to apply them to local problems. A Nepali developer who understands these techniques and also understands, intimately, the patterns of Nepali life — paddy yields, monsoon timing, migration flows, the particular ways our language bends — has an advantage no foreign team can match.

Second, AI systems trained elsewhere often fail here in ways their builders never see. A facial recognition system trained mostly on European faces will misidentify Nepali ones. A language model trained on English will hallucinate when asked about Nepali history. Building AI literacy in Nepal is not just about using what others have made. It is about knowing where it breaks, and being able to fix it — or replace it.

Check your understanding

Quick check

—

Which of the following is the most accurate description of what modern AI systems do?

1. AThey think and reason in ways similar to humans, just faster.
2. BThey find statistical patterns in large amounts of data and use those patterns to produce outputs.
3. CThey follow a long list of explicit rules written by programmers.
4. DThey are conscious but their consciousness is artificial rather than biological.

Check answer→

Quick check

—

A facial recognition system trained mostly on European faces is deployed in Kathmandu and frequently misidentifies people. What is the most likely cause?

1. AThe system has a bug in its code.
2. BThe system is not running on fast enough hardware.
3. CThe training data did not represent the population the system is being used on.
4. DNepali faces are intrinsically harder to distinguish.

Check answer→

What comes next

The next section looks at the neighbouring fields — machine learning, data science, robotics — and where the boundaries between them actually fall. The vocabulary matters: muddling these terms is how budgets get spent on the wrong things.

For now, sit with one question: where in your own life have you already used AI today without noticing?