The choices we are making — and who makes them · AI, Ethics, and Society

Most of the writing you will read about AI ethics is, in some way, about a film. It imagines a future intelligence that becomes too powerful, escapes its lab, and either saves or destroys humanity. It is dramatic, it is cinematic, and — for the actual decisions being made in Nepal this year — it is almost entirely beside the point. The ethical choices that will shape how millions of Nepalis are treated by automated systems over the next decade are being made right now, by people with names and salaries and email addresses, and almost none of those people would describe what they are doing as “ethics.”

The thing that is actually happening

A procurement officer at a commercial bank in Kathmandu is comparing two credit-scoring vendors this quarter. One is cheaper, has slick marketing, and was trained mostly on Indian consumer data. The other is more expensive, slower to integrate, and at least claims to have looked at Nepali repayment patterns. The procurement officer will pick one based on a spreadsheet and a slide deck. That spreadsheet does not have a column called “ethics.” Whichever vendor wins, however, is going to be making decisions — quietly, at scale — about which Nepali small-business owners get a loan and which ones get a polite, automated rejection they will never be able to appeal.

A few kilometres away, an IT manager at a metropolitan office is being pitched a CCTV system with “AI-powered face recognition” by a vendor who has done similar installations in three Indian cities. The pitch is about catching pickpockets and finding missing children. Nobody in the room is asking what the model’s false-positive rate is on faces from the eastern hills, or where the matched faces go, or who can subpoena that database in five years. The manager will sign or not sign. That signature is the ethics.

A private school in Lalitpur is rolling out an AI tutor — an English-language large language model with a thin wrapper — for its grade-six students, most of whom learn maths in Nepali at home and English-medium at school. The school’s marketing brochure will call this “world-class.” Nobody is asking what happens to the child who phrases her confusion in Nepali and gets back a confident, wrong answer in English. The decision to deploy is the ethics.

Why “AGI” is a convenient distraction

There is a particular framing of AI ethics — popular among lab CEOs, podcast hosts, and a certain kind of think-tank — that treats the field as primarily about existential risk: the danger that, some years from now, a sufficiently powerful AI will go wrong in a civilisation-ending way. This framing is not stupid. There are serious people who believe it, and some of their arguments deserve attention.

But notice what the framing does in a country like Nepal. It moves the conversation from the bank’s loan model — which is real, deployed, and rejecting people this week — to a hypothetical future system that does not exist. It shifts authority from the people who could actually intervene (regulators, procurement officers, civil society) to the people who claim special knowledge of the future (foreign labs, foreign think-tanks, foreign consultants). And it casts the present-tense harms as small, manageable, beneath the dignity of “real” ethics work.

This is not an argument that long-term risks do not matter. It is an argument that, for almost everyone reading this, the long-term frame is not where their leverage is. A schoolteacher in Birgunj is not going to align a frontier model. She might, however, be the only adult who notices that her students’ AI tutor systematically misunderstands the Maithili-inflected English in their questions. That noticing is the ethics. So is what she does about it, and whether anyone listens.

Who actually decides

It is worth being concrete about who, in practice, makes ethical AI decisions in Nepal today. The list is not what the seminars suggest.

1. Procurement officers in banks, hospitals, ministries, and large NGOs. These are the people who choose between vendors. Their choices determine which models touch which Nepalis. Most of them have no formal AI training and almost no leverage to ask hard questions of slick foreign sales teams.

2. Engineers and product managers at Nepali product companies — Khalti, eSewa, Pathao, Foodmandu, Daraz Nepal — and the dozens of smaller fintechs and logistics firms. They decide what to log, what to model, what to A/B test, what to ship. A single Slack-message decision to use a fraud-detection threshold of 0.7 instead of 0.8 will, over a year, decline tens of thousands of legitimate transactions.

3. Regulators at NRB, NTA, NITC and the line ministries. They write the rules vendors must follow. When a regulator’s directive simply says “use industry best practices,” that is an ethical decision — it has just outsourced the ethics to whoever the vendor’s lawyers say “industry” is.

4. Users, silently. Every time a Nepali clicks “accept” on a terms-of-service in English they do not read, every time a small-business owner agrees to give a payments app access to their contacts, every time a parent installs an “AI homework helper” without checking what it does with their child’s data — that is also a decision. It does not feel like one. It is one.

5. Almost nobody you would call an “AI ethicist.” The job barely exists in Nepal. The few people with that title are usually academics or consultants writing papers that the procurement officer in paragraph one will never read. This is not a complaint about academics. It is a description of where the actual decisions sit.

What this means for the rest of the course

If you accept the argument above — that the real ethical action is in the present, distributed, and bureaucratic — then the rest of this course has a particular shape. It will not spend much time on superintelligence. It will spend a lot of time on the kinds of decisions a procurement officer, a regulator, a teacher, a journalist, a software engineer, or an informed citizen can actually intervene in. It will assume that “the ethics” is mostly something that gets built into a system at the moment someone signs a contract, picks a vendor, sets a threshold, or decides not to translate the interface into Nepali because “everyone knows English anyway.”

You will not become an ethicist by the end of this. You will, hopefully, become harder to fool — and more capable of asking the one or two questions in a procurement meeting that change which model gets deployed and which Nepalis it touches. That is, in 2026, what ethics work in this country mostly looks like.

Check your understanding

Quick check

—

In practice, who is making most of the consequential AI ethics decisions in Nepal right now?

Academic AI ethicists at Nepali universities, through their published papers.
Ordinary users, by carefully reading terms-of-service before clicking accept.
Procurement officers, engineers, product managers, and regulators — through vendor selections, threshold settings, and directive language — most of whom would not describe their work as “ethics.”
Foreign AI labs, through their alignment research.

Quick check

—

What is the main problem with framing AI ethics primarily as a question about future super-intelligence or AGI?

It is technically inaccurate — super-intelligence is impossible.
It deflects attention and authority away from the present-tense decisions — vendor choices, deployment calls, regulatory language — where the actual harms in Nepal are happening this year.
It is unpopular with the public, so it makes ethics work harder to fund.
It is a problem only for rich countries and does not apply in Nepal at all.

What comes next

If ethics is mostly about present-day decisions, the next question is: how do we evaluate them? The next section introduces the three lenses this course returns to throughout — fairness, autonomy, and accountability — and applies them to one concrete worked example: a fraud-detection model at a Nepali bank that occasionally freezes the wrong accounts.