Why here is different · ailiteracy.nepal

Most writing about AI is, quietly, writing about America. The companies are American. The benchmarks are American. The journalists who interview the engineers are American, and the engineers, when they describe what they are trying to do, picture an American user. If you read this writing long enough, you start to think the questions it is asking are the only questions — that the world has one AI story, and you are catching up to it.

This course begins by disagreeing.

Two different conversations

When a venture capitalist in San Francisco worries about AI, they worry about super-intelligence, copyright, and whether the next model will eat the last one’s margins. When a teacher in Pokhara worries about AI, she worries about whether the new app her students are using has any idea who Pṛthvī Nārāyaṇ Shāh was — and whether the textbook she has to teach from will be obsolete before the term ends.

Both of these are real worries. They are not the same worry. And the second one is barely visible inside the first.

Three frictions you cannot import past

Three forces shape almost every AI question in this country. They are unglamorous and they are decisive.

1. The data is thin, and what exists is in the wrong places. A model is only as smart as the data it has seen. Nepal has comparatively little public, high-quality, machine-readable data about itself. Land records exist but are scanned. Health records exist but live in paper registers in hundreds of municipalities. Court judgments are PDFs of photocopies. Even a basic question — “how many small businesses are there in Province 2?” — does not have a clean answer you can hand a model. A foreign model, trained on the United States Census and JSTOR, knows the world better than it knows us.

2. The language is under-represented. Roughly 30 million people speak Nepali. On the internet, however, Nepali is a small language. Most large language models have read perhaps a few hundred megabytes of Nepali text against several terabytes of English. The model can write a serviceable Nepali email, but it will guess wildly about Madhesh politics, mis-decline honorifics, and quietly anglicise idioms. The further you get from नमस्ते — कस्तो छ? and into anything technical, legal, or culturally specific, the thinner the ice.

3. The deployment context is rural, mobile, intermittent. A surprising amount of AI tooling assumes a fast laptop on stable wifi. In Nepal that is a niche case. The median user is on a mid-range Android phone, on 4G that drops to 2G when the bus crosses a ridge, with a power supply that is back but not yet trusted. Anything you build for general use has to survive that environment — or it serves only the people who least need help.

Why borrowing the agenda wholesale fails

Borrowing the global AI agenda without translation does two specific damages.

First, it spends scarce attention on the wrong problems. The country has perhaps a few thousand people who could plausibly work in AI. If they spend their careers replicating ChatGPT for a domestic market that already uses ChatGPT, they leave the actual Nepali problems — Nepali OCR for the courts, Nepali ASR for the health posts, fraud detection for cooperatives — to nobody.

Second, it imports assumptions that quietly harm us. A copyright regime designed for Hollywood does very little for a country whose folk songs are the heritage. A facial-recognition framework designed around US privacy law does not anticipate a country with eight ethnic majorities and a recent civil conflict. If we adopt these frames whole, we will have written the wrong rules before we noticed we were writing rules at all.

What this course is, and is not

This course is not a Nepali translation of the global AI debate. It is a tour of where AI actually meets this country today — the systems already running in Khalti, the satellite images already segmenting Rasuwa, the OCR already failing on Nepali court records — and the systems that could run, and probably should, if we choose to build them.

It is opinionated. It will make claims about what Nepal should and should not build. You should disagree with some of them; that is the point.

Check your understanding

Quick check

—

Which of the following is the strongest reason a globally-trained AI model often performs worse in Nepal than in the US?

Nepali users are less technically skilled.
The model was trained on far less Nepali-language and Nepal-specific data than English-language and US-specific data.
Nepali devices have weaker chips than American devices.
Nepal has worse internet, which slows the model down.

Quick check

—

Why is it risky for Nepal to adopt the global AI policy agenda (copyright, safety frames, regulation) without modification?

It would offend Nepal's international partners.
Nepali policymakers are not capable of understanding it.
The global agenda is built around the problems and incentives of richer, more data-saturated countries — and silently codes those into the rules.
Global AI policy is always wrong and should be ignored.

What comes next

If the country is genuinely different — different data, different language, different devices — then the next question is concrete: what is already on the ground? The next section looks at the connectivity, devices, and existing datasets we actually have, and what they let you build today.